In order to allow sleeping during mmu notifier calls, we need to avoid
invoking them under the page table spinlock. This patch solves the
problem by calling invalidate_page notification after releasing the lock
(but before freeing the page itself), or by wrapping the page invalidation
with calls to invalidate_range_begin and invalidate_range_end.
To prevent accidental changes to the invalidate_range_end arguments after
the call to invalidate_range_begin, the patch introduces a convention of
saving the arguments in consistently named locals:
unsigned long mmun_start; /* For mmu_notifiers */
unsigned long mmun_end; /* For mmu_notifiers */
...
mmun_start = ...
mmun_end = ...
mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
...
mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
The patch changes code to use this convention for all calls to
mmu_notifier_invalidate_range_start/end, except those where the calls are
close enough so that anyone who glances at the code can see the values
aren't changing.
This patchset is a preliminary step towards on-demand paging design to be
added to the RDMA stack.
Why do we want on-demand paging for Infiniband?
Applications register memory with an RDMA adapter using system calls,
and subsequently post IO operations that refer to the corresponding
virtual addresses directly to HW. Until now, this was achieved by
pinning the memory during the registration calls. The goal of on demand
paging is to avoid pinning the pages of registered memory regions (MRs).
This will allow users the same flexibility they get when swapping any
other part of their processes address spaces. Instead of requiring the
entire MR to fit in physical memory, we can allow the MR to be larger,
and only fit the current working set in physical memory.
Why should anyone care? What problems are users currently experiencing?
This can make programming with RDMA much simpler. Today, developers
that are working with more data than their RAM can hold need either to
deregister and reregister memory regions throughout their process's
life, or keep a single memory region and copy the data to it. On demand
paging will allow these developers to register a single MR at the
beginning of their process's life, and let the operating system manage
which pages needs to be fetched at a given time. In the future, we
might be able to provide a single memory access key for each process
that would provide the entire process's address as one large memory
region, and the developers wouldn't need to register memory regions at
all.
Is there any prospect that any other subsystems will utilise these
infrastructural changes? If so, which and how, etc?
As for other subsystems, I understand that XPMEM wanted to sleep in
MMU notifiers, as Christoph Lameter wrote at
http://lkml.indiana.edu/hypermail/linux/kernel/0802.1/0460.html and
perhaps Andrea knows about other use cases.
Scheduling in mmu notifications is required since we need to sync the
hardware with the secondary page tables change. A TLB flush of an IO
device is inherently slower than a CPU TLB flush, so our design works by
sending the invalidation request to the device, and waiting for an
interrupt before exiting the mmu notifier handler.
Avi said:
kvm may be a buyer. kvm::mmu_lock, which serializes guest page
faults, also protects long operations such as destroying large ranges.
It would be good to convert it into a spinlock, but as it is used inside
mmu notifiers, this cannot be done.
(there are alternatives, such as keeping the spinlock and using a
generation counter to do the teardown in O(1), which is what the "may"
is doing up there).
[akpm@linux-foundation.orgpossible speed tweak in hugetlb_cow(), cleanups]
Signed-off-by: Andrea Arcangeli <andrea@qumranet.com>
Signed-off-by: Sagi Grimberg <sagig@mellanox.com>
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Or Gerlitz <ogerlitz@mellanox.com>
Cc: Haggai Eran <haggaie@mellanox.com>
Cc: Shachar Raindel <raindel@mellanox.com>
Cc: Liran Liss <liranl@mellanox.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Avi Kivity <avi@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We had thought that pages could no longer get freed while still marked as
mlocked; but Johannes Weiner posted this program to demonstrate that
truncating an mlocked private file mapping containing COWed pages is still
mishandled:
#include <sys/types.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <stdio.h>
int main(void)
{
char *map;
int fd;
system("grep mlockfreed /proc/vmstat");
fd = open("chigurh", O_CREAT|O_EXCL|O_RDWR);
unlink("chigurh");
ftruncate(fd, 4096);
map = mmap(NULL, 4096, PROT_WRITE, MAP_PRIVATE, fd, 0);
map[0] = 11;
mlock(map, sizeof(fd));
ftruncate(fd, 0);
close(fd);
munlock(map, sizeof(fd));
munmap(map, 4096);
system("grep mlockfreed /proc/vmstat");
return 0;
}
The anon COWed pages are not caught by truncation's clear_page_mlock() of
the pagecache pages; but unmap_mapping_range() unmaps them, so we ought to
look out for them there in page_remove_rmap(). Indeed, why should
truncation or invalidation be doing the clear_page_mlock() when removing
from pagecache? mlock is a property of mapping in userspace, not a
property of pagecache: an mlocked unmapped page is nonsensical.
Reported-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ying Han <yinghan@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
page_evictable(page, vma) is an irritant: almost all its callers pass
NULL for vma. Remove the vma arg and use mlocked_vma_newpage(vma, page)
explicitly in the couple of places it's needed. But in those places we
don't even need page_evictable() itself! They're dealing with a freshly
allocated anonymous page, which has no "mapping" and cannot be mlocked yet.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In file and anon rmap, we use interval trees to find potentially relevant
vmas and then call vma_address() to find the virtual address the given
page might be found at in these vmas. vma_address() used to include a
check that the returned address falls within the limits of the vma, but
this check isn't necessary now that we always use interval trees in rmap:
the interval tree just doesn't return any vmas which this check would find
to be irrelevant. As a result, we can replace the use of -EFAULT error
code (which then needed to be checked in every call site) with a
VM_BUG_ON().
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mremap() had a clever optimization where move_ptes() did not take the
anon_vma lock to avoid a race with anon rmap users such as page migration.
Instead, the avc's were ordered in such a way that the origin vma was
always visited by rmap before the destination. This ordering and the use
of page table locks rmap usage safe. However, we want to replace the use
of linked lists in anon rmap with an interval tree, and this will make it
harder to impose such ordering as the interval tree will always be sorted
by the avc->vma->vm_pgoff value. For now, let's replace the
anon_vma_moveto_tail() ordering function with proper anon_vma locking in
move_ptes(). Once we have the anon interval tree in place, we will
re-introduce an optimization to avoid taking these locks in the most
common cases.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Implement an interval tree as a replacement for the VMA prio_tree. The
algorithms are similar to lib/interval_tree.c; however that code can't be
directly reused as the interval endpoints are not explicitly stored in the
VMA. So instead, the common algorithm is moved into a template and the
details (node type, how to get interval endpoints from the node, etc) are
filled in using the C preprocessor.
Once the interval tree functions are available, using them as a
replacement to the VMA prio tree is a relatively simple, mechanical job.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The swap token code no longer fits in with the current VM model. It
does not play well with cgroups or the better NUMA placement code in
development, since we have only one swap token globally.
It also has the potential to mess with scalability of the system, by
increasing the number of non-reclaimable pages on the active and
inactive anon LRU lists.
Last but not least, the swap token code has been broken for a year
without complaints, as reported by Konstantin Khlebnikov. This suggests
we no longer have much use for it.
The days of sub-1G memory systems with heavy use of swap are over. If
we ever need thrashing reducing code in the future, we will have to
implement something that does scale.
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Hugh Dickins <hughd@google.com>
Acked-by: Bob Picco <bpicco@meloft.net>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Now, page-stat-per-memcg is recorded into per page_cgroup flag by
duplicating page's status into the flag. The reason is that memcg has a
feature to move a page from a group to another group and we have race
between "move" and "page stat accounting",
Under current logic, assume CPU-A and CPU-B. CPU-A does "move" and CPU-B
does "page stat accounting".
When CPU-A goes 1st,
CPU-A CPU-B
update "struct page" info.
move_lock_mem_cgroup(memcg)
see pc->flags
copy page stat to new group
overwrite pc->mem_cgroup.
move_unlock_mem_cgroup(memcg)
move_lock_mem_cgroup(mem)
set pc->flags
update page stat accounting
move_unlock_mem_cgroup(mem)
stat accounting is guarded by move_lock_mem_cgroup() and "move" logic
(CPU-A) doesn't see changes in "struct page" information.
But it's costly to have the same information both in 'struct page' and
'struct page_cgroup'. And, there is a potential problem.
For example, assume we have PG_dirty accounting in memcg.
PG_..is a flag for struct page.
PCG_ is a flag for struct page_cgroup.
(This is just an example. The same problem can be found in any
kind of page stat accounting.)
CPU-A CPU-B
TestSet PG_dirty
(delay) TestClear PG_dirty
if (TestClear(PCG_dirty))
memcg->nr_dirty--
if (TestSet(PCG_dirty))
memcg->nr_dirty++
Here, memcg->nr_dirty = +1, this is wrong. This race was reported by Greg
Thelen <gthelen@google.com>. Now, only FILE_MAPPED is supported but
fortunately, it's serialized by page table lock and this is not real bug,
_now_,
If this potential problem is caused by having duplicated information in
struct page and struct page_cgroup, we may be able to fix this by using
original 'struct page' information. But we'll have a problem in "move
account"
Assume we use only PG_dirty.
CPU-A CPU-B
TestSet PG_dirty
(delay) move_lock_mem_cgroup()
if (PageDirty(page))
new_memcg->nr_dirty++
pc->mem_cgroup = new_memcg;
move_unlock_mem_cgroup()
move_lock_mem_cgroup()
memcg = pc->mem_cgroup
new_memcg->nr_dirty++
accounting information may be double-counted. This was original reason to
have PCG_xxx flags but it seems PCG_xxx has another problem.
I think we need a bigger lock as
move_lock_mem_cgroup(page)
TestSetPageDirty(page)
update page stats (without any checks)
move_unlock_mem_cgroup(page)
This fixes both of problems and we don't have to duplicate page flag into
page_cgroup. Please note: move_lock_mem_cgroup() is held only when there
are possibility of "account move" under the system. So, in most path,
status update will go without atomic locks.
This patch introduces mem_cgroup_begin_update_page_stat() and
mem_cgroup_end_update_page_stat() both should be called at modifying
'struct page' information if memcg takes care of it. as
mem_cgroup_begin_update_page_stat()
modify page information
mem_cgroup_update_page_stat()
=> never check any 'struct page' info, just update counters.
mem_cgroup_end_update_page_stat().
This patch is slow because we need to call begin_update_page_stat()/
end_update_page_stat() regardless of accounted will be changed or not. A
following patch adds an easy optimization and reduces the cost.
[akpm@linux-foundation.org: s/lock/locked/]
[hughd@google.com: fix deadlock by avoiding stat lock when anon]
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Greg Thelen <gthelen@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Reduce code duplication by calling anon_vma_chain_link() from
anon_vma_prepare().
Also move anon_vmal_chain_link() to a more suitable location in the file.
Signed-off-by: Kautuk Consul <consul.kautuk@gmail.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Hugh Dickins <hughd@google.com>
Reviewed-by: KAMEZWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mgorman@suse.de>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Since commit 2a11c8ea20 ("kconfig: Introduce IS_ENABLED(),
IS_BUILTIN() and IS_MODULE()") there is a generic grep-friendly method
for checking config options in C expressions.
Signed-off-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
migrate was doing an rmap_walk with speculative lock-less access on
pagetables. That could lead it to not serializing properly against mremap
PT locks. But a second problem remains in the order of vmas in the
same_anon_vma list used by the rmap_walk.
If vma_merge succeeds in copy_vma, the src vma could be placed after the
dst vma in the same_anon_vma list. That could still lead to migrate
missing some pte.
This patch adds an anon_vma_moveto_tail() function to force the dst vma at
the end of the list before mremap starts to solve the problem.
If the mremap is very large and there are a lots of parents or childs
sharing the anon_vma root lock, this should still scale better than taking
the anon_vma root lock around every pte copy practically for the whole
duration of mremap.
Update: Hugh noticed special care is needed in the error path where
move_page_tables goes in the reverse direction, a second
anon_vma_moveto_tail() call is needed in the error path.
This program exercises the anon_vma_moveto_tail:
===
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
printf("%p\n", p);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
if (memcmp(p, p2, SIZE))
printf("error\n");
p4 = mremap(p+SIZE/2, SIZE/2, SIZE/2, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
if (p4 != p3)
perror("mremap"), exit(1);
p4 = mremap(p4, SIZE/2, SIZE/2, MREMAP_FIXED|MREMAP_MAYMOVE, p+SIZE/2);
if (p4 != p+SIZE/2)
perror("mremap"), exit(1);
if (memcmp(p, p2, SIZE))
printf("error\n");
printf("ok\n");
return 0;
}
===
$ perf probe -a anon_vma_moveto_tail
Add new event:
probe:anon_vma_moveto_tail (on anon_vma_moveto_tail)
You can now use it on all perf tools, such as:
perf record -e probe:anon_vma_moveto_tail -aR sleep 1
$ perf record -e probe:anon_vma_moveto_tail -aR ./anon_vma_moveto_tail
0x7f2ca2800000
ok
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.043 MB perf.data (~1860 samples) ]
$ perf report --stdio
100.00% anon_vma_moveto [kernel.kallsyms] [k] anon_vma_moveto_tail
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Nai Xia <nai.xia@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Pawel Sikora <pluto@agmk.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'modsplit-Oct31_2011' of git://git.kernel.org/pub/scm/linux/kernel/git/paulg/linux: (230 commits)
Revert "tracing: Include module.h in define_trace.h"
irq: don't put module.h into irq.h for tracking irqgen modules.
bluetooth: macroize two small inlines to avoid module.h
ip_vs.h: fix implicit use of module_get/module_put from module.h
nf_conntrack.h: fix up fallout from implicit moduleparam.h presence
include: replace linux/module.h with "struct module" wherever possible
include: convert various register fcns to macros to avoid include chaining
crypto.h: remove unused crypto_tfm_alg_modname() inline
uwb.h: fix implicit use of asm/page.h for PAGE_SIZE
pm_runtime.h: explicitly requires notifier.h
linux/dmaengine.h: fix implicit use of bitmap.h and asm/page.h
miscdevice.h: fix up implicit use of lists and types
stop_machine.h: fix implicit use of smp.h for smp_processor_id
of: fix implicit use of errno.h in include/linux/of.h
of_platform.h: delete needless include <linux/module.h>
acpi: remove module.h include from platform/aclinux.h
miscdevice.h: delete unnecessary inclusion of module.h
device_cgroup.h: delete needless include <linux/module.h>
net: sch_generic remove redundant use of <linux/module.h>
net: inet_timewait_sock doesnt need <linux/module.h>
...
Fix up trivial conflicts (other header files, and removal of the ab3550 mfd driver) in
- drivers/media/dvb/frontends/dibx000_common.c
- drivers/media/video/{mt9m111.c,ov6650.c}
- drivers/mfd/ab3550-core.c
- include/linux/dmaengine.h
The files changed within are only using the EXPORT_SYMBOL
macro variants. They are not using core modular infrastructure
and hence don't need module.h but only the export.h header.
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
On x86 a page without a mapper is by definition not referenced / old.
The s390 architecture keeps the reference bit in the storage key and
the current code will check the storage key for page without a mapper.
This leads to an interesting effect: the first time an s390 system
needs to write pages to swap it only finds referenced pages. This
causes a lot of pages to get added and written to the swap device.
To avoid this behaviour change page_referenced to query the storage
key only if there is a mapper of the page.
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
i_alloc_sem is a rather special rw_semaphore. It's the last one that may
be released by a non-owner, and it's write side is always mirrored by
real exclusion. It's intended use it to wait for all pending direct I/O
requests to finish before starting a truncate.
Replace it with a hand-grown construct:
- exclusion for truncates is already guaranteed by i_mutex, so it can
simply fall way
- the reader side is replaced by an i_dio_count member in struct inode
that counts the number of pending direct I/O requests. Truncate can't
proceed as long as it's non-zero
- when i_dio_count reaches non-zero we wake up a pending truncate using
wake_up_bit on a new bit in i_flags
- new references to i_dio_count can't appear while we are waiting for
it to read zero because the direct I/O count always needs i_mutex
(or an equivalent like XFS's i_iolock) for starting a new operation.
This scheme is much simpler, and saves the space of a spinlock_t and a
struct list_head in struct inode (typically 160 bits on a non-debug 64-bit
system).
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
We cannot take a mutex while holding a spinlock, so flip the order and
fix the locking documentation.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Hugh Dickins points out that lockdep (correctly) spots a potential
deadlock on the anon_vma lock, because we now do a GFP_KERNEL allocation
of anon_vma_chain while doing anon_vma_clone(). The problem is that
page reclaim will want to take the anon_vma lock of any anonymous pages
that it will try to reclaim.
So re-organize the code in anon_vma_clone() slightly: first do just a
GFP_NOWAIT allocation, which will usually work fine. But if that fails,
let's just drop the lock and re-do the allocation, now with GFP_KERNEL.
End result: not only do we avoid the locking problem, this also ends up
getting better concurrency in case the allocation does need to block.
Tim Chen reports that with all these anon_vma locking tweaks, we're now
almost back up to the spinlock performance.
Reported-and-tested-by: Hugh Dickins <hughd@google.com>
Tested-by: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This matches the anon_vma_clone() case, and uses the same lock helper
functions. Because of the need to potentially release the anon_vma's,
it's a bit more complex, though.
We traverse the 'vma->anon_vma_chain' in two phases: the first loop gets
the anon_vma lock (with the helper function that only takes the lock
once for the whole loop), and removes any entries that don't need any
more processing.
The second phase just traverses the remaining list entries (without
holding the anon_vma lock), and does any actual freeing of the
anon_vma's that is required.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Tested-by: Hugh Dickins <hughd@google.com>
Tested-by: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In anon_vma_clone() we traverse the vma->anon_vma_chain of the source
vma, locking the anon_vma for each entry.
But they are all going to have the same root entry, which means that
we're locking and unlocking the same lock over and over again. Which is
expensive in locked operations, but can get _really_ expensive when that
root entry sees any kind of lock contention.
In fact, Tim Chen reports a big performance regression due to this: when
we switched to use a mutex instead of a spinlock, the contention case
gets much worse.
So to alleviate this all, this commit creates a small helper function
(lock_anon_vma_root()) that can be used to take the lock just once
rather than taking and releasing it over and over again.
We still have the same "take the lock and release" it behavior in the
exit path (in unlink_anon_vmas()), but that one is a bit harder to fix
since we're actually freeing the anon_vma entries as we go, and that
will touch the lock too.
Reported-and-tested-by: Tim Chen <tim.c.chen@linux.intel.com>
Tested-by: Hugh Dickins <hughd@google.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Split the global inode_wb_list_lock into a per-bdi_writeback list_lock,
as it's currently the most contended lock in the system for metadata
heavy workloads. It won't help for single-filesystem workloads for
which we'll need the I/O-less balance_dirty_pages, but at least we
can dedicate a cpu to spinning on each bdi now for larger systems.
Based on earlier patches from Nick Piggin and Dave Chinner.
It reduces lock contentions to 1/4 in this test case:
10 HDD JBOD, 100 dd on each disk, XFS, 6GB ram
lock_stat version 0.3
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
class name con-bounces contentions waittime-min waittime-max waittime-total acq-bounces acquisitions holdtime-min holdtime-max holdtime-total
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
vanilla 2.6.39-rc3:
inode_wb_list_lock: 42590 44433 0.12 147.74 144127.35 252274 886792 0.08 121.34 917211.23
------------------
inode_wb_list_lock 2 [<ffffffff81165da5>] bdev_inode_switch_bdi+0x29/0x85
inode_wb_list_lock 34 [<ffffffff8115bd0b>] inode_wb_list_del+0x22/0x49
inode_wb_list_lock 12893 [<ffffffff8115bb53>] __mark_inode_dirty+0x170/0x1d0
inode_wb_list_lock 10702 [<ffffffff8115afef>] writeback_single_inode+0x16d/0x20a
------------------
inode_wb_list_lock 2 [<ffffffff81165da5>] bdev_inode_switch_bdi+0x29/0x85
inode_wb_list_lock 19 [<ffffffff8115bd0b>] inode_wb_list_del+0x22/0x49
inode_wb_list_lock 5550 [<ffffffff8115bb53>] __mark_inode_dirty+0x170/0x1d0
inode_wb_list_lock 8511 [<ffffffff8115b4ad>] writeback_sb_inodes+0x10f/0x157
2.6.39-rc3 + patch:
&(&wb->list_lock)->rlock: 11383 11657 0.14 151.69 40429.51 90825 527918 0.11 145.90 556843.37
------------------------
&(&wb->list_lock)->rlock 10 [<ffffffff8115b189>] inode_wb_list_del+0x5f/0x86
&(&wb->list_lock)->rlock 1493 [<ffffffff8115b1ed>] writeback_inodes_wb+0x3d/0x150
&(&wb->list_lock)->rlock 3652 [<ffffffff8115a8e9>] writeback_sb_inodes+0x123/0x16f
&(&wb->list_lock)->rlock 1412 [<ffffffff8115a38e>] writeback_single_inode+0x17f/0x223
------------------------
&(&wb->list_lock)->rlock 3 [<ffffffff8110b5af>] bdi_lock_two+0x46/0x4b
&(&wb->list_lock)->rlock 6 [<ffffffff8115b189>] inode_wb_list_del+0x5f/0x86
&(&wb->list_lock)->rlock 2061 [<ffffffff8115af97>] __mark_inode_dirty+0x173/0x1cf
&(&wb->list_lock)->rlock 2629 [<ffffffff8115a8e9>] writeback_sb_inodes+0x123/0x16f
hughd@google.com: fix recursive lock when bdi_lock_two() is called with new the same as old
akpm@linux-foundation.org: cleanup bdev_inode_switch_bdi() comment
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Inspired by an analysis from Hugh on why again all this doesn't explode
in our face.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
On one machine I've been getting hangs, a page fault's anon_vma_prepare()
waiting in anon_vma_lock(), other processes waiting for that page's lock.
This is a replay of last year's f18194275c "mm: fix hang on
anon_vma->root->lock".
The new page_lock_anon_vma() places too much faith in its refcount: when
it has acquired the mutex_trylock(), it's possible that a racing task in
anon_vma_alloc() has just reallocated the struct anon_vma, set refcount
to 1, and is about to reset its anon_vma->root.
Fix this by saving anon_vma->root, and relying on the usual page_mapped()
check instead of a refcount check: if page is still mapped, the anon_vma
is still ours; if page is not still mapped, we're no longer interested.
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I've hit the "address >= vma->vm_end" check in do_page_add_anon_rmap()
just once. The stack showed khugepaged allocation trying to compact
pages: the call to page_add_anon_rmap() coming from remove_migration_pte().
That path holds anon_vma lock, but does not hold mmap_sem: it can
therefore race with a split_vma(), and in commit 5f70b962cc "mmap:
avoid unnecessary anon_vma lock" we just took away the anon_vma lock
protection when adjusting vma->vm_end.
I don't think that particular BUG_ON ever caught anything interesting,
so better replace it by a comment, than reinstate the anon_vma locking.
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Optimize the page_lock_anon_vma() fast path to be one atomic op, instead
of two.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Straightforward conversion of anon_vma->lock to a mutex.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Hugh Dickins <hughd@google.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Tony Luck <tony.luck@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Convert page_lock_anon_vma() over to use refcounts. This is done to
prepare for the conversion of anon_vma from spinlock to mutex.
Sadly this inceases the cost of page_lock_anon_vma() from one to two
atomics, a follow up patch addresses this, lets keep that simple for now.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
A slightly more verbose comment to go along with the trickery in
page_lock_anon_vma().
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Its beyond ugly and gets in the way.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Tony Luck <tony.luck@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Namhyung Kim <namhyung@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Straightforward conversion of i_mmap_lock to a mutex.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Tony Luck <tony.luck@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The page_clear_dirty primitive always sets the default storage key
which resets the access control bits and the fetch protection bit.
That will surprise a KVM guest that sets non-zero access control
bits or the fetch protection bit. Merge page_test_dirty and
page_clear_dirty back to a single function and only clear the
dirty bit from the storage key.
In addition move the function page_test_and_clear_dirty and
page_test_and_clear_young to page.h where they belong. This
requires to change the parameter from a struct page * to a page
frame number.
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Protect the inode writeback list with a new global lock
inode_wb_list_lock and use it to protect the list manipulations and
traversals. This lock replaces the inode_lock as the inodes on the
list can be validity checked while holding the inode->i_lock and
hence the inode_lock is no longer needed to protect the list.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Protect inode state transitions and validity checks with the
inode->i_lock. This enables us to make inode state transitions
independently of the inode_lock and is the first step to peeling
away the inode_lock from the code.
This requires that __iget() is done atomically with i_state checks
during list traversals so that we don't race with another thread
marking the inode I_FREEING between the state check and grabbing the
reference.
Also remove the unlock_new_inode() memory barrier optimisation
required to avoid taking the inode_lock when clearing I_NEW.
Simplify the code by simply taking the inode->i_lock around the
state change and wakeup. Because the wakeup is no longer tricky,
remove the wake_up_inode() function and open code the wakeup where
necessary.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This patch changes the anon_vma refcount to be 0 when the object is free.
It does this by adding 1 ref to being in use in the anon_vma structure
(iow. the anon_vma->head list is not empty).
This allows a simpler release scheme without having to check both the
refcount and the list as well as avoids taking a ref for each entry on the
list.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The normal code pattern used in the kernel is: get/put.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When vmscan.c calls page_referenced(), if an anon page was created
before a process forked, rmap will search for it in both of the
processes, even though one of them might have since broken COW.
If the child process mlocks the vma where the COWed page belongs to,
page_referenced() running on the page mapped by the parent would lead to
*vm_flags getting VM_LOCKED set erroneously (leading to the references
on the parent page being ignored and evicting the parent page too
early).
*mapcount would also be decremented by page_referenced_one even if the
page wasn't found by page_check_address.
This also lets pmdp_clear_flush_young_notify() go ahead on a
pmd_trans_splitting() pmd.
We hold the page_table_lock so __split_huge_page_map() must wait the
pmdp_clear_flush_young_notify() to complete before it can modify the
pmd. The pmd is also still mapped in userland so the young bit may
materialize through a tlb miss before split_huge_page_map runs.
This will provide a more accurate page_referenced() behavior during
split_huge_page().
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Michel Lespinasse <walken@google.com>
Reviewed-by: Michel Lespinasse <walken@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel<riel@redhat.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Replace usage of the mem_cgroup_update_file_mapped() memcg
statistic update routine with two new routines:
* mem_cgroup_inc_page_stat()
* mem_cgroup_dec_page_stat()
As before, only the file_mapped statistic is managed. However, these more
general interfaces allow for new statistics to be more easily added. New
statistics are added with memcg dirty page accounting.
Signed-off-by: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrea Righi <arighi@develer.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
hugetlbfs was changed to allow memory failure to migrate the hugetlbfs
pages and that broke THP as split_huge_page was then called on hugetlbfs
pages too.
compound_head/order was also run unsafe on THP pages that can be splitted
at any time.
All compound_head() invocations in memory-failure.c that are run on pages
that aren't pinned and that can be freed and reused from under us (while
compound_head is running) are buggy because compound_head can return a
dangling pointer, but I'm not fixing this as this is a generic
memory-failure bug not specific to THP but it applies to hugetlbfs too, so
I can fix it later after THP is merged upstream.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add hugepage stat information to /proc/vmstat and /proc/meminfo.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This documents how split_huge_page is safe vs new vma inserctions into the
anon_vma that may have already released the anon_vma->lock but not
established pmds yet when split_huge_page starts.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Paging logic that splits the page before it is unmapped and added to swap
to ensure backwards compatibility with the legacy swap code. Eventually
swap should natively pageout the hugepages to increase performance and
decrease seeking and fragmentation of swap space. swapoff can just skip
over huge pmd as they cannot be part of swap yet. In add_to_swap be
careful to split the page only if we got a valid swap entry so we don't
split hugepages with a full swap.
In theory we could split pages before isolating them during the lru scan,
but for khugepaged to be safe, I'm relying on either mmap_sem write mode,
or PG_lock taken, so split_huge_page has to run either with mmap_sem
read/write mode or PG_lock taken. Calling it from isolate_lru_page would
make locking more complicated, in addition to that split_huge_page would
deadlock if called by __isolate_lru_page because it has to take the lru
lock to add the tail pages.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Make anon_vma_chain_free() static. It is called only in rmap.c and the
corresponding alloc function is already static.
Signed-off-by: Namhyung Kim <namhyung@gmail.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The page_check_address() conditionally grabs *@ptlp in case of returning
non-NULL. Rename and wrap it using __cond_lock() removes following
warnings from sparse:
mm/rmap.c:472:9: warning: context imbalance in 'page_mapped_in_vma' - unexpected unlock
mm/rmap.c:524:9: warning: context imbalance in 'page_referenced_one' - unexpected unlock
mm/rmap.c:706:9: warning: context imbalance in 'page_mkclean_one' - unexpected unlock
mm/rmap.c:1066:9: warning: context imbalance in 'try_to_unmap_one' - unexpected unlock
Signed-off-by: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The page_lock_anon_vma() conditionally grabs RCU and anon_vma lock but
page_unlock_anon_vma() releases them unconditionally. This leads sparse
to complain about context imbalance. Annotate them.
Signed-off-by: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
