android_kernel_xiaomi_sm8350/kernel/futex.c

2104 lines
49 KiB
C
Raw Normal View History

/*
* Fast Userspace Mutexes (which I call "Futexes!").
* (C) Rusty Russell, IBM 2002
*
* Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
* (C) Copyright 2003 Red Hat Inc, All Rights Reserved
*
* Removed page pinning, fix privately mapped COW pages and other cleanups
* (C) Copyright 2003, 2004 Jamie Lokier
*
* Robust futex support started by Ingo Molnar
* (C) Copyright 2006 Red Hat Inc, All Rights Reserved
* Thanks to Thomas Gleixner for suggestions, analysis and fixes.
*
* PI-futex support started by Ingo Molnar and Thomas Gleixner
* Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
* Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
*
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
* PRIVATE futexes by Eric Dumazet
* Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
*
* Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
* enough at me, Linus for the original (flawed) idea, Matthew
* Kirkwood for proof-of-concept implementation.
*
* "The futexes are also cursed."
* "But they come in a choice of three flavours!"
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
#include <linux/slab.h>
#include <linux/poll.h>
#include <linux/fs.h>
#include <linux/file.h>
#include <linux/jhash.h>
#include <linux/init.h>
#include <linux/futex.h>
#include <linux/mount.h>
#include <linux/pagemap.h>
#include <linux/syscalls.h>
#include <linux/signal.h>
#include <linux/module.h>
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
#include <asm/futex.h>
#include "rtmutex_common.h"
#define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
/*
* Priority Inheritance state:
*/
struct futex_pi_state {
/*
* list of 'owned' pi_state instances - these have to be
* cleaned up in do_exit() if the task exits prematurely:
*/
struct list_head list;
/*
* The PI object:
*/
struct rt_mutex pi_mutex;
struct task_struct *owner;
atomic_t refcount;
union futex_key key;
};
/*
* We use this hashed waitqueue instead of a normal wait_queue_t, so
* we can wake only the relevant ones (hashed queues may be shared).
*
* A futex_q has a woken state, just like tasks have TASK_RUNNING.
* It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
* The order of wakup is always to make the first condition true, then
* wake up q->waiters, then make the second condition true.
*/
struct futex_q {
struct plist_node list;
wait_queue_head_t waiters;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/* Which hash list lock to use: */
spinlock_t *lock_ptr;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/* Key which the futex is hashed on: */
union futex_key key;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/* For fd, sigio sent using these: */
int fd;
struct file *filp;
/* Optional priority inheritance state: */
struct futex_pi_state *pi_state;
struct task_struct *task;
};
/*
* Split the global futex_lock into every hash list lock.
*/
struct futex_hash_bucket {
spinlock_t lock;
struct plist_head chain;
};
static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
/* Futex-fs vfsmount entry: */
static struct vfsmount *futex_mnt;
/*
* We hash on the keys returned from get_futex_key (see below).
*/
static struct futex_hash_bucket *hash_futex(union futex_key *key)
{
u32 hash = jhash2((u32*)&key->both.word,
(sizeof(key->both.word)+sizeof(key->both.ptr))/4,
key->both.offset);
return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
}
/*
* Return 1 if two futex_keys are equal, 0 otherwise.
*/
static inline int match_futex(union futex_key *key1, union futex_key *key2)
{
return (key1->both.word == key2->both.word
&& key1->both.ptr == key2->both.ptr
&& key1->both.offset == key2->both.offset);
}
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
/**
* get_futex_key - Get parameters which are the keys for a futex.
* @uaddr: virtual address of the futex
* @shared: NULL for a PROCESS_PRIVATE futex,
* &current->mm->mmap_sem for a PROCESS_SHARED futex
* @key: address where result is stored.
*
* Returns a negative error code or 0
* The key words are stored in *key on success.
*
* For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
* offset_within_page). For private mappings, it's (uaddr, current->mm).
* We can usually work out the index without swapping in the page.
*
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
* fshared is NULL for PROCESS_PRIVATE futexes
* For other futexes, it points to &current->mm->mmap_sem and
* caller must have taken the reader lock. but NOT any spinlocks.
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
union futex_key *key)
{
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
unsigned long address = (unsigned long)uaddr;
struct mm_struct *mm = current->mm;
struct vm_area_struct *vma;
struct page *page;
int err;
/*
* The futex address must be "naturally" aligned.
*/
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
key->both.offset = address % PAGE_SIZE;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (unlikely((address % sizeof(u32)) != 0))
return -EINVAL;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
address -= key->both.offset;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
/*
* PROCESS_PRIVATE futexes are fast.
* As the mm cannot disappear under us and the 'key' only needs
* virtual address, we dont even have to find the underlying vma.
* Note : We do have to check 'uaddr' is a valid user address,
* but access_ok() should be faster than find_vma()
*/
if (!fshared) {
if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
return -EFAULT;
key->private.mm = mm;
key->private.address = address;
return 0;
}
/*
* The futex is hashed differently depending on whether
* it's in a shared or private mapping. So check vma first.
*/
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
vma = find_extend_vma(mm, address);
if (unlikely(!vma))
return -EFAULT;
/*
* Permissions.
*/
if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
/*
* Private mappings are handled in a simple way.
*
* NOTE: When userspace waits on a MAP_SHARED mapping, even if
* it's a read-only handle, it's expected that futexes attach to
* the object not the particular process. Therefore we use
* VM_MAYSHARE here, not VM_SHARED which is restricted to shared
* mappings of _writable_ handles.
*/
if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
key->private.mm = mm;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
key->private.address = address;
return 0;
}
/*
* Linear file mappings are also simple.
*/
key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
+ vma->vm_pgoff);
return 0;
}
/*
* We could walk the page table to read the non-linear
* pte, and get the page index without fetching the page
* from swap. But that's a lot of code to duplicate here
* for a rare case, so we simply fetch the page.
*/
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
if (err >= 0) {
key->shared.pgoff =
page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
put_page(page);
return 0;
}
return err;
}
EXPORT_SYMBOL_GPL(get_futex_key);
/*
* Take a reference to the resource addressed by a key.
* Can be called while holding spinlocks.
*
*/
inline void get_futex_key_refs(union futex_key *key)
{
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (key->both.ptr == 0)
return;
switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
case FUT_OFF_INODE:
atomic_inc(&key->shared.inode->i_count);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
break;
case FUT_OFF_MMSHARED:
atomic_inc(&key->private.mm->mm_count);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
break;
}
}
EXPORT_SYMBOL_GPL(get_futex_key_refs);
/*
* Drop a reference to the resource addressed by a key.
* The hash bucket spinlock must not be held.
*/
void drop_futex_key_refs(union futex_key *key)
{
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (key->both.ptr == 0)
return;
switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
case FUT_OFF_INODE:
iput(key->shared.inode);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
break;
case FUT_OFF_MMSHARED:
mmdrop(key->private.mm);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
break;
}
}
EXPORT_SYMBOL_GPL(drop_futex_key_refs);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
{
int ret;
pagefault_disable();
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
pagefault_enable();
return ret ? -EFAULT : 0;
}
/*
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
* Fault handling.
* if fshared is non NULL, current->mm->mmap_sem is already held
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
static int futex_handle_fault(unsigned long address,
struct rw_semaphore *fshared, int attempt)
{
struct vm_area_struct * vma;
struct mm_struct *mm = current->mm;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
int ret = -EFAULT;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (attempt > 2)
return ret;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (!fshared)
down_read(&mm->mmap_sem);
vma = find_vma(mm, address);
if (vma && address >= vma->vm_start &&
(vma->vm_flags & VM_WRITE)) {
switch (handle_mm_fault(mm, vma, address, 1)) {
case VM_FAULT_MINOR:
ret = 0;
current->min_flt++;
break;
case VM_FAULT_MAJOR:
ret = 0;
current->maj_flt++;
break;
}
}
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (!fshared)
up_read(&mm->mmap_sem);
return ret;
}
/*
* PI code:
*/
static int refill_pi_state_cache(void)
{
struct futex_pi_state *pi_state;
if (likely(current->pi_state_cache))
return 0;
pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
if (!pi_state)
return -ENOMEM;
INIT_LIST_HEAD(&pi_state->list);
/* pi_mutex gets initialized later */
pi_state->owner = NULL;
atomic_set(&pi_state->refcount, 1);
current->pi_state_cache = pi_state;
return 0;
}
static struct futex_pi_state * alloc_pi_state(void)
{
struct futex_pi_state *pi_state = current->pi_state_cache;
WARN_ON(!pi_state);
current->pi_state_cache = NULL;
return pi_state;
}
static void free_pi_state(struct futex_pi_state *pi_state)
{
if (!atomic_dec_and_test(&pi_state->refcount))
return;
/*
* If pi_state->owner is NULL, the owner is most probably dying
* and has cleaned up the pi_state already
*/
if (pi_state->owner) {
spin_lock_irq(&pi_state->owner->pi_lock);
list_del_init(&pi_state->list);
spin_unlock_irq(&pi_state->owner->pi_lock);
rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
}
if (current->pi_state_cache)
kfree(pi_state);
else {
/*
* pi_state->list is already empty.
* clear pi_state->owner.
* refcount is at 0 - put it back to 1.
*/
pi_state->owner = NULL;
atomic_set(&pi_state->refcount, 1);
current->pi_state_cache = pi_state;
}
}
/*
* Look up the task based on what TID userspace gave us.
* We dont trust it.
*/
static struct task_struct * futex_find_get_task(pid_t pid)
{
struct task_struct *p;
rcu_read_lock();
p = find_task_by_pid(pid);
if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
p = ERR_PTR(-ESRCH);
else
get_task_struct(p);
rcu_read_unlock();
return p;
}
/*
* This task is holding PI mutexes at exit time => bad.
* Kernel cleans up PI-state, but userspace is likely hosed.
* (Robust-futex cleanup is separate and might save the day for userspace.)
*/
void exit_pi_state_list(struct task_struct *curr)
{
struct list_head *next, *head = &curr->pi_state_list;
struct futex_pi_state *pi_state;
struct futex_hash_bucket *hb;
union futex_key key;
/*
* We are a ZOMBIE and nobody can enqueue itself on
* pi_state_list anymore, but we have to be careful
* versus waiters unqueueing themselves:
*/
spin_lock_irq(&curr->pi_lock);
while (!list_empty(head)) {
next = head->next;
pi_state = list_entry(next, struct futex_pi_state, list);
key = pi_state->key;
hb = hash_futex(&key);
spin_unlock_irq(&curr->pi_lock);
spin_lock(&hb->lock);
spin_lock_irq(&curr->pi_lock);
/*
* We dropped the pi-lock, so re-check whether this
* task still owns the PI-state:
*/
if (head->next != next) {
spin_unlock(&hb->lock);
continue;
}
WARN_ON(pi_state->owner != curr);
WARN_ON(list_empty(&pi_state->list));
list_del_init(&pi_state->list);
pi_state->owner = NULL;
spin_unlock_irq(&curr->pi_lock);
rt_mutex_unlock(&pi_state->pi_mutex);
spin_unlock(&hb->lock);
spin_lock_irq(&curr->pi_lock);
}
spin_unlock_irq(&curr->pi_lock);
}
static int
lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
union futex_key *key, struct futex_pi_state **ps)
{
struct futex_pi_state *pi_state = NULL;
struct futex_q *this, *next;
struct plist_head *head;
struct task_struct *p;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
pid_t pid = uval & FUTEX_TID_MASK;
head = &hb->chain;
plist_for_each_entry_safe(this, next, head, list) {
if (match_futex(&this->key, key)) {
/*
* Another waiter already exists - bump up
* the refcount and return its pi_state:
*/
pi_state = this->pi_state;
/*
* Userspace might have messed up non PI and PI futexes
*/
if (unlikely(!pi_state))
return -EINVAL;
WARN_ON(!atomic_read(&pi_state->refcount));
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
WARN_ON(pid && pi_state->owner &&
pi_state->owner->pid != pid);
atomic_inc(&pi_state->refcount);
*ps = pi_state;
return 0;
}
}
/*
* We are the first waiter - try to look up the real owner and attach
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
* the new pi_state to it, but bail out when TID = 0
*/
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
if (!pid)
return -ESRCH;
p = futex_find_get_task(pid);
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
if (IS_ERR(p))
return PTR_ERR(p);
/*
* We need to look at the task state flags to figure out,
* whether the task is exiting. To protect against the do_exit
* change of the task flags, we do this protected by
* p->pi_lock:
*/
spin_lock_irq(&p->pi_lock);
if (unlikely(p->flags & PF_EXITING)) {
/*
* The task is on the way out. When PF_EXITPIDONE is
* set, we know that the task has finished the
* cleanup:
*/
int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
spin_unlock_irq(&p->pi_lock);
put_task_struct(p);
return ret;
}
pi_state = alloc_pi_state();
/*
* Initialize the pi_mutex in locked state and make 'p'
* the owner of it:
*/
rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
/* Store the key for possible exit cleanups: */
pi_state->key = *key;
WARN_ON(!list_empty(&pi_state->list));
list_add(&pi_state->list, &p->pi_state_list);
pi_state->owner = p;
spin_unlock_irq(&p->pi_lock);
put_task_struct(p);
*ps = pi_state;
return 0;
}
/*
* The hash bucket lock must be held when this is called.
* Afterwards, the futex_q must not be accessed.
*/
static void wake_futex(struct futex_q *q)
{
plist_del(&q->list, &q->list.plist);
if (q->filp)
send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
/*
* The lock in wake_up_all() is a crucial memory barrier after the
* plist_del() and also before assigning to q->lock_ptr.
*/
wake_up_all(&q->waiters);
/*
* The waiting task can free the futex_q as soon as this is written,
* without taking any locks. This must come last.
*
* A memory barrier is required here to prevent the following store
* to lock_ptr from getting ahead of the wakeup. Clearing the lock
* at the end of wake_up_all() does not prevent this store from
* moving.
*/
smp_wmb();
q->lock_ptr = NULL;
}
static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
{
struct task_struct *new_owner;
struct futex_pi_state *pi_state = this->pi_state;
u32 curval, newval;
if (!pi_state)
return -EINVAL;
spin_lock(&pi_state->pi_mutex.wait_lock);
new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
/*
* This happens when we have stolen the lock and the original
* pending owner did not enqueue itself back on the rt_mutex.
* Thats not a tragedy. We know that way, that a lock waiter
* is on the fly. We make the futex_q waiter the pending owner.
*/
if (!new_owner)
new_owner = this->task;
/*
* We pass it to the next owner. (The WAITERS bit is always
* kept enabled while there is PI state around. We must also
* preserve the owner died bit.)
*/
if (!(uval & FUTEX_OWNER_DIED)) {
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
int ret = 0;
newval = FUTEX_WAITERS | new_owner->pid;
pagefault_disable();
curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
pagefault_enable();
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
if (curval == -EFAULT)
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
ret = -EFAULT;
if (curval != uval)
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
ret = -EINVAL;
if (ret) {
spin_unlock(&pi_state->pi_mutex.wait_lock);
return ret;
}
}
spin_lock_irq(&pi_state->owner->pi_lock);
WARN_ON(list_empty(&pi_state->list));
list_del_init(&pi_state->list);
spin_unlock_irq(&pi_state->owner->pi_lock);
spin_lock_irq(&new_owner->pi_lock);
WARN_ON(!list_empty(&pi_state->list));
list_add(&pi_state->list, &new_owner->pi_state_list);
pi_state->owner = new_owner;
spin_unlock_irq(&new_owner->pi_lock);
spin_unlock(&pi_state->pi_mutex.wait_lock);
rt_mutex_unlock(&pi_state->pi_mutex);
return 0;
}
static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
{
u32 oldval;
/*
* There is no waiter, so we unlock the futex. The owner died
* bit has not to be preserved here. We are the owner:
*/
pagefault_disable();
oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
pagefault_enable();
if (oldval == -EFAULT)
return oldval;
if (oldval != uval)
return -EAGAIN;
return 0;
}
/*
* Express the locking dependencies for lockdep:
*/
static inline void
double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
{
if (hb1 <= hb2) {
spin_lock(&hb1->lock);
if (hb1 < hb2)
spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
} else { /* hb1 > hb2 */
spin_lock(&hb2->lock);
spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
}
}
/*
* Wake up all waiters hashed on the physical page that is mapped
* to this virtual address:
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
int nr_wake)
{
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
struct futex_hash_bucket *hb;
struct futex_q *this, *next;
struct plist_head *head;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
union futex_key key;
int ret;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
down_read(fshared);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = get_futex_key(uaddr, fshared, &key);
if (unlikely(ret != 0))
goto out;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
hb = hash_futex(&key);
spin_lock(&hb->lock);
head = &hb->chain;
plist_for_each_entry_safe(this, next, head, list) {
if (match_futex (&this->key, &key)) {
if (this->pi_state) {
ret = -EINVAL;
break;
}
wake_futex(this);
if (++ret >= nr_wake)
break;
}
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
spin_unlock(&hb->lock);
out:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
return ret;
}
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
/*
* Wake up all waiters hashed on the physical page that is mapped
* to this virtual address:
*/
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
static int
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
u32 __user *uaddr2,
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
int nr_wake, int nr_wake2, int op)
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
{
union futex_key key1, key2;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
struct futex_hash_bucket *hb1, *hb2;
struct plist_head *head;
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
struct futex_q *this, *next;
int ret, op_ret, attempt = 0;
retryfull:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
down_read(fshared);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = get_futex_key(uaddr1, fshared, &key1);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
if (unlikely(ret != 0))
goto out;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = get_futex_key(uaddr2, fshared, &key2);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
if (unlikely(ret != 0))
goto out;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
hb1 = hash_futex(&key1);
hb2 = hash_futex(&key2);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
retry:
double_lock_hb(hb1, hb2);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
op_ret = futex_atomic_op_inuser(op, uaddr2);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
if (unlikely(op_ret < 0)) {
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
u32 dummy;
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
spin_unlock(&hb1->lock);
if (hb1 != hb2)
spin_unlock(&hb2->lock);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
#ifndef CONFIG_MMU
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/*
* we don't get EFAULT from MMU faults if we don't have an MMU,
* but we might get them from range checking
*/
ret = op_ret;
goto out;
#endif
if (unlikely(op_ret != -EFAULT)) {
ret = op_ret;
goto out;
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/*
* futex_atomic_op_inuser needs to both read and write
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
* *(int __user *)uaddr2, but we can't modify it
* non-atomically. Therefore, if get_user below is not
* enough, we need to handle the fault ourselves, while
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
* still holding the mmap_sem.
*/
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
if (attempt++) {
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_handle_fault((unsigned long)uaddr2,
fshared, attempt);
if (ret)
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
goto out;
goto retry;
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/*
* If we would have faulted, release mmap_sem,
* fault it in and start all over again.
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
ret = get_user(dummy, uaddr2);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
if (ret)
return ret;
goto retryfull;
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
head = &hb1->chain;
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
plist_for_each_entry_safe(this, next, head, list) {
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
if (match_futex (&this->key, &key1)) {
wake_futex(this);
if (++ret >= nr_wake)
break;
}
}
if (op_ret > 0) {
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
head = &hb2->chain;
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
op_ret = 0;
plist_for_each_entry_safe(this, next, head, list) {
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
if (match_futex (&this->key, &key2)) {
wake_futex(this);
if (++op_ret >= nr_wake2)
break;
}
}
ret += op_ret;
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
spin_unlock(&hb1->lock);
if (hb1 != hb2)
spin_unlock(&hb2->lock);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
out:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
return ret;
}
/*
* Requeue all waiters hashed on one physical page to another
* physical page.
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
u32 __user *uaddr2,
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
int nr_wake, int nr_requeue, u32 *cmpval)
{
union futex_key key1, key2;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
struct futex_hash_bucket *hb1, *hb2;
struct plist_head *head1;
struct futex_q *this, *next;
int ret, drop_count = 0;
retry:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
down_read(fshared);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = get_futex_key(uaddr1, fshared, &key1);
if (unlikely(ret != 0))
goto out;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = get_futex_key(uaddr2, fshared, &key2);
if (unlikely(ret != 0))
goto out;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
hb1 = hash_futex(&key1);
hb2 = hash_futex(&key2);
double_lock_hb(hb1, hb2);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
if (likely(cmpval != NULL)) {
u32 curval;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
ret = get_futex_value_locked(&curval, uaddr1);
if (unlikely(ret)) {
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
spin_unlock(&hb1->lock);
if (hb1 != hb2)
spin_unlock(&hb2->lock);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/*
* If we would have faulted, release mmap_sem, fault
* it in and start all over again.
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
ret = get_user(curval, uaddr1);
if (!ret)
goto retry;
return ret;
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
if (curval != *cmpval) {
ret = -EAGAIN;
goto out_unlock;
}
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
head1 = &hb1->chain;
plist_for_each_entry_safe(this, next, head1, list) {
if (!match_futex (&this->key, &key1))
continue;
if (++ret <= nr_wake) {
wake_futex(this);
} else {
/*
* If key1 and key2 hash to the same bucket, no need to
* requeue.
*/
if (likely(head1 != &hb2->chain)) {
plist_del(&this->list, &hb1->chain);
plist_add(&this->list, &hb2->chain);
this->lock_ptr = &hb2->lock;
#ifdef CONFIG_DEBUG_PI_LIST
this->list.plist.lock = &hb2->lock;
#endif
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
}
this->key = key2;
get_futex_key_refs(&key2);
drop_count++;
if (ret - nr_wake >= nr_requeue)
break;
}
}
out_unlock:
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
spin_unlock(&hb1->lock);
if (hb1 != hb2)
spin_unlock(&hb2->lock);
/* drop_futex_key_refs() must be called outside the spinlocks. */
while (--drop_count >= 0)
drop_futex_key_refs(&key1);
out:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
return ret;
}
/* The key must be already stored in q->key. */
static inline struct futex_hash_bucket *
queue_lock(struct futex_q *q, int fd, struct file *filp)
{
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
struct futex_hash_bucket *hb;
q->fd = fd;
q->filp = filp;
init_waitqueue_head(&q->waiters);
get_futex_key_refs(&q->key);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
hb = hash_futex(&q->key);
q->lock_ptr = &hb->lock;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
spin_lock(&hb->lock);
return hb;
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
{
int prio;
/*
* The priority used to register this element is
* - either the real thread-priority for the real-time threads
* (i.e. threads with a priority lower than MAX_RT_PRIO)
* - or MAX_RT_PRIO for non-RT threads.
* Thus, all RT-threads are woken first in priority order, and
* the others are woken last, in FIFO order.
*/
prio = min(current->normal_prio, MAX_RT_PRIO);
plist_node_init(&q->list, prio);
#ifdef CONFIG_DEBUG_PI_LIST
q->list.plist.lock = &hb->lock;
#endif
plist_add(&q->list, &hb->chain);
q->task = current;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
spin_unlock(&hb->lock);
}
static inline void
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
{
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
spin_unlock(&hb->lock);
drop_futex_key_refs(&q->key);
}
/*
* queue_me and unqueue_me must be called as a pair, each
* exactly once. They are called with the hashed spinlock held.
*/
/* The key must be already stored in q->key. */
static void queue_me(struct futex_q *q, int fd, struct file *filp)
{
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
struct futex_hash_bucket *hb;
hb = queue_lock(q, fd, filp);
__queue_me(q, hb);
}
/* Return 1 if we were still queued (ie. 0 means we were woken) */
static int unqueue_me(struct futex_q *q)
{
spinlock_t *lock_ptr;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
int ret = 0;
/* In the common case we don't take the spinlock, which is nice. */
retry:
lock_ptr = q->lock_ptr;
[PATCH] bug in futex unqueue_me This patch adds a barrier() in futex unqueue_me to avoid aliasing of two pointers. On my s390x system I saw the following oops: Unable to handle kernel pointer dereference at virtual kernel address 0000000000000000 Oops: 0004 [#1] CPU: 0 Not tainted Process mytool (pid: 13613, task: 000000003ecb6ac0, ksp: 00000000366bdbd8) Krnl PSW : 0704d00180000000 00000000003c9ac2 (_spin_lock+0xe/0x30) Krnl GPRS: 00000000ffffffff 000000003ecb6ac0 0000000000000000 0700000000000000 0000000000000000 0000000000000000 000001fe00002028 00000000000c091f 000001fe00002054 000001fe00002054 0000000000000000 00000000366bddc0 00000000005ef8c0 00000000003d00e8 0000000000144f91 00000000366bdcb8 Krnl Code: ba 4e 20 00 12 44 b9 16 00 3e a7 84 00 08 e3 e0 f0 88 00 04 Call Trace: ([<0000000000144f90>] unqueue_me+0x40/0xe4) [<0000000000145a0c>] do_futex+0x33c/0xc40 [<000000000014643e>] sys_futex+0x12e/0x144 [<000000000010bb00>] sysc_noemu+0x10/0x16 [<000002000003741c>] 0x2000003741c The code in question is: static int unqueue_me(struct futex_q *q) { int ret = 0; spinlock_t *lock_ptr; /* In the common case we don't take the spinlock, which is nice. */ retry: lock_ptr = q->lock_ptr; if (lock_ptr != 0) { spin_lock(lock_ptr); /* * q->lock_ptr can change between reading it and * spin_lock(), causing us to take the wrong lock. This * corrects the race condition. [...] and my compiler (gcc 4.1.0) makes the following out of it: 00000000000003c8 <unqueue_me>: 3c8: eb bf f0 70 00 24 stmg %r11,%r15,112(%r15) 3ce: c0 d0 00 00 00 00 larl %r13,3ce <unqueue_me+0x6> 3d0: R_390_PC32DBL .rodata+0x2a 3d4: a7 f1 1e 00 tml %r15,7680 3d8: a7 84 00 01 je 3da <unqueue_me+0x12> 3dc: b9 04 00 ef lgr %r14,%r15 3e0: a7 fb ff d0 aghi %r15,-48 3e4: b9 04 00 b2 lgr %r11,%r2 3e8: e3 e0 f0 98 00 24 stg %r14,152(%r15) 3ee: e3 c0 b0 28 00 04 lg %r12,40(%r11) /* write q->lock_ptr in r12 */ 3f4: b9 02 00 cc ltgr %r12,%r12 3f8: a7 84 00 4b je 48e <unqueue_me+0xc6> /* if r12 is zero then jump over the code.... */ 3fc: e3 20 b0 28 00 04 lg %r2,40(%r11) /* write q->lock_ptr in r2 */ 402: c0 e5 00 00 00 00 brasl %r14,402 <unqueue_me+0x3a> 404: R_390_PC32DBL _spin_lock+0x2 /* use r2 as parameter for spin_lock */ So the code becomes more or less: if (q->lock_ptr != 0) spin_lock(q->lock_ptr) instead of if (lock_ptr != 0) spin_lock(lock_ptr) Which caused the oops from above. After adding a barrier gcc creates code without this problem: [...] (the same) 3ee: e3 c0 b0 28 00 04 lg %r12,40(%r11) 3f4: b9 02 00 cc ltgr %r12,%r12 3f8: b9 04 00 2c lgr %r2,%r12 3fc: a7 84 00 48 je 48c <unqueue_me+0xc4> 400: c0 e5 00 00 00 00 brasl %r14,400 <unqueue_me+0x38> 402: R_390_PC32DBL _spin_lock+0x2 As a general note, this code of unqueue_me seems a bit fishy. The retry logic of unqueue_me only works if we can guarantee, that the original value of q->lock_ptr is always a spinlock (Otherwise we overwrite kernel memory). We know that q->lock_ptr can change. I dont know what happens with the original spinlock, as I am not an expert with the futex code. Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Thomas Gleixner <tglx@timesys.com> Signed-off-by: Christian Borntraeger <borntrae@de.ibm.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-08-05 15:13:52 -04:00
barrier();
if (lock_ptr != 0) {
spin_lock(lock_ptr);
/*
* q->lock_ptr can change between reading it and
* spin_lock(), causing us to take the wrong lock. This
* corrects the race condition.
*
* Reasoning goes like this: if we have the wrong lock,
* q->lock_ptr must have changed (maybe several times)
* between reading it and the spin_lock(). It can
* change again after the spin_lock() but only if it was
* already changed before the spin_lock(). It cannot,
* however, change back to the original value. Therefore
* we can detect whether we acquired the correct lock.
*/
if (unlikely(lock_ptr != q->lock_ptr)) {
spin_unlock(lock_ptr);
goto retry;
}
WARN_ON(plist_node_empty(&q->list));
plist_del(&q->list, &q->list.plist);
BUG_ON(q->pi_state);
spin_unlock(lock_ptr);
ret = 1;
}
drop_futex_key_refs(&q->key);
return ret;
}
/*
* PI futexes can not be requeued and must remove themself from the
* hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
* and dropped here.
*/
static void unqueue_me_pi(struct futex_q *q)
{
WARN_ON(plist_node_empty(&q->list));
plist_del(&q->list, &q->list.plist);
BUG_ON(!q->pi_state);
free_pi_state(q->pi_state);
q->pi_state = NULL;
spin_unlock(q->lock_ptr);
drop_futex_key_refs(&q->key);
}
/*
* Fixup the pi_state owner with current.
*
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
* Must be called with hash bucket lock held and mm->sem held for non
* private futexes.
*/
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
struct task_struct *curr)
{
u32 newtid = curr->pid | FUTEX_WAITERS;
struct futex_pi_state *pi_state = q->pi_state;
u32 uval, curval, newval;
int ret;
/* Owner died? */
if (pi_state->owner != NULL) {
spin_lock_irq(&pi_state->owner->pi_lock);
WARN_ON(list_empty(&pi_state->list));
list_del_init(&pi_state->list);
spin_unlock_irq(&pi_state->owner->pi_lock);
} else
newtid |= FUTEX_OWNER_DIED;
pi_state->owner = curr;
spin_lock_irq(&curr->pi_lock);
WARN_ON(!list_empty(&pi_state->list));
list_add(&pi_state->list, &curr->pi_state_list);
spin_unlock_irq(&curr->pi_lock);
/*
* We own it, so we have to replace the pending owner
* TID. This must be atomic as we have preserve the
* owner died bit here.
*/
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
ret = get_futex_value_locked(&uval, uaddr);
while (!ret) {
newval = (uval & FUTEX_OWNER_DIED) | newtid;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
pagefault_disable();
curval = futex_atomic_cmpxchg_inatomic(uaddr,
uval, newval);
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
pagefault_enable();
if (curval == -EFAULT)
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
ret = -EFAULT;
if (curval == uval)
break;
uval = curval;
}
return ret;
}
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
/*
* In case we must use restart_block to restart a futex_wait,
* we encode in the 'arg3' shared capability
*/
#define ARG3_SHARED 1
static long futex_wait_restart(struct restart_block *restart);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
u32 val, ktime_t *abs_time)
{
struct task_struct *curr = current;
DECLARE_WAITQUEUE(wait, curr);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
struct futex_hash_bucket *hb;
struct futex_q q;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
u32 uval;
int ret;
struct hrtimer_sleeper t;
int rem = 0;
q.pi_state = NULL;
retry:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
down_read(fshared);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = get_futex_key(uaddr, fshared, &q.key);
if (unlikely(ret != 0))
goto out_release_sem;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
hb = queue_lock(&q, -1, NULL);
/*
* Access the page AFTER the futex is queued.
* Order is important:
*
* Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
* Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
*
* The basic logical guarantee of a futex is that it blocks ONLY
* if cond(var) is known to be true at the time of blocking, for
* any cond. If we queued after testing *uaddr, that would open
* a race condition where we could block indefinitely with
* cond(var) false, which would violate the guarantee.
*
* A consequence is that futex_wait() can return zero and absorb
* a wakeup when *uaddr != val on entry to the syscall. This is
* rare, but normal.
*
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
* for shared futexes, we hold the mmap semaphore, so the mapping
* cannot have changed since we looked it up in get_futex_key.
*/
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
ret = get_futex_value_locked(&uval, uaddr);
if (unlikely(ret)) {
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
queue_unlock(&q, hb);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/*
* If we would have faulted, release mmap_sem, fault it in and
* start all over again.
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
ret = get_user(uval, uaddr);
if (!ret)
goto retry;
return ret;
}
ret = -EWOULDBLOCK;
if (uval != val)
goto out_unlock_release_sem;
/* Only actually queue if *uaddr contained val. */
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
__queue_me(&q, hb);
/*
* Now the futex is queued and we have checked the data, we
* don't want to hold mmap_sem while we sleep.
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
/*
* There might have been scheduling since the queue_me(), as we
* cannot hold a spinlock across the get_user() in case it
* faults, and we cannot just set TASK_INTERRUPTIBLE state when
* queueing ourselves into the futex hash. This code thus has to
* rely on the futex_wake() code removing us from hash when it
* wakes us up.
*/
/* add_wait_queue is the barrier after __set_current_state. */
__set_current_state(TASK_INTERRUPTIBLE);
add_wait_queue(&q.waiters, &wait);
/*
* !plist_node_empty() is safe here without any lock.
* q.lock_ptr != 0 is not safe, because of ordering against wakeup.
*/
if (likely(!plist_node_empty(&q.list))) {
if (!abs_time)
schedule();
else {
hrtimer_init(&t.timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
hrtimer_init_sleeper(&t, current);
t.timer.expires = *abs_time;
hrtimer_start(&t.timer, t.timer.expires, HRTIMER_MODE_ABS);
/*
* the timer could have already expired, in which
* case current would be flagged for rescheduling.
* Don't bother calling schedule.
*/
if (likely(t.task))
schedule();
hrtimer_cancel(&t.timer);
/* Flag if a timeout occured */
rem = (t.task == NULL);
}
}
__set_current_state(TASK_RUNNING);
/*
* NOTE: we don't remove ourselves from the waitqueue because
* we are the only user of it.
*/
/* If we were woken (and unqueued), we succeeded, whatever. */
if (!unqueue_me(&q))
return 0;
if (rem)
return -ETIMEDOUT;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
/*
* We expect signal_pending(current), but another thread may
* have handled it for us already.
*/
if (!abs_time)
return -ERESTARTSYS;
else {
struct restart_block *restart;
restart = &current_thread_info()->restart_block;
restart->fn = futex_wait_restart;
restart->arg0 = (unsigned long)uaddr;
restart->arg1 = (unsigned long)val;
restart->arg2 = (unsigned long)abs_time;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
restart->arg3 = 0;
if (fshared)
restart->arg3 |= ARG3_SHARED;
return -ERESTART_RESTARTBLOCK;
}
out_unlock_release_sem:
queue_unlock(&q, hb);
out_release_sem:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
return ret;
}
static long futex_wait_restart(struct restart_block *restart)
{
u32 __user *uaddr = (u32 __user *)restart->arg0;
u32 val = (u32)restart->arg1;
ktime_t *abs_time = (ktime_t *)restart->arg2;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
struct rw_semaphore *fshared = NULL;
restart->fn = do_no_restart_syscall;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (restart->arg3 & ARG3_SHARED)
fshared = &current->mm->mmap_sem;
return (long)futex_wait(uaddr, fshared, val, abs_time);
}
/*
* Userspace tried a 0 -> TID atomic transition of the futex value
* and failed. The kernel side here does the whole locking operation:
* if there are waiters then it will block, it does PI, etc. (Due to
* races the kernel might see a 0 value of the futex too.)
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
int detect, ktime_t *time, int trylock)
{
struct hrtimer_sleeper timeout, *to = NULL;
struct task_struct *curr = current;
struct futex_hash_bucket *hb;
u32 uval, newval, curval;
struct futex_q q;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
int ret, lock_taken, ownerdied = 0, attempt = 0;
if (refill_pi_state_cache())
return -ENOMEM;
if (time) {
to = &timeout;
hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
hrtimer_init_sleeper(to, current);
to->timer.expires = *time;
}
q.pi_state = NULL;
retry:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
down_read(fshared);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = get_futex_key(uaddr, fshared, &q.key);
if (unlikely(ret != 0))
goto out_release_sem;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
retry_unlocked:
hb = queue_lock(&q, -1, NULL);
retry_locked:
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
ret = lock_taken = 0;
/*
* To avoid races, we attempt to take the lock here again
* (by doing a 0 -> TID atomic cmpxchg), while holding all
* the locks. It will most likely not succeed.
*/
newval = current->pid;
pagefault_disable();
curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
pagefault_enable();
if (unlikely(curval == -EFAULT))
goto uaddr_faulted;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
/*
* Detect deadlocks. In case of REQUEUE_PI this is a valid
* situation and we return success to user space.
*/
if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
ret = -EDEADLK;
goto out_unlock_release_sem;
}
/*
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
* Surprise - we got the lock. Just return to userspace:
*/
if (unlikely(!curval))
goto out_unlock_release_sem;
uval = curval;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
/*
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
* Set the WAITERS flag, so the owner will know it has someone
* to wake at next unlock
*/
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
newval = curval | FUTEX_WAITERS;
/*
* There are two cases, where a futex might have no owner (the
* owner TID is 0): OWNER_DIED. We take over the futex in this
* case. We also do an unconditional take over, when the owner
* of the futex died.
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
*
* This is safe as we are protected by the hash bucket lock !
*/
if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
/* Keep the OWNER_DIED bit */
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
newval = (curval & ~FUTEX_TID_MASK) | current->pid;
ownerdied = 0;
lock_taken = 1;
}
pagefault_disable();
curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
pagefault_enable();
if (unlikely(curval == -EFAULT))
goto uaddr_faulted;
if (unlikely(curval != uval))
goto retry_locked;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
/*
* We took the lock due to owner died take over.
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
*/
if (unlikely(lock_taken))
goto out_unlock_release_sem;
/*
* We dont have the lock. Look up the PI state (or create it if
* we are the first waiter):
*/
ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
if (unlikely(ret)) {
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
switch (ret) {
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
case -EAGAIN:
/*
* Task is exiting and we just wait for the
* exit to complete.
*/
queue_unlock(&q, hb);
if (fshared)
up_read(fshared);
cond_resched();
goto retry;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
case -ESRCH:
/*
* No owner found for this futex. Check if the
* OWNER_DIED bit is set to figure out whether
* this is a robust futex or not.
*/
if (get_futex_value_locked(&curval, uaddr))
goto uaddr_faulted;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
/*
* We simply start over in case of a robust
* futex. The code above will take the futex
* and return happy.
*/
if (curval & FUTEX_OWNER_DIED) {
ownerdied = 1;
goto retry_locked;
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
}
default:
goto out_unlock_release_sem;
}
}
/*
* Only actually queue now that the atomic ops are done:
*/
__queue_me(&q, hb);
/*
* Now the futex is queued and we have checked the data, we
* don't want to hold mmap_sem while we sleep.
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
WARN_ON(!q.pi_state);
/*
* Block on the PI mutex:
*/
if (!trylock)
ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
else {
ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
/* Fixup the trylock return value: */
ret = ret ? 0 : -EWOULDBLOCK;
}
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
down_read(fshared);
spin_lock(q.lock_ptr);
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
if (!ret) {
/*
* Got the lock. We might not be the anticipated owner
* if we did a lock-steal - fix up the PI-state in
* that case:
*/
if (q.pi_state->owner != curr)
ret = fixup_pi_state_owner(uaddr, &q, curr);
} else {
/*
* Catch the rare case, where the lock was released
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
* when we were on the way back before we locked the
* hash bucket.
*/
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
if (q.pi_state->owner == curr &&
rt_mutex_trylock(&q.pi_state->pi_mutex)) {
ret = 0;
} else {
/*
* Paranoia check. If we did not take the lock
* in the trylock above, then we should not be
* the owner of the rtmutex, neither the real
* nor the pending one:
*/
if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
printk(KERN_ERR "futex_lock_pi: ret = %d "
"pi-mutex: %p pi-state %p\n", ret,
q.pi_state->pi_mutex.owner,
q.pi_state->owner);
}
}
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
/* Unqueue and drop the lock */
unqueue_me_pi(&q);
if (fshared)
up_read(fshared);
return ret != -EINTR ? ret : -ERESTARTNOINTR;
out_unlock_release_sem:
queue_unlock(&q, hb);
out_release_sem:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
return ret;
uaddr_faulted:
/*
* We have to r/w *(int __user *)uaddr, but we can't modify it
* non-atomically. Therefore, if get_user below is not
* enough, we need to handle the fault ourselves, while
* still holding the mmap_sem.
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
*
* ... and hb->lock. :-) --ANK
*/
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
queue_unlock(&q, hb);
if (attempt++) {
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_handle_fault((unsigned long)uaddr, fshared,
attempt);
if (ret)
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
goto out_release_sem;
goto retry_unlocked;
}
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
ret = get_user(uval, uaddr);
if (!ret && (uval != -EFAULT))
goto retry;
return ret;
}
/*
* Userspace attempted a TID -> 0 atomic transition, and failed.
* This is the in-kernel slowpath: we look up the PI state (if any),
* and do the rt-mutex unlock.
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
{
struct futex_hash_bucket *hb;
struct futex_q *this, *next;
u32 uval;
struct plist_head *head;
union futex_key key;
int ret, attempt = 0;
retry:
if (get_user(uval, uaddr))
return -EFAULT;
/*
* We release only a lock we actually own:
*/
if ((uval & FUTEX_TID_MASK) != current->pid)
return -EPERM;
/*
* First take all the futex related locks:
*/
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
down_read(fshared);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = get_futex_key(uaddr, fshared, &key);
if (unlikely(ret != 0))
goto out;
hb = hash_futex(&key);
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
retry_unlocked:
spin_lock(&hb->lock);
/*
* To avoid races, try to do the TID -> 0 atomic transition
* again. If it succeeds then we can return without waking
* anyone else up:
*/
if (!(uval & FUTEX_OWNER_DIED)) {
pagefault_disable();
uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
pagefault_enable();
}
if (unlikely(uval == -EFAULT))
goto pi_faulted;
/*
* Rare case: we managed to release the lock atomically,
* no need to wake anyone else up:
*/
if (unlikely(uval == current->pid))
goto out_unlock;
/*
* Ok, other tasks may need to be woken up - check waiters
* and do the wakeup if necessary:
*/
head = &hb->chain;
plist_for_each_entry_safe(this, next, head, list) {
if (!match_futex (&this->key, &key))
continue;
ret = wake_futex_pi(uaddr, uval, this);
/*
* The atomic access to the futex value
* generated a pagefault, so retry the
* user-access and the wakeup:
*/
if (ret == -EFAULT)
goto pi_faulted;
goto out_unlock;
}
/*
* No waiters - kernel unlocks the futex:
*/
if (!(uval & FUTEX_OWNER_DIED)) {
ret = unlock_futex_pi(uaddr, uval);
if (ret == -EFAULT)
goto pi_faulted;
}
out_unlock:
spin_unlock(&hb->lock);
out:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
return ret;
pi_faulted:
/*
* We have to r/w *(int __user *)uaddr, but we can't modify it
* non-atomically. Therefore, if get_user below is not
* enough, we need to handle the fault ourselves, while
* still holding the mmap_sem.
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
*
* ... and hb->lock. --ANK
*/
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
spin_unlock(&hb->lock);
if (attempt++) {
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_handle_fault((unsigned long)uaddr, fshared,
attempt);
if (ret)
pi-futex: fix exit races and locking problems 1. New entries can be added to tsk->pi_state_list after task completed exit_pi_state_list(). The result is memory leakage and deadlocks. 2. handle_mm_fault() is called under spinlock. The result is obvious. 3. results in self-inflicted deadlock inside glibc. Sometimes futex_lock_pi returns -ESRCH, when it is not expected and glibc enters to for(;;) sleep() to simulate deadlock. This problem is quite obvious and I think the patch is right. Though it looks like each "if" in futex_lock_pi() got some stupid special case "else if". :-) 4. sometimes futex_lock_pi() returns -EDEADLK, when nobody has the lock. The reason is also obvious (see comment in the patch), but correct fix is far beyond my comprehension. I guess someone already saw this, the chunk: if (rt_mutex_trylock(&q.pi_state->pi_mutex)) ret = 0; is obviously from the same opera. But it does not work, because the rtmutex is really taken at this point: wake_futex_pi() of previous owner reassigned it to us. My fix works. But it looks very stupid. I would think about removal of shift of ownership in wake_futex_pi() and making all the work in context of process taking lock. From: Thomas Gleixner <tglx@linutronix.de> Fix 1) Avoid the tasklist lock variant of the exit race fix by adding an additional state transition to the exit code. This fixes also the issue, when a task with recursive segfaults is not able to release the futexes. Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH problem finally. Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup in the lock protected section by using the in_atomic userspace access functions. This removes also the ugly lock drop / unqueue inside of fixup_pi_state() Fix 4) Fix a stale lock in the error path of futex_wake_pi() Added some error checks for verification. The -EDEADLK problem is solved by the rtmutex fixups. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-06-08 16:47:00 -04:00
goto out;
goto retry_unlocked;
}
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (fshared)
up_read(fshared);
ret = get_user(uval, uaddr);
if (!ret && (uval != -EFAULT))
goto retry;
return ret;
}
static int futex_close(struct inode *inode, struct file *filp)
{
struct futex_q *q = filp->private_data;
unqueue_me(q);
kfree(q);
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
return 0;
}
/* This is one-shot: once it's gone off you need a new fd */
static unsigned int futex_poll(struct file *filp,
struct poll_table_struct *wait)
{
struct futex_q *q = filp->private_data;
int ret = 0;
poll_wait(filp, &q->waiters, wait);
/*
* plist_node_empty() is safe here without any lock.
* q->lock_ptr != 0 is not safe, because of ordering against wakeup.
*/
if (plist_node_empty(&q->list))
ret = POLLIN | POLLRDNORM;
return ret;
}
static const struct file_operations futex_fops = {
.release = futex_close,
.poll = futex_poll,
};
/*
* Signal allows caller to avoid the race which would occur if they
* set the sigio stuff up afterwards.
*/
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
static int futex_fd(u32 __user *uaddr, int signal)
{
struct futex_q *q;
struct file *filp;
int ret, err;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
struct rw_semaphore *fshared;
static unsigned long printk_interval;
if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
"will be removed from the kernel in June 2007\n",
current->comm);
}
ret = -EINVAL;
if (!valid_signal(signal))
goto out;
ret = get_unused_fd();
if (ret < 0)
goto out;
filp = get_empty_filp();
if (!filp) {
put_unused_fd(ret);
ret = -ENFILE;
goto out;
}
filp->f_op = &futex_fops;
filp->f_path.mnt = mntget(futex_mnt);
filp->f_path.dentry = dget(futex_mnt->mnt_root);
filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
if (signal) {
err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
if (err < 0) {
goto error;
}
filp->f_owner.signum = signal;
}
q = kmalloc(sizeof(*q), GFP_KERNEL);
if (!q) {
err = -ENOMEM;
goto error;
}
q->pi_state = NULL;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
fshared = &current->mm->mmap_sem;
down_read(fshared);
err = get_futex_key(uaddr, fshared, &q->key);
if (unlikely(err != 0)) {
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
up_read(fshared);
kfree(q);
goto error;
}
/*
* queue_me() must be called before releasing mmap_sem, because
* key->shared.inode needs to be referenced while holding it.
*/
filp->private_data = q;
queue_me(q, ret, filp);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
up_read(fshared);
/* Now we map fd to filp, so userspace can access it */
fd_install(ret, filp);
out:
return ret;
error:
put_unused_fd(ret);
put_filp(filp);
ret = err;
goto out;
}
/*
* Support for robust futexes: the kernel cleans up held futexes at
* thread exit time.
*
* Implementation: user-space maintains a per-thread list of locks it
* is holding. Upon do_exit(), the kernel carefully walks this list,
* and marks all locks that are owned by this thread with the
* FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
* always manipulated with the lock held, so the list is private and
* per-thread. Userspace also maintains a per-thread 'list_op_pending'
* field, to allow the kernel to clean up if the thread dies after
* acquiring the lock, but just before it could have added itself to
* the list. There can only be one such pending lock.
*/
/**
* sys_set_robust_list - set the robust-futex list head of a task
* @head: pointer to the list-head
* @len: length of the list-head, as userspace expects
*/
asmlinkage long
sys_set_robust_list(struct robust_list_head __user *head,
size_t len)
{
/*
* The kernel knows only one size for now:
*/
if (unlikely(len != sizeof(*head)))
return -EINVAL;
current->robust_list = head;
return 0;
}
/**
* sys_get_robust_list - get the robust-futex list head of a task
* @pid: pid of the process [zero for current task]
* @head_ptr: pointer to a list-head pointer, the kernel fills it in
* @len_ptr: pointer to a length field, the kernel fills in the header size
*/
asmlinkage long
sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
size_t __user *len_ptr)
{
struct robust_list_head __user *head;
unsigned long ret;
if (!pid)
head = current->robust_list;
else {
struct task_struct *p;
ret = -ESRCH;
rcu_read_lock();
p = find_task_by_pid(pid);
if (!p)
goto err_unlock;
ret = -EPERM;
if ((current->euid != p->euid) && (current->euid != p->uid) &&
!capable(CAP_SYS_PTRACE))
goto err_unlock;
head = p->robust_list;
rcu_read_unlock();
}
if (put_user(sizeof(*head), len_ptr))
return -EFAULT;
return put_user(head, head_ptr);
err_unlock:
rcu_read_unlock();
return ret;
}
/*
* Process a futex-list entry, check whether it's owned by the
* dying task, and do notification if so:
*/
int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
{
u32 uval, nval, mval;
retry:
if (get_user(uval, uaddr))
return -1;
if ((uval & FUTEX_TID_MASK) == curr->pid) {
/*
* Ok, this dying thread is truly holding a futex
* of interest. Set the OWNER_DIED bit atomically
* via cmpxchg, and if the value had FUTEX_WAITERS
* set, wake up a waiter (if any). (We have to do a
* futex_wake() even if OWNER_DIED is already set -
* to handle the rare but possible case of recursive
* thread-death.) The rest of the cleanup is done in
* userspace.
*/
mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
if (nval == -EFAULT)
return -1;
if (nval != uval)
goto retry;
/*
* Wake robust non-PI futexes here. The wakeup of
* PI futexes happens in exit_pi_state():
*/
if (!pi) {
if (uval & FUTEX_WAITERS)
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
futex_wake(uaddr, &curr->mm->mmap_sem, 1);
}
}
return 0;
}
/*
* Fetch a robust-list pointer. Bit 0 signals PI futexes:
*/
static inline int fetch_robust_entry(struct robust_list __user **entry,
struct robust_list __user * __user *head,
int *pi)
{
unsigned long uentry;
if (get_user(uentry, (unsigned long __user *)head))
return -EFAULT;
*entry = (void __user *)(uentry & ~1UL);
*pi = uentry & 1;
return 0;
}
/*
* Walk curr->robust_list (very carefully, it's a userspace list!)
* and mark any locks found there dead, and notify any waiters.
*
* We silently return on any sign of list-walking problem.
*/
void exit_robust_list(struct task_struct *curr)
{
struct robust_list_head __user *head = curr->robust_list;
struct robust_list __user *entry, *pending;
unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
unsigned long futex_offset;
/*
* Fetch the list head (which was registered earlier, via
* sys_set_robust_list()):
*/
if (fetch_robust_entry(&entry, &head->list.next, &pi))
return;
/*
* Fetch the relative futex offset:
*/
if (get_user(futex_offset, &head->futex_offset))
return;
/*
* Fetch any possibly pending lock-add first, and handle it
* if it exists:
*/
if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
return;
if (pending)
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
handle_futex_death((void __user *)pending + futex_offset,
curr, pip);
while (entry != &head->list) {
/*
* A pending lock might already be on the list, so
* don't process it twice:
*/
if (entry != pending)
if (handle_futex_death((void __user *)entry + futex_offset,
curr, pi))
return;
/*
* Fetch the next entry in the list:
*/
if (fetch_robust_entry(&entry, &entry->next, &pi))
return;
/*
* Avoid excessively long or circular lists:
*/
if (!--limit)
break;
cond_resched();
}
}
long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
u32 __user *uaddr2, u32 val2, u32 val3)
{
int ret;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
int cmd = op & FUTEX_CMD_MASK;
struct rw_semaphore *fshared = NULL;
if (!(op & FUTEX_PRIVATE_FLAG))
fshared = &current->mm->mmap_sem;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
switch (cmd) {
case FUTEX_WAIT:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_wait(uaddr, fshared, val, timeout);
break;
case FUTEX_WAKE:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_wake(uaddr, fshared, val);
break;
case FUTEX_FD:
/* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
ret = futex_fd(uaddr, val);
break;
case FUTEX_REQUEUE:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
break;
case FUTEX_CMP_REQUEUE:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
break;
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
case FUTEX_WAKE_OP:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter (which at least on UP usually means an immediate context switch to one of the waiter threads). This waiter wakes up and after a few instructions it attempts to acquire the cv internal lock, but that lock is still held by the thread calling pthread_cond_signal. So it goes to sleep and eventually the signalling thread is scheduled in, unlocks the internal lock and wakes the waiter again. Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal to avoid this performance issue, but it was removed when locks were redesigned to the 3 state scheme (unlocked, locked uncontended, locked contended). Following scenario shows why simply using FUTEX_REQUEUE in pthread_cond_signal together with using lll_mutex_unlock_force in place of lll_mutex_unlock is not enough and probably why it has been disabled at that time: The number is value in cv->__data.__lock. thr1 thr2 thr3 0 pthread_cond_wait 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) 0 lll_futex_wait (&cv->__data.__futex, futexval) 0 pthread_cond_signal 1 lll_mutex_lock (cv->__data.__lock) 1 pthread_cond_signal 2 lll_mutex_lock (cv->__data.__lock) 2 lll_futex_wait (&cv->__data.__lock, 2) 2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock) # FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE 2 lll_mutex_unlock_force (cv->__data.__lock) 0 cv->__data.__lock = 0 0 lll_futex_wake (&cv->__data.__lock, 1) 1 lll_mutex_lock (cv->__data.__lock) 0 lll_mutex_unlock (cv->__data.__lock) # Here, lll_mutex_unlock doesn't know there are threads waiting # on the internal cv's lock Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal, but it will cost us not one, but 2 extra syscalls and, what's worse, one of these extra syscalls will be done for every single waiting loop in pthread_cond_*wait. We would need to use lll_mutex_unlock_force in pthread_cond_signal after requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait. Another alternative is to do the unlocking pthread_cond_signal needs to do (the lock can't be unlocked before lll_futex_wake, as that is racy) in the kernel. I have implemented both variants, futex-requeue-glibc.patch is the first one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel. The kernel interface allows userland to specify how exactly an unlocking operation should look like (some atomic arithmetic operation with optional constant argument and comparison of the previous futex value with another constant). It has been implemented just for ppc*, x86_64 and i?86, for other architectures I'm including just a stub header which can be used as a starting point by maintainers to write support for their arches and ATM will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been (lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running 32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL. With the following benchmark on UP x86-64 I get: for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \ for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done time elf/ld.so --library-path .:nptl-orig /tmp/bench real 0m0.655s user 0m0.253s sys 0m0.403s real 0m0.657s user 0m0.269s sys 0m0.388s time elf/ld.so --library-path .:nptl-requeue /tmp/bench real 0m0.496s user 0m0.225s sys 0m0.271s real 0m0.531s user 0m0.242s sys 0m0.288s time elf/ld.so --library-path .:nptl-wake_op /tmp/bench real 0m0.380s user 0m0.176s sys 0m0.204s real 0m0.382s user 0m0.175s sys 0m0.207s The benchmark is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt Older futex-requeue-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt Older futex-wake_op-glibc.patch version is at: http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt Will post a new version (just x86-64 fixes so that the patch applies against pthread_cond_signal.S) to libc-hacker ml soon. Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded testcase that will not test the atomicity of the operation, but at least check if the threads that should have been woken up are woken up and whether the arithmetic operation in the kernel gave the expected results. Acked-by: Ingo Molnar <mingo@redhat.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jamie Lokier <jamie@shareable.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 18:16:25 -04:00
break;
case FUTEX_LOCK_PI:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
break;
case FUTEX_UNLOCK_PI:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_unlock_pi(uaddr, fshared);
break;
case FUTEX_TRYLOCK_PI:
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
break;
default:
ret = -ENOSYS;
}
return ret;
}
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
struct timespec __user *utime, u32 __user *uaddr2,
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
u32 val3)
{
struct timespec ts;
ktime_t t, *tp = NULL;
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
u32 val2 = 0;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
int cmd = op & FUTEX_CMD_MASK;
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI)) {
if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
return -EFAULT;
if (!timespec_valid(&ts))
return -EINVAL;
t = timespec_to_ktime(ts);
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
if (cmd == FUTEX_WAIT)
t = ktime_add(ktime_get(), t);
tp = &t;
}
/*
FUTEX: new PRIVATE futexes Analysis of current linux futex code : -------------------------------------- A central hash table futex_queues[] holds all contexts (futex_q) of waiting threads. Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to perform lookups or insert/deletion of a futex_q. When a futex_wait() is done, calling thread has to : 1) - Obtain a read lock on mmap_sem to be able to validate the user pointer (calling find_vma()). This validation tells us if the futex uses an inode based store (mapped file), or mm based store (anonymous mem) 2) - compute a hash key 3) - Atomic increment of reference counter on an inode or a mm_struct 4) - lock part of futex_queues[] hash table 5) - perform the test on value of futex. (rollback is value != expected_value, returns EWOULDBLOCK) (various loops if test triggers mm faults) 6) queue the context into hash table, release the lock got in 4) 7) - release the read_lock on mmap_sem <block> 8) Eventually unqueue the context (but rarely, as this part  may be done by the futex_wake()) Futexes were designed to improve scalability but current implementation has various problems : - Central hashtable : This means scalability problems if many processes/threads want to use futexes at the same time. This means NUMA unbalance because this hashtable is located on one node. - Using mmap_sem on every futex() syscall : Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic ops on mmap_sem, dirtying cache line : - lot of cache line ping pongs on SMP configurations. mmap_sem is also extensively used by mm code (page faults, mmap()/munmap()) Highly threaded processes might suffer from mmap_sem contention. mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded programs because of contention on the mmap_sem cache line. - Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter: It's also a cache line ping pong on SMP. It also increases mmap_sem hold time because of cache misses. Most of these scalability problems come from the fact that futexes are in one global namespace. As we use a central hash table, we must make sure they are all using the same reference (given by the mm subsystem). We chose to force all futexes be 'shared'. This has a cost. But fact is POSIX defined PRIVATE and SHARED, allowing clear separation, and optimal performance if carefuly implemented. Time has come for linux to have better threading performance. The goal is to permit new futex commands to avoid : - Taking the mmap_sem semaphore, conflicting with other subsystems. - Modifying a ref_count on mm or an inode, still conflicting with mm or fs. This is possible because, for one process using PTHREAD_PROCESS_PRIVATE futexes, we only need to distinguish futexes by their virtual address, no matter the underlying mm storage is. If glibc wants to exploit this new infrastructure, it should use new _PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be prepared to fallback on old subcommands for old kernels. Using one global variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK. PTHREAD_PROCESS_SHARED futexes should still use the old subcommands. Compatibility with old applications is preserved, they still hit the scalability problems, but new applications can fly :) Note : the same SHARED futex (mapped on a file) can be used by old binaries *and* new binaries, because both binaries will use the old subcommands. Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic, as this is the default semantic. Almost all applications should benefit of this changes (new kernel and updated libc) Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine) /* calling futex_wait(addr, value) with value != *addr */ 433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes) 424 cycles per futex(FUTEX_WAIT) call (using one futex) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes) 334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex) For reference : 187 cycles per getppid() call 188 cycles per umask() call 181 cycles per ni_syscall() call Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Pierre Peiffer <pierre.peiffer@bull.net> Cc: "Ulrich Drepper" <drepper@gmail.com> Cc: "Nick Piggin" <nickpiggin@yahoo.com.au> Cc: "Ingo Molnar" <mingo@elte.hu> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 05:35:04 -04:00
* requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
*/
if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE)
[PATCH] pi-futex: futex code cleanups We are pleased to announce "lightweight userspace priority inheritance" (PI) support for futexes. The following patchset and glibc patch implements it, ontop of the robust-futexes patchset which is included in 2.6.16-mm1. We are calling it lightweight for 3 reasons: - in the user-space fastpath a PI-enabled futex involves no kernel work (or any other PI complexity) at all. No registration, no extra kernel calls - just pure fast atomic ops in userspace. - in the slowpath (in the lock-contention case), the system call and scheduling pattern is in fact better than that of normal futexes, due to the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down] - the in-kernel PI implementation is streamlined around the mutex abstraction, with strict rules that keep the implementation relatively simple: only a single owner may own a lock (i.e. no read-write lock support), only the owner may unlock a lock, no recursive locking, etc. Priority Inheritance - why, oh why??? ------------------------------------- Many of you heard the horror stories about the evil PI code circling Linux for years, which makes no real sense at all and is only used by buggy applications and which has horrible overhead. Some of you have dreaded this very moment, when someone actually submits working PI code ;-) So why would we like to see PI support for futexes? We'd like to see it done purely for technological reasons. We dont think it's a buggy concept, we think it's useful functionality to offer to applications, which functionality cannot be achieved in other ways. We also think it's the right thing to do, and we think we've got the right arguments and the right numbers to prove that. We also believe that we can address all the counter-arguments as well. For these reasons (and the reasons outlined below) we are submitting this patch-set for upstream kernel inclusion. What are the benefits of PI? The short reply: ---------------- User-space PI helps achieving/improving determinism for user-space applications. In the best-case, it can help achieve determinism and well-bound latencies. Even in the worst-case, PI will improve the statistical distribution of locking related application delays. The longer reply: ----------------- Firstly, sharing locks between multiple tasks is a common programming technique that often cannot be replaced with lockless algorithms. As we can see it in the kernel [which is a quite complex program in itself], lockless structures are rather the exception than the norm - the current ratio of lockless vs. locky code for shared data structures is somewhere between 1:10 and 1:100. Lockless is hard, and the complexity of lockless algorithms often endangers to ability to do robust reviews of said code. I.e. critical RT apps often choose lock structures to protect critical data structures, instead of lockless algorithms. Furthermore, there are cases (like shared hardware, or other resource limits) where lockless access is mathematically impossible. Media players (such as Jack) are an example of reasonable application design with multiple tasks (with multiple priority levels) sharing short-held locks: for example, a highprio audio playback thread is combined with medium-prio construct-audio-data threads and low-prio display-colory-stuff threads. Add video and decoding to the mix and we've got even more priority levels. So once we accept that synchronization objects (locks) are an unavoidable fact of life, and once we accept that multi-task userspace apps have a very fair expectation of being able to use locks, we've got to think about how to offer the option of a deterministic locking implementation to user-space. Most of the technical counter-arguments against doing priority inheritance only apply to kernel-space locks. But user-space locks are different, there we cannot disable interrupts or make the task non-preemptible in a critical section, so the 'use spinlocks' argument does not apply (user-space spinlocks have the same priority inversion problems as other user-space locking constructs). Fact is, pretty much the only technique that currently enables good determinism for userspace locks (such as futex-based pthread mutexes) is priority inheritance: Currently (without PI), if a high-prio and a low-prio task shares a lock [this is a quite common scenario for most non-trivial RT applications], even if all critical sections are coded carefully to be deterministic (i.e. all critical sections are short in duration and only execute a limited number of instructions), the kernel cannot guarantee any deterministic execution of the high-prio task: any medium-priority task could preempt the low-prio task while it holds the shared lock and executes the critical section, and could delay it indefinitely. Implementation: --------------- As mentioned before, the userspace fastpath of PI-enabled pthread mutexes involves no kernel work at all - they behave quite similarly to normal futex-based locks: a 0 value means unlocked, and a value==TID means locked. (This is the same method as used by list-based robust futexes.) Userspace uses atomic ops to lock/unlock these mutexes without entering the kernel. To handle the slowpath, we have added two new futex ops: FUTEX_LOCK_PI FUTEX_UNLOCK_PI If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work: if there is no futex-queue attached to the futex address yet then the code looks up the task that owns the futex [it has put its own TID into the futex value], and attaches a 'PI state' structure to the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, kernel-based synchronization object. The 'other' task is made the owner of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex acquired, and it sets the futex value to its own TID and returns. Userspace has no other work to perform - it now owns the lock, and futex value contains FUTEX_WAITERS|TID. If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID -> 0 atomic transition of the futex value], then no kernel work is triggered. If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes up any potential waiters. Note that under this approach, contrary to other PI-futex approaches, there is no prior 'registration' of a PI-futex. [which is not quite possible anyway, due to existing ABI properties of pthread mutexes.] Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties of futexes, and all four combinations are possible: futex, robust-futex, PI-futex, robust+PI-futex. glibc support: -------------- Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes (and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no additional kernel changes are needed for that). [NOTE: The glibc patch is obviously inofficial and unsupported without matching upstream kernel functionality.] the patch-queue and the glibc patch can also be downloaded from: http://redhat.com/~mingo/PI-futex-patches/ Many thanks go to the people who helped us create this kernel feature: Steven Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan van de Ven, Oleg Nesterov and others. Credits for related prior projects goes to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others. Clean up the futex code, before adding more features to it: - use u32 as the futex field type - that's the ABI - use __user and pointers to u32 instead of unsigned long - code style / comment style cleanups - rename hash-bucket name from 'bh' to 'hb'. I checked the pre and post futex.o object files to make sure this patch has no code effects. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Ulrich Drepper <drepper@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 05:54:47 -04:00
val2 = (u32) (unsigned long) utime;
return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
}
[PATCH] VFS: Permit filesystem to override root dentry on mount Extend the get_sb() filesystem operation to take an extra argument that permits the VFS to pass in the target vfsmount that defines the mountpoint. The filesystem is then required to manually set the superblock and root dentry pointers. For most filesystems, this should be done with simple_set_mnt() which will set the superblock pointer and then set the root dentry to the superblock's s_root (as per the old default behaviour). The get_sb() op now returns an integer as there's now no need to return the superblock pointer. This patch permits a superblock to be implicitly shared amongst several mount points, such as can be done with NFS to avoid potential inode aliasing. In such a case, simple_set_mnt() would not be called, and instead the mnt_root and mnt_sb would be set directly. The patch also makes the following changes: (*) the get_sb_*() convenience functions in the core kernel now take a vfsmount pointer argument and return an integer, so most filesystems have to change very little. (*) If one of the convenience function is not used, then get_sb() should normally call simple_set_mnt() to instantiate the vfsmount. This will always return 0, and so can be tail-called from get_sb(). (*) generic_shutdown_super() now calls shrink_dcache_sb() to clean up the dcache upon superblock destruction rather than shrink_dcache_anon(). This is required because the superblock may now have multiple trees that aren't actually bound to s_root, but that still need to be cleaned up. The currently called functions assume that the whole tree is rooted at s_root, and that anonymous dentries are not the roots of trees which results in dentries being left unculled. However, with the way NFS superblock sharing are currently set to be implemented, these assumptions are violated: the root of the filesystem is simply a dummy dentry and inode (the real inode for '/' may well be inaccessible), and all the vfsmounts are rooted on anonymous[*] dentries with child trees. [*] Anonymous until discovered from another tree. (*) The documentation has been adjusted, including the additional bit of changing ext2_* into foo_* in the documentation. [akpm@osdl.org: convert ipath_fs, do other stuff] Signed-off-by: David Howells <dhowells@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Nathan Scott <nathans@sgi.com> Cc: Roland Dreier <rolandd@cisco.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-23 05:02:57 -04:00
static int futexfs_get_sb(struct file_system_type *fs_type,
int flags, const char *dev_name, void *data,
struct vfsmount *mnt)
{
[PATCH] VFS: Permit filesystem to override root dentry on mount Extend the get_sb() filesystem operation to take an extra argument that permits the VFS to pass in the target vfsmount that defines the mountpoint. The filesystem is then required to manually set the superblock and root dentry pointers. For most filesystems, this should be done with simple_set_mnt() which will set the superblock pointer and then set the root dentry to the superblock's s_root (as per the old default behaviour). The get_sb() op now returns an integer as there's now no need to return the superblock pointer. This patch permits a superblock to be implicitly shared amongst several mount points, such as can be done with NFS to avoid potential inode aliasing. In such a case, simple_set_mnt() would not be called, and instead the mnt_root and mnt_sb would be set directly. The patch also makes the following changes: (*) the get_sb_*() convenience functions in the core kernel now take a vfsmount pointer argument and return an integer, so most filesystems have to change very little. (*) If one of the convenience function is not used, then get_sb() should normally call simple_set_mnt() to instantiate the vfsmount. This will always return 0, and so can be tail-called from get_sb(). (*) generic_shutdown_super() now calls shrink_dcache_sb() to clean up the dcache upon superblock destruction rather than shrink_dcache_anon(). This is required because the superblock may now have multiple trees that aren't actually bound to s_root, but that still need to be cleaned up. The currently called functions assume that the whole tree is rooted at s_root, and that anonymous dentries are not the roots of trees which results in dentries being left unculled. However, with the way NFS superblock sharing are currently set to be implemented, these assumptions are violated: the root of the filesystem is simply a dummy dentry and inode (the real inode for '/' may well be inaccessible), and all the vfsmounts are rooted on anonymous[*] dentries with child trees. [*] Anonymous until discovered from another tree. (*) The documentation has been adjusted, including the additional bit of changing ext2_* into foo_* in the documentation. [akpm@osdl.org: convert ipath_fs, do other stuff] Signed-off-by: David Howells <dhowells@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Nathan Scott <nathans@sgi.com> Cc: Roland Dreier <rolandd@cisco.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-23 05:02:57 -04:00
return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
}
static struct file_system_type futex_fs_type = {
.name = "futexfs",
.get_sb = futexfs_get_sb,
.kill_sb = kill_anon_super,
};
static int __init init(void)
{
int i = register_filesystem(&futex_fs_type);
if (i)
return i;
futex_mnt = kern_mount(&futex_fs_type);
if (IS_ERR(futex_mnt)) {
unregister_filesystem(&futex_fs_type);
return PTR_ERR(futex_mnt);
}
for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
spin_lock_init(&futex_queues[i].lock);
}
return 0;
}
__initcall(init);