The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The injector filter requires stable_page_flags() which is supplied
by procfs. So make it dependent on that.
Also add ifdefs around the filter code in memory-failure.c so that
when the filter is disabled due to missing dependencies the whole
code still builds.
Reported-by: Ingo Molnar
Signed-off-by: Andi Kleen <ak@linux.intel.com>
This is a simpler, gentler variant of memory_failure() for soft page
offlining controlled from user space. It doesn't kill anything, just
tries to invalidate and if that doesn't work migrate the
page away.
This is useful for predictive failure analysis, where a page has
a high rate of corrected errors, but hasn't gone bad yet. Instead
it can be offlined early and avoided.
The offlining is controlled from sysfs, including a new generic
entry point for hard page offlining for symmetry too.
We use the page isolate facility to prevent re-allocation
race. Normally this is only used by memory hotplug. To avoid
races with memory allocation I am using lock_system_sleep().
This avoids the situation where memory hotplug is about
to isolate a page range and then hwpoison undoes that work.
This is a big hammer currently, but the simplest solution
currently.
When the page is not free or LRU we try to free pages
from slab and other caches. The slab freeing is currently
quite dumb and does not try to focus on the specific slab
cache which might own the page. This could be potentially
improved later.
Thanks to Fengguang Wu and Haicheng Li for some fixes.
[Added fix from Andrew Morton to adapt to new migrate_pages prototype]
Signed-off-by: Andi Kleen <ak@linux.intel.com>
shake_page handles more types of page caches than
the much simpler lru_add_drain_all:
- slab (quite inefficiently for now)
- any other caches with a shrinker callback
- per cpu page allocator pages
- per CPU LRU
Use this call to try to turn pages into free or LRU pages.
Then handle the case of the page becoming free after drain everything.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
In some use cases, user doesn't need extra filtering. E.g. user program
can inject errors through madvise syscall to its own pages, however it
might not know what the page state exactly is or which inode the page
belongs to.
So introduce an one-off interface "corrupt-filter-enable".
Echo 0 to switch off page filters, and echo 1 to switch on the filters.
[AK: changed default to 0]
Signed-off-by: Haicheng Li <haicheng.li@linux.intel.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
The hwpoison test suite need to inject hwpoison to a collection of
selected task pages, and must not touch pages not owned by them and
thus kill important system processes such as init. (But it's OK to
mis-hwpoison free/unowned pages as well as shared clean pages.
Mis-hwpoison of shared dirty pages will kill all tasks, so the test
suite will target all or non of such tasks in the first place.)
The memory cgroup serves this purpose well. We can put the target
processes under the control of a memory cgroup, and tell the hwpoison
injection code to only kill pages associated with some active memory
cgroup.
The prerequisite for doing hwpoison stress tests with mem_cgroup is,
the mem_cgroup code tracks task pages _accurately_ (unless page is
locked). Which we believe is/should be true.
The benefits are simplification of hwpoison injector code. Also the
mem_cgroup code will automatically be tested by hwpoison test cases.
The alternative interfaces pin-pfn/unpin-pfn can also delegate the
(process and page flags) filtering functions reliably to user space.
However prototype implementation shows that this scheme adds more
complexity than we wanted.
Example test case:
mkdir /cgroup/hwpoison
usemem -m 100 -s 1000 &
echo `jobs -p` > /cgroup/hwpoison/tasks
memcg_ino=$(ls -id /cgroup/hwpoison | cut -f1 -d' ')
echo $memcg_ino > /debug/hwpoison/corrupt-filter-memcg
page-types -p `pidof init` --hwpoison # shall do nothing
page-types -p `pidof usemem` --hwpoison # poison its pages
[AK: Fix documentation]
[Add fix for problem noticed by Li Zefan <lizf@cn.fujitsu.com>;
dentry in the css could be NULL]
CC: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
CC: Hugh Dickins <hugh.dickins@tiscali.co.uk>
CC: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
CC: Balbir Singh <balbir@linux.vnet.ibm.com>
CC: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
CC: Li Zefan <lizf@cn.fujitsu.com>
CC: Paul Menage <menage@google.com>
CC: Nick Piggin <npiggin@suse.de>
CC: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
When specified, only poison pages if ((page_flags & mask) == value).
- corrupt-filter-flags-mask
- corrupt-filter-flags-value
This allows stress testing of many kinds of pages.
Strictly speaking, the buddy pages requires taking zone lock, to avoid
setting PG_hwpoison on a "was buddy but now allocated to someone" page.
However we can just do nothing because we set PG_locked in the beginning,
this prevents the page allocator from allocating it to someone. (It will
BUG() on the unexpected PG_locked, which is fine for hwpoison testing.)
[AK: Add select PROC_PAGE_MONITOR to satisfy dependency]
CC: Nick Piggin <npiggin@suse.de>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Filesystem data/metadata present the most tricky-to-isolate pages.
It requires careful code review and stress testing to get them right.
The fs/device filter helps to target the stress tests to some specific
filesystem pages. The filter condition is block device's major/minor
numbers:
- corrupt-filter-dev-major
- corrupt-filter-dev-minor
When specified (non -1), only page cache pages that belong to that
device will be poisoned.
The filters are checked reliably on the locked and refcounted page.
Haicheng: clear PG_hwpoison and drop bad page count if filter not OK
AK: Add documentation
CC: Haicheng Li <haicheng.li@intel.com>
CC: Nick Piggin <npiggin@suse.de>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Return 0 to indicate success, when
- action result is RECOVERED or DELAYED
- no extra page reference
Note that dirty swapcache pages are kept in swapcache, so can have one
more reference count.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Change semantics for
- IGNORED: not handled; it may well be _unsafe_
- DELAYED: to be handled later; it is _safe_
With this change,
- IGNORED/FAILED mean (maybe) Error
- DELAYED/RECOVERED mean Success
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
The unpoisoning interface is useful for stress testing tools to
reclaim poisoned pages (to prevent OOM)
There is no hardware level unpoisioning, so this
cannot be used for real memory errors, only for software injected errors.
Note that it may leak pages silently - those who have been removed from
LRU cache, but not isolated from page cache/swap cache at hwpoison time.
Especially the stress test of dirty swap cache pages shall reboot system
before exhausting memory.
AK: Fix comments, add documentation, add printks, rename symbol
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Most free pages in the buddy system have no PG_buddy set.
Introduce is_free_buddy_page() for detecting them reliably.
CC: Nick Piggin <npiggin@suse.de>
CC: Mel Gorman <mel@linux.vnet.ibm.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
The buddy page has already be handled in the very beginning.
So remove redundant code.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Introduce delete_from_lru_cache() to
- clear PG_active, PG_unevictable to avoid complains at unpoison time
- move the isolate_lru_page() call back to the handlers instead of the
entrance of __memory_failure(), this is more hwpoison filter friendly
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Don't try to isolate a still mapped page. Otherwise we will hit the
BUG_ON(page_mapped(page)) in __remove_from_page_cache().
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Now that "ref" is just a boolean turn it into
a flags argument. First step is only a single flag
that makes the code's intention more clear, but more
may follow.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
If page is double referenced in madvise_hwpoison() and __memory_failure(),
remove_mapping() will fail because it expects page_count=2. Fix it by
not grabbing extra page count in __memory_failure().
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Use a different errno than the usual EIO for invalid page numbers.
This is mainly for better reporting for the injector.
This also avoids calling action_result() with invalid pfn.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
(PG_swapbacked && !PG_lru) pages should not happen.
Better to treat them as unknown pages.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
shake_page handles more types of page caches than lru_drain_all()
- per cpu page allocator pages
- per CPU LRU
Stops early when the page became free.
Used in followon patches.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Remove three degrees of obfuscation, left over from when we had
CONFIG_UNEVICTABLE_LRU. MLOCK_PAGES is CONFIG_HAVE_MLOCKED_PAGE_BIT is
CONFIG_HAVE_MLOCK is CONFIG_MMU. rmap.o (and memory-failure.o) are only
built when CONFIG_MMU, so don't need such conditions at all.
Somehow, I feel no compulsion to remove the CONFIG_HAVE_MLOCK* lines from
169 defconfigs: leave those to evolve in due course.
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Izik Eidus <ieidus@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Nick Piggin <npiggin@suse.de>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The madvise injector already holds a reference when passing in a page
to the memory-failure code. The code corrects for this additional reference
for its checks, but the final printk output didn't. Fix that.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Memory failure on a KSM page currently oopses on its NULL anon_vma in
page_lock_anon_vma(): that may not be much worse than the consequence
of ignoring it, but it is better to be consistent with how ZERO_PAGE
and hugetlb pages and other awkward cases are treated. Just skip it.
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Right now we have some trouble with non atomic access
to page flags when locking the page. To plug this hole
for now, limit error recovery to LRU pages for now.
This could be better fixed by defining a suitable protocol,
but let's go this simple way for now
This avoids unnecessary races with __set_page_locked() and
__SetPageSlab*() and maybe more non-atomic page flag operations.
This loses isolated pages which are currently in page reclaim, but these
are relatively limited compared to the total memory.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
[AK: new description, bug fixes, cleanups]
Add the high level memory handler that poisons pages
that got corrupted by hardware (typically by a two bit flip in a DIMM
or a cache) on the Linux level. The goal is to prevent everyone
from accessing these pages in the future.
This done at the VM level by marking a page hwpoisoned
and doing the appropriate action based on the type of page
it is.
The code that does this is portable and lives in mm/memory-failure.c
To quote the overview comment:
High level machine check handler. Handles pages reported by the
hardware as being corrupted usually due to a 2bit ECC memory or cache
failure.
This focuses on pages detected as corrupted in the background.
When the current CPU tries to consume corruption the currently
running process can just be killed directly instead. This implies
that if the error cannot be handled for some reason it's safe to
just ignore it because no corruption has been consumed yet. Instead
when that happens another machine check will happen.
Handles page cache pages in various states. The tricky part
here is that we can access any page asynchronous to other VM
users, because memory failures could happen anytime and anywhere,
possibly violating some of their assumptions. This is why this code
has to be extremely careful. Generally it tries to use normal locking
rules, as in get the standard locks, even if that means the
error handling takes potentially a long time.
Some of the operations here are somewhat inefficient and have non
linear algorithmic complexity, because the data structures have not
been optimized for this case. This is in particular the case
for the mapping from a vma to a process. Since this case is expected
to be rare we hope we can get away with this.
There are in principle two strategies to kill processes on poison:
- just unmap the data and wait for an actual reference before
killing
- kill as soon as corruption is detected.
Both have advantages and disadvantages and should be used
in different situations. Right now both are implemented and can
be switched with a new sysctl vm.memory_failure_early_kill
The default is early kill.
The patch does some rmap data structure walking on its own to collect
processes to kill. This is unusual because normally all rmap data structure
knowledge is in rmap.c only. I put it here for now to keep
everything together and rmap knowledge has been seeping out anyways
Includes contributions from Johannes Weiner, Chris Mason, Fengguang Wu,
Nick Piggin (who did a lot of great work) and others.
Cc: npiggin@suse.de
Cc: riel@redhat.com
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Hidehiro Kawai <hidehiro.kawai.ez@hitachi.com>
