Introduce an atomic_inc_not_zero operation. Make this a special case of
atomic_add_unless because lockless pagecache actually wants
atomic_inc_not_negativeone due to its offset refcount.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Minor simplification to the sparc64 tlb_flush_mmu: tlb_remove_page
set need_flush only after handling the tlb_fast_mode case, then
tlb_flush_mmu need not consider whether it's tlb_fast_mode.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
I don't know if we ever implemented this, but the only user in any 2.6
tree are the compat ioctls.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The old keyboard driver is gone in 2.6, so the only user left are the
compat ioctls.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The old sound drivers are gone in 2.6, so the only user left are the
compat ioctls.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The TIOCPKT_ macros are defined by all other architectures in asm/ioctls.h
and so does sparc and sparc64, so reomve the duplicates in asm/termios.h.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
It isn't needed any longer, as noted by Hugh Dickins.
We still need the flush routines, due to the one remaining
call site in hugetlb_prefault_arch_hook(). That can be
eliminated at some later point, however.
Signed-off-by: David S. Miller <davem@davemloft.net>
sparc64 is unique among architectures in taking the page_table_lock in
its context switch (well, cris does too, but erroneously, and it's not
yet SMP anyway).
This seems to be a private affair between switch_mm and activate_mm,
using page_table_lock as a per-mm lock, without any relation to its uses
elsewhere. That's fine, but comment it as such; and unlock sooner in
switch_mm, more like in activate_mm (preemption is disabled here).
There is a block of "if (0)"ed code in smp_flush_tlb_pending which would
have liked to rely on the page_table_lock, in switch_mm and elsewhere;
but its comment explains how dup_mmap's flush_tlb_mm defeated it. And
though that could have been changed at any time over the past few years,
now the chance vanishes as we push the page_table_lock downwards, and
perhaps split it per page table page. Just delete that block of code.
Which leaves the mysterious spin_unlock_wait(&oldmm->page_table_lock)
in kernel/fork.c copy_mm. Textual analysis (supported by Nick Piggin)
suggests that the comment was written by DaveM, and that it relates to
the defeated approach in the sparc64 smp_flush_tlb_pending. Just delete
this block too.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Sparc64 changes to track kprobe execution on a per-cpu basis. We now track
the kprobe state machine independently on each cpu using an arch specific
kprobe control block.
Signed-off-by: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Signed-off-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The sys_ptrace boilerplate code (everything outside the big switch
statement for the arch-specific requests) is shared by most architectures.
This patch moves it to kernel/ptrace.c and leaves the arch-specific code as
arch_ptrace.
Some architectures have a too different ptrace so we have to exclude them.
They continue to keep their implementations. For sh64 I had to add a
sh64_ptrace wrapper because it does some initialization on the first call.
For um I removed an ifdefed SUBARCH_PTRACE_SPECIAL block, but
SUBARCH_PTRACE_SPECIAL isn't defined anywhere in the tree.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Paul Mackerras <paulus@samba.org>
Acked-by: Ralf Baechle <ralf@linux-mips.org>
Acked-By: David Howells <dhowells@redhat.com>
Acked-by: Russell King <rmk+kernel@arm.linux.org.uk>
Acked-by: Paul Mundt <lethal@linux-sh.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
__MUTEX_INITIALIZER() has no users, and equates to the more commonly used
DECLARE_MUTEX(), thus making it pretty much redundant. Remove it for good.
Signed-off-by: Arthur Othieno <a.othieno@bluewin.ch>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch removes page_pte_prot and page_pte macros from all
architectures. Some architectures define both, some only page_pte (broken)
and others none. These macros are not used anywhere.
page_pte_prot(page, prot) is identical to mk_pte(page, prot) and
page_pte(page) is identical to page_pte_prot(page, __pgprot(0)).
* The following architectures define both page_pte_prot and page_pte
arm, arm26, ia64, sh64, sparc, sparc64
* The following architectures define only page_pte (broken)
frv, i386, m32r, mips, sh, x86-64
* All other architectures define neither
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
zap_pte_range has been counting the pages it frees in tlb->freed, then
tlb_finish_mmu has used that to update the mm's rss. That got stranger when I
added anon_rss, yet updated it by a different route; and stranger when rss and
anon_rss became mm_counters with special access macros. And it would no
longer be viable if we're relying on page_table_lock to stabilize the
mm_counter, but calling tlb_finish_mmu outside that lock.
Remove the mmu_gather's freed field, let tlb_finish_mmu stick to its own
business, just decrement the rss mm_counter in zap_pte_range (yes, there was
some point to batching the update, and a subsequent patch restores that). And
forget the anal paranoia of first reading the counter to avoid going negative
- if rss does go negative, just fix that bug.
Remove the mmu_gather's flushes and avoided_flushes from arm and arm26: no use
was being made of them. But arm26 alone was actually using the freed, in the
way some others use need_flush: give it a need_flush. arm26 seems to prefer
spaces to tabs here: respect that.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
tlb_is_full_mm? What does that mean? The TLB is full? No, it means that the
mm's last user has gone and the whole mm is being torn down. And it's an
inline function because sparc64 uses a different (slightly better)
"tlb_frozen" name for the flag others call "fullmm".
And now the ptep_get_and_clear_full macro used in zap_pte_range refers
directly to tlb->fullmm, which would be wrong for sparc64. Rather than
correct that, I'd prefer to scrap tlb_is_full_mm altogether, and change
sparc64 to just use the same poor name as everyone else - is that okay?
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
tlb_gather_mmu dates from before kernel preemption was allowed, and uses
smp_processor_id or __get_cpu_var to find its per-cpu mmu_gather. That works
because it's currently only called after getting page_table_lock, which is not
dropped until after the matching tlb_finish_mmu. But don't rely on that, it
will soon change: now disable preemption internally by proper get_cpu_var in
tlb_gather_mmu, put_cpu_var in tlb_finish_mmu.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add sem_is_read/write_locked functions to the read/write semaphores, along the
same lines of the *_is_locked spinlock functions. The swap token tuning patch
uses sem_is_read_locked; sem_is_write_locked is added for completeness.
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The hairy fast allocator in the sparc64 PCI IOMMU code
has a hard limit of 256 pages. Certain devices can
exceed this when performing very large I/Os.
So replace with a more simple allocator, based largely
upon the arch/ppc64/kernel/iommu.c code.
Signed-off-by: David S. Miller <davem@davemloft.net>
All the PCI controller drivers were doing the same thing
setting up the IOMMU software state, put it all in one spot.
Signed-off-by: David S. Miller <davem@davemloft.net>
Delete all of the code working with sp_banks[] and replace
with clean acquisition and sorting of physical memory
parameters from the firmware.
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of doing byte-at-a-time user accesses to figure
out where the fault occurred, read the saved fault_address
from the current thread structure.
For the sake of defensive programming, if the fault_address
does not fall into the user buffer range, simply assume the
whole area faulted. This will cause the fixup for
copy_from_user() to clear the entire kernel side buffer.
Signed-off-by: David S. Miller <davem@davemloft.net>
We were not calling kernel_mna_trap_fault() correctly.
Instead of being fancy, just return 0 vs. -EFAULT from
the assembler stubs, and handle that return value as
appropriate.
Create an "__retl_efault" stub for assembler exception
table entries and use it where possible.
Signed-off-by: David S. Miller <davem@davemloft.net>
The funny "range" exception table entries we had were only
used by the compat layer socketcall assembly, and it wasn't
even needed there.
For free we now get proper exception table sorting and fast
binary searching.
Signed-off-by: David S. Miller <davem@davemloft.net>
Also, the us3_cpufreq driver can work on Ultra-IV and IV+.
They use the SAFARI bus register to control the clock divider
just like Ultra-III and III+ do.
Signed-off-by: David S. Miller <davem@davemloft.net>
At boot time, determine the D-cache, I-cache and E-cache size and
line-size. Use them in cache flushes when appropriate.
This change was motivated by discovering that the D-cache on
UltraSparc-IIIi and later are 64K not 32K, and the flushes done by the
Cheetah error handlers were assuming a 32K size.
There are still some pieces of code that are hard coding things and
will need to be fixed up at some point.
While we're here, fix the D-cache and I-cache parity error handlers
to run with interrupts disabled, and when the trap occurs at trap
level > 1 log the event via a counter displayed in /proc/cpuinfo.
Signed-off-by: David S. Miller <davem@davemloft.net>
The trick is that we do the kernel linear mapping TLB miss starting
with an instruction sequence like this:
ba,pt %xcc, kvmap_load
xor %g2, %g4, %g5
succeeded by an instruction sequence which performs a full page table
walk starting at swapper_pg_dir.
We first take over the trap table from the firmware. Then, using this
constant PTE generation for the linear mapping area above, we build
the kernel page tables for the linear mapping.
After this is setup, we patch that branch above into a "nop", which
will cause TLB misses to fall through to the full page table walk.
With this, the page unmapping for CONFIG_DEBUG_PAGEALLOC is trivial.
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of all of this cpu-specific code to remap the kernel
to the correct location, use portable firmware calls to do
this instead.
What we do now is the following in position independant
assembler:
chosen_node = prom_finddevice("/chosen");
prom_mmu_ihandle_cache = prom_getint(chosen_node, "mmu");
vaddr = 4MB_ALIGN(current_text_addr());
prom_translate(vaddr, &paddr_high, &paddr_low, &mode);
prom_boot_mapping_mode = mode;
prom_boot_mapping_phys_high = paddr_high;
prom_boot_mapping_phys_low = paddr_low;
prom_map(-1, 8 * 1024 * 1024, KERNBASE, paddr_low);
and that replaces the massive amount of by-hand TLB probing and
programming we used to do here.
The new code should also handle properly the case where the kernel
is mapped at the correct address already (think: future kexec
support).
Consequently, the bulk of remap_kernel() dies as does the entirety
of arch/sparc64/prom/map.S
We try to share some strings in the PROM library with the ones used
at bootup, and while we're here mark input strings to oplib.h routines
with "const" when appropriate.
There are many more simplifications now possible. For one thing, we
can consolidate the two copies we now have of a lot of cpu setup code
sitting in head.S and trampoline.S.
This is a significant step towards CONFIG_DEBUG_PAGEALLOC support.
Signed-off-by: David S. Miller <davem@davemloft.net>
As recently done by Russell King for ARM, commit
4732efbeb9 introduces a generic asm/futex.h copied
along most arches, which includes a "-ENOSYS support" to be changed if needed.
However, it includes an unused var (taken from the "real" version) which GCC
warns about.
Remove it from all arches having that file version (i.e. same GIT id).
$ git-diff-tree -r HEAD
and
$ git-ls-tree -r HEAD include/|grep 9feff4ce14
may be more interesting than looking at the patch itself, to make sure I've
just copied the arm header to all other archs having the original dummy version
of this file.
Cc: Jakub Jelinek <jakub@redhat.com>
Signed-off-by: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Arrange the modules, OBP, and vmalloc areas such that a range
verification can be done quite minimally.
Signed-off-by: David S. Miller <davem@davemloft.net>
This showed that arch/sparc64/kernel/ptrace.c was not getting
the define properly, and thus the code protected by this ifdef
was never actually compiled before. So fix that too.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch (written by me and also containing many suggestions of Arjan van
de Ven) does a major cleanup of the spinlock code. It does the following
things:
- consolidates and enhances the spinlock/rwlock debugging code
- simplifies the asm/spinlock.h files
- encapsulates the raw spinlock type and moves generic spinlock
features (such as ->break_lock) into the generic code.
- cleans up the spinlock code hierarchy to get rid of the spaghetti.
Most notably there's now only a single variant of the debugging code,
located in lib/spinlock_debug.c. (previously we had one SMP debugging
variant per architecture, plus a separate generic one for UP builds)
Also, i've enhanced the rwlock debugging facility, it will now track
write-owners. There is new spinlock-owner/CPU-tracking on SMP builds too.
All locks have lockup detection now, which will work for both soft and hard
spin/rwlock lockups.
The arch-level include files now only contain the minimally necessary
subset of the spinlock code - all the rest that can be generalized now
lives in the generic headers:
include/asm-i386/spinlock_types.h | 16
include/asm-x86_64/spinlock_types.h | 16
I have also split up the various spinlock variants into separate files,
making it easier to see which does what. The new layout is:
SMP | UP
----------------------------|-----------------------------------
asm/spinlock_types_smp.h | linux/spinlock_types_up.h
linux/spinlock_types.h | linux/spinlock_types.h
asm/spinlock_smp.h | linux/spinlock_up.h
linux/spinlock_api_smp.h | linux/spinlock_api_up.h
linux/spinlock.h | linux/spinlock.h
/*
* here's the role of the various spinlock/rwlock related include files:
*
* on SMP builds:
*
* asm/spinlock_types.h: contains the raw_spinlock_t/raw_rwlock_t and the
* initializers
*
* linux/spinlock_types.h:
* defines the generic type and initializers
*
* asm/spinlock.h: contains the __raw_spin_*()/etc. lowlevel
* implementations, mostly inline assembly code
*
* (also included on UP-debug builds:)
*
* linux/spinlock_api_smp.h:
* contains the prototypes for the _spin_*() APIs.
*
* linux/spinlock.h: builds the final spin_*() APIs.
*
* on UP builds:
*
* linux/spinlock_type_up.h:
* contains the generic, simplified UP spinlock type.
* (which is an empty structure on non-debug builds)
*
* linux/spinlock_types.h:
* defines the generic type and initializers
*
* linux/spinlock_up.h:
* contains the __raw_spin_*()/etc. version of UP
* builds. (which are NOPs on non-debug, non-preempt
* builds)
*
* (included on UP-non-debug builds:)
*
* linux/spinlock_api_up.h:
* builds the _spin_*() APIs.
*
* linux/spinlock.h: builds the final spin_*() APIs.
*/
All SMP and UP architectures are converted by this patch.
arm, i386, ia64, ppc, ppc64, s390/s390x, x64 was build-tested via
crosscompilers. m32r, mips, sh, sparc, have not been tested yet, but should
be mostly fine.
From: Grant Grundler <grundler@parisc-linux.org>
Booted and lightly tested on a500-44 (64-bit, SMP kernel, dual CPU).
Builds 32-bit SMP kernel (not booted or tested). I did not try to build
non-SMP kernels. That should be trivial to fix up later if necessary.
I converted bit ops atomic_hash lock to raw_spinlock_t. Doing so avoids
some ugly nesting of linux/*.h and asm/*.h files. Those particular locks
are well tested and contained entirely inside arch specific code. I do NOT
expect any new issues to arise with them.
If someone does ever need to use debug/metrics with them, then they will
need to unravel this hairball between spinlocks, atomic ops, and bit ops
that exist only because parisc has exactly one atomic instruction: LDCW
(load and clear word).
From: "Luck, Tony" <tony.luck@intel.com>
ia64 fix
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Arjan van de Ven <arjanv@infradead.org>
Signed-off-by: Grant Grundler <grundler@parisc-linux.org>
Cc: Matthew Wilcox <willy@debian.org>
Signed-off-by: Hirokazu Takata <takata@linux-m32r.org>
Signed-off-by: Mikael Pettersson <mikpe@csd.uu.se>
Signed-off-by: Benoit Boissinot <benoit.boissinot@ens-lyon.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
There were three changes necessary in order to allow
sparc64 to use setup-res.c:
1) Sparc64 roots the PCI I/O and MEM address space using
parent resources contained in the PCI controller structure.
I'm actually surprised no other platforms do this, especially
ones like Alpha and PPC{,64}. These resources get linked into the
iomem/ioport tree when PCI controllers are probed.
So the hierarchy looks like this:
iomem --|
PCI controller 1 MEM space --|
device 1
device 2
etc.
PCI controller 2 MEM space --|
...
ioport --|
PCI controller 1 IO space --|
...
PCI controller 2 IO space --|
...
You get the idea. The drivers/pci/setup-res.c code allocates
using plain iomem_space and ioport_space as the root, so that
wouldn't work with the above setup.
So I added a pcibios_select_root() that is used to handle this.
It uses the PCI controller struct's io_space and mem_space on
sparc64, and io{port,mem}_resource on every other platform to
keep current behavior.
2) quirk_io_region() is buggy. It takes in raw BUS view addresses
and tries to use them as a PCI resource.
pci_claim_resource() expects the resource to be fully formed when
it gets called. The sparc64 implementation would do the translation
but that's absolutely wrong, because if the same resource gets
released then re-claimed we'll adjust things twice.
So I fixed up quirk_io_region() to do the proper pcibios_bus_to_resource()
conversion before passing it on to pci_claim_resource().
3) I was mistakedly __init'ing the function methods the PCI controller
drivers provide on sparc64 to implement some parts of these
routines. This was, of course, easy to fix.
So we end up with the following, and that nasty SPARC64 makefile
ifdef in drivers/pci/Makefile is finally zapped.
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Since GCC has to emit a call and a delay slot to the
out-of-line "membar" routines in arch/sparc64/lib/mb.S
it is much better to just do the necessary predicted
branch inline instead as:
ba,pt %xcc, 1f
membar #whatever
1:
instead of the current:
call membar_foo
dslot
because this way GCC is not required to allocate a stack
frame if the function can be a leaf function.
This also makes this bug fix easier to backport to 2.4.x
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch just gathers together all the struct flock definitions except
xtensa into asm-generic/fcntl.h.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch puts the most popular of each fcntl operation/flag into
asm-generic/fcntl.h and cleans up the arch files.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch puts the most popular of each open flag into asm-generic/fcntl.h
and cleans up the arch files.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This set of patches creates asm-generic/fcntl.h and consolidates as much as
possible from the asm-*/fcntl.h files into it.
This patch just gathers all the identical bits of the asm-*/fcntl.h files into
asm-generic/fcntl.h.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.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>
unused and useless..
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The size of auxiliary vector is fixed at 42 in linux/sched.h. But it isn't
very obvious when looking at linux/elf.h. This patch adds AT_VECTOR_SIZE
so that we can change it if necessary when a new vector is added.
Because of include file ordering problems, doing this necessitated the
extraction of the AT_* symbols into a standalone header file.
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
When I first wrote the compat layer patches, I was somewhat cavalier about
the definition of compat_uid_t and compat_gid_t (or maybe I just
misunderstood :-)). This patch makes the compat types much more consistent
with the types we are being compatible with and hopefully will fix a few
bugs along the way.
compat type type in compat arch
__compat_[ug]id_t __kernel_[ug]id_t
__compat_[ug]id32_t __kernel_[ug]id32_t
compat_[ug]id_t [ug]id_t
The difference is that compat_uid_t is always 32 bits (for the archs we
care about) but __compat_uid_t may be 16 bits on some.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
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>
