This patch adds a pv_op for flush_tlb_others. Linux running on native
hardware uses cross-CPU IPIs to flush the TLB on any CPU which may
have a particular mm's pagetable entries cached in its TLB. This is
inefficient in a paravirtualized environment, since the hypervisor
knows which real CPUs actually contain cached mappings, which may be a
small subset of a guest's VCPUs.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Implement the actual patching machinery. paravirt_patch_default()
contains the logic to automatically patch a callsite based on a few
simple rules:
- if the paravirt_op function is paravirt_nop, then patch nops
- if the paravirt_op function is a jmp target, then jmp to it
- if the paravirt_op function is callable and doesn't clobber too much
for the callsite, call it directly
paravirt_patch_default is suitable as a default implementation of
paravirt_ops.patch, will remove most of the expensive indirect calls
in favour of either a direct call or a pile of nops.
Backends may implement their own patcher, however. There are several
helper functions to help with this:
paravirt_patch_nop nop out a callsite
paravirt_patch_ignore leave the callsite as-is
paravirt_patch_call patch a call if the caller and callee
have compatible clobbers
paravirt_patch_jmp patch in a jmp
paravirt_patch_insns patch some literal instructions over
the callsite, if they fit
This patch also implements more direct patches for the native case, so
that when running on native hardware many common operations are
implemented inline.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Zachary Amsden <zach@vmware.com>
Cc: Anthony Liguori <anthony@codemonkey.ws>
Acked-by: Ingo Molnar <mingo@elte.hu>
Clean things up, and broadly document:
- the paravirt_ops functions themselves
- the patching mechanism
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Wrap a set of interesting paravirt_ops calls in a wrapper which makes
the callsites available for patching. Unfortunately this is pretty
ugly because there's no way to get gcc to generate a function call,
but also wrap just the callsite itself with the necessary labels.
This patch supports functions with 0-4 arguments, and either void or
returning a value. 64-bit arguments must be split into a pair of
32-bit arguments (lower word first). Small structures are returned in
registers.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Zachary Amsden <zach@vmware.com>
Cc: Anthony Liguori <anthony@codemonkey.ws>
Fix a few clobbers to include the return register. The clobbers set
is the set of all registers modified (or may be modified) by the code
snippet, regardless of whether it was deliberate or accidental.
Also, make sure that callsites which are used in contexts which don't
allow clobbers actually save and restore all clobberable registers.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Zachary Amsden <zach@vmware.com>
Use patch type identifiers derived from the offset of the operation in
the paravirt_ops structure. This avoids having to maintain a separate
enum for patch site types.
Also, since the identifier is derived from the offset into
paravirt_ops, the offset can be derived from the identifier. This is
used to remove replicated information in the various callsite macros,
which has been a source of bugs in the past.
This patch also drops the fused save_fl+cli operation, which doesn't
really add much and makes things more complex - specifically because
it breaks the 1:1 relationship between identifiers and offsets. If
this operation turns out to be particularly beneficial, then the right
answer is to define a new entrypoint for it.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Zachary Amsden <zach@vmware.com>
Rename struct paravirt_patch to paravirt_patch_site, so that it
clearly refers to a callsite, and not the patch which may be applied
to that callsite.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Zachary Amsden <zach@vmware.com>
Add hooks to allow a paravirt implementation to track the lifetime of
an mm. Paravirtualization requires three hooks, but only two are
needed in common code. They are:
arch_dup_mmap, which is called when a new mmap is created at fork
arch_exit_mmap, which is called when the last process reference to an
mm is dropped, which typically happens on exit and exec.
The third hook is activate_mm, which is called from the arch-specific
activate_mm() macro/function, and so doesn't need stub versions for
other architectures. It's called when an mm is first used.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: linux-arch@vger.kernel.org
Cc: James Bottomley <James.Bottomley@SteelEye.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Normally when running in PAE mode, the 4th PMD maps the kernel address space,
which can be shared among all processes (since they all need the same kernel
mappings).
Xen, however, does not allow guests to have the kernel pmd shared between page
tables, so parameterize pgtable.c to allow both modes of operation.
There are several side-effects of this. One is that vmalloc will update the
kernel address space mappings, and those updates need to be propagated into
all processes if the kernel mappings are not intrinsically shared. In the
non-PAE case, this is done by maintaining a pgd_list of all processes; this
list is used when all process pagetables must be updated. pgd_list is
threaded via otherwise unused entries in the page structure for the pgd, which
means that the pgd must be page-sized for this to work.
Normally the PAE pgd is only 4x64 byte entries large, but Xen requires the PAE
pgd to page aligned anyway, so this patch forces the pgd to be page
aligned+sized when the kernel pmd is unshared, to accomodate both these
requirements.
Also, since there may be several distinct kernel pmds (if the user/kernel
split is below 3G), there's no point in allocating them from a slab cache;
they're just allocated with get_free_page and initialized appropriately. (Of
course the could be cached if there is just a single kernel pmd - which is the
default with a 3G user/kernel split - but it doesn't seem worthwhile to add
yet another case into this code).
[ Many thanks to wli for review comments. ]
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: William Lee Irwin III <wli@holomorphy.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Zachary Amsden <zach@vmware.com>
Cc: Christoph Lameter <clameter@sgi.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Allocate a fixmap slot for use by a paravirt_ops implementation. This
is intended for early-boot bootstrap mappings. Once the zones and
allocator have been set up, it would be better to use get_vm_area() to
allocate some virtual space.
Xen uses this to map the hypervisor's shared info page, which doesn't
have a pseudo-physical page number, and therefore can't be mapped
ordinarily. It is needed early because it contains the vcpu state,
including the interrupt mask.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
This patch introduces paravirt_ops hooks to control how the kernel's
initial pagetable is set up.
In the case of a native boot, the very early bootstrap code creates a
simple non-PAE pagetable to map the kernel and physical memory. When
the VM subsystem is initialized, it creates a proper pagetable which
respects the PAE mode, large pages, etc.
When booting under a hypervisor, there are many possibilities for what
paging environment the hypervisor establishes for the guest kernel, so
the constructon of the kernel's pagetable depends on the hypervisor.
In the case of Xen, the hypervisor boots the kernel with a fully
constructed pagetable, which is already using PAE if necessary. Also,
Xen requires particular care when constructing pagetables to make sure
all pagetables are always mapped read-only.
In order to make this easier, kernel's initial pagetable construction
has been changed to only allocate and initialize a pagetable page if
there's no page already present in the pagetable. This allows the Xen
paravirt backend to make a copy of the hypervisor-provided pagetable,
allowing the kernel to establish any more mappings it needs while
keeping the existing ones.
A slightly subtle point which is worth highlighting here is that Xen
requires all kernel mappings to share the same pte_t pages between all
pagetables, so that updating a kernel page's mapping in one pagetable
is reflected in all other pagetables. This makes it possible to
allocate a page and attach it to a pagetable without having to
explicitly enumerate that page's mapping in all pagetables.
And:
+From: "Eric W. Biederman" <ebiederm@xmission.com>
If we don't set the leaf page table entries it is quite possible that
will inherit and incorrect page table entry from the initial boot
page table setup in head.S. So we need to redo the effort here,
so we pick up PSE, PGE and the like.
Hypervisors like Xen require that their page tables be read-only,
which is slightly incompatible with our low identity mappings, however
I discussed this with Jeremy he has modified the Xen early set_pte
function to avoid problems in this area.
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: William Irwin <bill.irwin@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Add a set of accessors to pack, unpack and modify page table entries
(at all levels). This allows a paravirt implementation to control the
contents of pgd/pmd/pte entries. For example, Xen uses this to
convert the (pseudo-)physical address into a machine address when
populating a pagetable entry, and converting back to pphys address
when an entry is read.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Add a _paravirt_nop function for use as a stub for no-op operations,
and paravirt_nop #defined void * version to make using it easier
(since all its uses are as a void *).
This is useful to allow the patcher to automatically identify noop
operations so it can simply nop out the callsite.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
[mingo] but only as a cleanup of the current open-coded (void *) casts.
My problem with this is that it loses the types. Not that there is much
to check for, but still, this adds some assumptions about how function
calls look like
On Thu, 2007-03-29 at 13:16 +0200, Andi Kleen wrote:
> Please clean it up properly with two structs.
Not sure about this, now I've done it. Running it here.
If you like it, I can do x86-64 as well.
==
lguest defines its own TSS struct because the "struct tss_struct"
contains linux-specific additions. Andi asked me to split the struct
in processor.h.
Unfortunately it makes usage a little awkward.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
The .smp_altinstructions section and its corresponding symbols are
completely unused, so remove them.
Also, remove stray #ifdef __KENREL__ in asm-i386/alternative.h
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
This patch is based on Rusty's recent cleanup of the EFLAGS-related
macros; it extends the same kind of cleanup to control registers and
MSRs.
It also unifies these between i386 and x86-64; at least with regards
to MSRs, the two had definitely gotten out of sync.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Andi Kleen <ak@suse.de>
It doesn't put the CPU into deeper sleep states, so it's better to use the standard
idle loop to save power. But allow to reenable it anyways for benchmarking.
I also removed the obsolete idle=halt on i386
Cc: andreas.herrmann@amd.com
Signed-off-by: Andi Kleen <ak@suse.de>
Most of asm-x86_64/bugs.h is code which should be in a C file, so put it there.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Now that relocation of the VDSO for COMPAT_VDSO users is done at
runtime rather than compile time, it is possible to enable/disable
compat mode at runtime.
This patch allows you to enable COMPAT_VDSO mode with "vdso=2" on the
kernel command line, or via sysctl. (Switching on a running system
shouldn't be done lightly; any process which was relying on the compat
VDSO will be upset if it goes away.)
The COMPAT_VDSO config option still exists, but if enabled it just
makes vdso_enabled default to VDSO_COMPAT.
+From: Hugh Dickins <hugh@veritas.com>
Fix oops from i386-make-compat_vdso-runtime-selectable.patch.
Even mingetty at system startup finds it easy to trigger an oops
while reading /proc/PID/maps: though it has a good hold on the mm
itself, that cannot stop exit_mm() from resetting tsk->mm to NULL.
(It is usually show_map()'s call to get_gate_vma() which oopses,
and I expect we could change that to check priv->tail_vma instead;
but no matter, even m_start()'s call just after get_task_mm() is racy.)
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Zachary Amsden <zach@vmware.com>
Cc: "Jan Beulich" <JBeulich@novell.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Andi Kleen <ak@suse.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Roland McGrath <roland@redhat.com>
Some versions of libc can't deal with a VDSO which doesn't have its
ELF headers matching its mapped address. COMPAT_VDSO maps the VDSO at
a specific system-wide fixed address. Previously this was all done at
build time, on the grounds that the fixed VDSO address is always at
the top of the address space. However, a hypervisor may reserve some
of that address space, pushing the fixmap address down.
This patch does the adjustment dynamically at runtime, depending on
the runtime location of the VDSO fixmap.
[ Patch has been through several hands: Jan Beulich wrote the orignal
version; Zach reworked it, and Jeremy converted it to relocate phdrs
as well as sections. ]
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Zachary Amsden <zach@vmware.com>
Cc: "Jan Beulich" <JBeulich@novell.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Andi Kleen <ak@suse.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Roland McGrath <roland@redhat.com>
identify_cpu() is used to identify both the boot CPU and secondary
CPUs, but it performs some actions which only apply to the boot CPU.
Those functions are therefore really __init functions, but because
they're called by identify_cpu(), they must be marked __cpuinit.
This patch splits identify_cpu() into identify_boot_cpu() and
identify_secondary_cpu(), and calls the appropriate init functions
from each. Also, identify_boot_cpu() and all the functions it
dominates are marked __init.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Most of asm-i386/bugs.h is code which should be in a C file, so put it there.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
As per i386 patch: move X86_EFLAGS_IF et al out to a new header:
processor-flags.h, so we can include it from irqflags.h and use it in
raw_irqs_disabled_flags().
As a side-effect, we could now use these flags in .S files.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
Under CONFIG_DISCONTIGMEM, assuming that a !pfn_valid() implies all
subsequent pfn-s are also invalid is wrong. Thus replace this by
explicitly checking against the E820 map.
AK: make e820 on x86-64 not initdata
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Mark Langsdorf <mark.langsdorf@amd.com>
Rather than using a single constant PERCPU_ENOUGH_ROOM, compute it as
the sum of kernel_percpu + PERCPU_MODULE_RESERVE. This is now common
to all architectures; if an architecture wants to set
PERCPU_ENOUGH_ROOM to something special, then it may do so (ia64 is
the only one which does).
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Andi Kleen <ak@suse.de>
machine_ops is an interface for the machine_* functions defined in
<linux/reboot.h>. This is intended to allow hypervisors to intercept
the reboot process, but it could be used to implement other x86
subarchtecture reboots.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Add a smp_ops interface. This abstracts the API defined by
<linux/smp.h> for use within arch/i386. The primary intent is that it
be used by a paravirtualizing hypervisor to implement SMP, but it
could also be used by non-APIC-using sub-architectures.
This is related to CONFIG_PARAVIRT, but is implemented unconditionally
since it is simpler that way and not a highly performance-sensitive
interface.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Now we have an explicit per-cpu GDT variable, we don't need to keep the
descriptors around to use them to find the GDT: expose cpu_gdt directly.
We could go further and make load_gdt() pack the descriptor for us, or even
assume it means "load the current cpu's GDT" which is what it always does.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
- there's no reason for duplicating the prototype from
include/linux/syscalls.h in include/asm-x86_64/unistd.h
- every file should #include the headers containing the prototypes for
it's global functions
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Andi Kleen <ak@suse.de>
x86_64 currently simulates a list using the index and private fields of the
page struct. Seems that the code was inherited from i386. But x86_64 does
not use the slab to allocate pgds and pmds etc. So the lru field is not
used by the slab and therefore available.
This patch uses standard list operations on page->lru to realize pgd
tracking.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Move X86_EFLAGS_IF et al out to a new header: processor-flags.h, so we
can include it from irqflags.h and use it in raw_irqs_disabled_flags().
As a side-effect, we could now use these flags in .S files.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
On x86-64, kernel memory freed after init can be entirely unmapped instead
of just getting 'poisoned' by overwriting with a debug pattern.
On i386 and x86-64 (under CONFIG_DEBUG_RODATA), kernel text and bug table
can also be write-protected.
Compared to the first version, this one prevents re-creating deleted
mappings in the kernel image range on x86-64, if those got removed
previously. This, together with the original changes, prevents temporarily
having inconsistent mappings when cacheability attributes are being
changed on such pages (e.g. from AGP code). While on i386 such duplicate
mappings don't exist, the same change is done there, too, both for
consistency and because checking pte_present() before using various other
pte_XXX functions is a requirement anyway. At once, i386 code gets
adjusted to use pte_huge() instead of open coding this.
AK: split out cpa() changes
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Fix various broken corner cases in i386 and x86-64 change_page_attr.
AK: split off from tighten kernel image access rights
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andi Kleen <ak@suse.de>
paravirt.c used to implement native versions of all low-level
functions. Far cleaner is to have the native versions exposed in the
headers and as inline native_XXX, and if !CONFIG_PARAVIRT, then simply
#define XXX native_XXX.
There are several nice side effects:
1) write_dt_entry() now takes the correct "struct Xgt_desc_struct *"
not "void *".
2) load_TLS is reintroduced to the for loop, not manually unrolled
with a #error in case the bounds ever change.
3) Macros become inlines, with type checking.
4) Access to the native versions is trivial for KVM, lguest, Xen and
others who might want it.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@muc.de>
Cc: Avi Kivity <avi@qumranet.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
We now have cpu_init() and secondary_cpu_init() doing nothing but calling
_cpu_init() with the same arguments. Rename _cpu_init() to cpu_init() and use
it as a replcement for secondary_cpu_init().
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Now we are no longer dynamically allocating the GDT, we don't need the
"cpu_gdt_table" at all: we can switch straight from "boot_gdt_table" to the
per-cpu GDT. This means initializing the cpu_gdt array in C.
The boot CPU uses the per-cpu var directly, then in smp_prepare_cpus() it
switches to the per-cpu copy just allocated. For secondary CPUs, the
early_gdt_descr is set to point directly to their per-cpu copy.
For UP the code is very simple: it keeps using the "per-cpu" GDT as per SMP,
but we never have to move.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Allocating PDA and GDT at boot is a pain. Using simple per-cpu variables adds
happiness (although we need the GDT page-aligned for Xen, which we do in a
followup patch).
[akpm@linux-foundation.org: build fix]
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The specific case I am encountering is kdump under Xen with a 64 bit
hypervisor and 32 bit kernel/userspace. The dump created is 64 bit due to
the hypervisor but the dump kernel is 32 bit for maximum compatibility.
It's possibly less likely to be useful in a purely native scenario but I
see no reason to disallow it.
[akpm@linux-foundation.org: build fix]
Signed-off-by: Ian Campbell <ian.campbell@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Vivek Goyal <vgoyal@in.ibm.com>
Cc: Horms <horms@verge.net.au>
Cc: Magnus Damm <magnus.damm@gmail.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
GCC (4.1 at least) unrolls it anyway, but I can't believe this code
was ever justifiable. (I've also submitted a patch which cleans up
i386, which is even uglier).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Whenever we schedule, __switch_to calls load_esp0 which does:
tss->esp0 = thread->esp0;
This is never initialized for the initial thread (ie "swapper"), so when we're
scheduling that, we end up setting esp0 to 0. This is fine: the swapper never
leaves ring 0, so this field is never used.
lguest, however, gets upset that we're trying to used an unmapped page as our
kernel stack. Rather than work around it there, let's initialize it.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Extends the numa=fake x86_64 command-line option to allow for configurable
node sizes. These nodes can be used in conjunction with cpusets for coarse
memory resource management.
The old command-line option is still supported:
numa=fake=32 gives 32 fake NUMA nodes, ignoring the NUMA setup of the
actual machine.
But now you may configure your system for the node sizes of your choice:
numa=fake=2*512,1024,2*256
gives two 512M nodes, one 1024M node, two 256M nodes, and
the rest of system memory to a sixth node.
The existing hash function is maintained to support the various node sizes
that are possible with this implementation.
Each node of the same size receives roughly the same amount of available
pages, regardless of any reserved memory with its address range. The total
available pages on the system is calculated and divided by the number of equal
nodes to allocate. These nodes are then dynamically allocated and their
borders extended until such time as their number of available pages reaches
the required size.
Configurable node sizes are recommended when used in conjunction with cpusets
for memory control because it eliminates the overhead associated with scanning
the zonelists of many smaller full nodes on page_alloc().
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Change mark_tsc_unstable() so it takes a string argument, which holds the
reason the TSC was marked unstable.
This is then displayed the first time mark_tsc_unstable is called.
This should help us better debug why the TSC was marked unstable on certain
systems and allow us to make sure we're not being overly paranoid when
throwing out this troublesome clocksource.
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Andi Kleen <ak@suse.de>
o Modpost generates warnings for i386 if compiled with CONFIG_RELOCATABLE=y
WARNING: vmlinux - Section mismatch: reference to .init.text:find_unisys_acpi_oem_table from .text between 'acpi_madt_oem_check' (at offset 0xc0101eda) and 'enable_apic_mode'
WARNING: vmlinux - Section mismatch: reference to .init.text:acpi_get_table_header_early from .text between 'acpi_madt_oem_check' (at offset 0xc0101ef0) and 'enable_apic_mode'
WARNING: vmlinux - Section mismatch: reference to .init.text:parse_unisys_oem from .text between 'acpi_madt_oem_check' (at offset 0xc0101f2e) and 'enable_apic_mode'
WARNING: vmlinux - Section mismatch: reference to .init.text:setup_unisys from .text between 'acpi_madt_oem_check' (at offset 0xc0101f37) and 'enable_apic_mode'WARNING: vmlinux - Section mismatch: reference to .init.text:parse_unisys_oem from .text between 'mps_oem_check' (at offset 0xc0101ec7) and 'acpi_madt_oem_check'
WARNING: vmlinux - Section mismatch: reference to .init.text:es7000_sw_apic from .text between 'enable_apic_mode' (at offset 0xc0101f48) and 'check_apicid_present'
o Some functions which are inline (acpi_madt_oem_check) are not inlined by
compiler as these functions are accessed using function pointer. These
functions are put in .text section and they in-turn access __init type
functions hence modpost generates warnings.
o Do not iniline acpi_madt_oem_check, instead make it __init.
Signed-off-by: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Enable system hashtable memory to be distributed among nodes on x86_64 NUMA
Forcing the kernel to use node interleaved vmalloc instead of bootmem for
the system hashtable memory (alloc_large_system_hash) reduces the memory
imbalance on node 0 by around 40MB on a 8 node x86_64 NUMA box:
Before the following patch, on bootup of a 8 node box:
Node 0 MemTotal: 3407488 kB
Node 0 MemFree: 3206296 kB
Node 0 MemUsed: 201192 kB
Node 0 Active: 7012 kB
Node 0 Inactive: 512 kB
Node 0 Dirty: 0 kB
Node 0 Writeback: 0 kB
Node 0 FilePages: 1912 kB
Node 0 Mapped: 420 kB
Node 0 AnonPages: 5612 kB
Node 0 PageTables: 468 kB
Node 0 NFS_Unstable: 0 kB
Node 0 Bounce: 0 kB
Node 0 Slab: 5408 kB
Node 0 SReclaimable: 644 kB
Node 0 SUnreclaim: 4764 kB
After the patch (or using hashdist=1 on the kernel command line):
Node 0 MemTotal: 3407488 kB
Node 0 MemFree: 3247608 kB
Node 0 MemUsed: 159880 kB
Node 0 Active: 3012 kB
Node 0 Inactive: 616 kB
Node 0 Dirty: 0 kB
Node 0 Writeback: 0 kB
Node 0 FilePages: 2424 kB
Node 0 Mapped: 380 kB
Node 0 AnonPages: 1200 kB
Node 0 PageTables: 396 kB
Node 0 NFS_Unstable: 0 kB
Node 0 Bounce: 0 kB
Node 0 Slab: 6304 kB
Node 0 SReclaimable: 1596 kB
Node 0 SUnreclaim: 4708 kB
I guess it is a good idea to keep HASHDIST_DEFAULT "on" for x86_64 NUMA
since x86_64 has no dearth of vmalloc space? Or maybe enable hash
distribution for all 64bit NUMA arches? The following patch does it only
for x86_64.
I ran a HPC MPI benchmark -- 'Ansys wingsolid', which takes up quite a bit of
memory and uses up tlb entries. This was on a 4 way, 2 socket
Tyan AMD box (non vsmp), with 8G total memory (4G pernode).
The results with and without hash distribution are:
1. Vanilla - runtime of 1188.000s
2. With hashdist=1 runtime of 1154.000s
Oprofile output for the duration of run is:
1. Vanilla:
PU: AMD64 processors, speed 2411.16 MHz (estimated)
Counted L1_AND_L2_DTLB_MISSES events (L1 and L2 DTLB misses) with a unit
mask of 0x00 (No unit mask) count 500
samples % app name symbol name
163054 6.5513 libansys1.so MultiFront::decompose(int, int,
Elemset *, int *, int, int, int)
162061 6.5114 libansys3.so blockSaxpy6L_fd
162042 6.5107 libansys3.so blockInnerProduct6L_fd
156286 6.2794 libansys3.so maxb33_
87879 3.5309 libansys1.so elmatrixmultpcg_
84857 3.4095 libansys4.so saxpy_pcg
58637 2.3560 libansys4.so .st4560
46612 1.8728 libansys4.so .st4282
43043 1.7294 vmlinux-t copy_user_generic_string
41326 1.6604 libansys3.so blockSaxpyBackSolve6L_fd
41288 1.6589 libansys3.so blockInnerProductBackSolve6L_fd
2. With hashdist=1
CPU: AMD64 processors, speed 2411.13 MHz (estimated)
Counted L1_AND_L2_DTLB_MISSES events (L1 and L2 DTLB misses) with a unit
mask of 0x00 (No unit mask) count 500
samples % app name symbol name
162993 6.9814 libansys1.so MultiFront::decompose(int, int,
Elemset *, int *, int, int, int)
160799 6.8874 libansys3.so blockInnerProduct6L_fd
160459 6.8729 libansys3.so blockSaxpy6L_fd
156018 6.6826 libansys3.so maxb33_
84700 3.6279 libansys4.so saxpy_pcg
83434 3.5737 libansys1.so elmatrixmultpcg_
58074 2.4875 libansys4.so .st4560
46000 1.9703 libansys4.so .st4282
41166 1.7632 libansys3.so blockSaxpyBackSolve6L_fd
41033 1.7575 libansys3.so blockInnerProductBackSolve6L_fd
35762 1.5318 libansys1.so inner_product_sub
35591 1.5245 libansys1.so inner_product_sub2
28259 1.2104 libansys4.so addVectors
Signed-off-by: Pravin B. Shelar <pravin.shelar@calsoftinc.com>
Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Christoph Lameter <clameter@engr.sgi.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Fix for the following patch. Provide dummy cpufreq functions when
CPUFREQ is not compiled in.
Cc: Andi Kleen <ak@suse.de>
Cc: Dave Jones <davej@codemonkey.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Andi Kleen <ak@suse.de>
o X86_64 kernel should run from 2MB aligned address for two reasons.
- Performance.
- For relocatable kernels, page tables are updated based on difference
between compile time address and load time physical address.
This difference should be multiple of 2MB as kernel text and data
is mapped using 2MB pages and PMD should be pointing to a 2MB
aligned address. Life is simpler if both compile time and load time
kernel addresses are 2MB aligned.
o Flag the error at compile time if one is trying to build a kernel which
does not meet alignment restrictions.
Signed-off-by: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This patch modifies the x86_64 kernel so that it can be loaded and run
at any 2M aligned address, below 512G. The technique used is to
compile the decompressor with -fPIC and modify it so the decompressor
is fully relocatable. For the main kernel the page tables are
modified so the kernel remains at the same virtual address. In
addition a variable phys_base is kept that holds the physical address
the kernel is loaded at. __pa_symbol is modified to add that when
we take the address of a kernel symbol.
When loaded with a normal bootloader the decompressor will decompress
the kernel to 2M and it will run there. This both ensures the
relocation code is always working, and makes it easier to use 2M
pages for the kernel and the cpu.
AK: changed to not make RELOCATABLE default in Kconfig
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Currently __pa_symbol is for use with symbols in the kernel address
map and __pa is for use with pointers into the physical memory map.
But the code is implemented so you can usually interchange the two.
__pa which is much more common can be implemented much more cheaply
if it is it doesn't have to worry about any other kernel address
spaces. This is especially true with a relocatable kernel as
__pa_symbol needs to peform an extra variable read to resolve
the address.
There is a third macro that is added for the vsyscall data
__pa_vsymbol for finding the physical addesses of vsyscall pages.
Most of this patch is simply sorting through the references to
__pa or __pa_symbol and using the proper one. A little of
it is continuing to use a physical address when we have it
instead of recalculating it several times.
swapper_pgd is now NULL. leave_mm now uses init_mm.pgd
and init_mm.pgd is initialized at boot (instead of compile time)
to the physmem virtual mapping of init_level4_pgd. The
physical address changed.
Except for the for EMPTY_ZERO page all of the remaining references
to __pa_symbol appear to be during kernel initialization. So this
should reduce the cost of __pa in the common case, even on a relocated
kernel.
As this is technically a semantic change we need to be on the lookout
for anything I missed. But it works for me (tm).
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
With the rewrite of the SMP trampoline and the early page
allocator there is nothing that needs identity mapped pages,
once we start executing C code.
So add zap_identity_mappings into head64.c and remove
zap_low_mappings() from much later in the code. The functions
are subtly different thus the name change.
This also kills boot_level4_pgt which was from an earlier
attempt to move the identity mappings as early as possible,
and is now no longer needed. Essentially I have replaced
boot_level4_pgt with trampoline_level4_pgt in trampoline.S
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
