[PATCH 00/11] oom detection rework v5

[Michal Hocko]

Hi,

This is v5 of the series. The previous version was posted [1]. There
was a number of changes since then. Mostly driven by premature oom
killer invocation for higher order requests reported by Hugh [2]. That
forced me to look into compaction because it became clear that we have
to consider its feedback to make any retry decisions especially for
!costly high order requests.  As a result we have 7 compaction related
patches because the code wasn't really prepared to tell its caller
about a feedback. I have tried hard to not change compaction logic
and profile it for my specific test cases and rather come up with a
generic feedback mechanism which would be mostly independent on the
current implementation.

As pointed by Linus [3][4] relying on zone_reclaimable as a way to
communicate the reclaim progress is rater dubious. I tend to agree,
not only it is really obscure, it is not hard to imagine cases where a
single page freed in the loop keeps all the reclaimers looping without
getting any progress because their gfp_mask wouldn't allow to get that
page anyway (e.g. single GFP_ATOMIC alloc and free loop). This is rather
rare so it doesn't happen in the practice but the current logic which we
have is rather obscure and hard to follow a also non-deterministic.

This is an attempt to make the OOM detection more deterministic and
easier to follow because each reclaimer basically tracks its own
progress which is implemented at the page allocator layer rather spread
out between the allocator and the reclaim. The more on the implementation
is described in the first patch.

I have tested several different scenarios but it should be clear that
testing OOM killer is quite hard to be representative. There is usually
a tiny gap between almost OOM and full blown OOM which is often time
sensitive. Anyway, I have tested the following 2 scenarios and I would
appreciate if there are more to test.

Testing environment: a virtual machine with 2G of RAM and 2CPUs without
any swap to make the OOM more deterministic.

1) 2 writers (each doing dd with 4M blocks to an xfs partition with 1G
   file size, removes the files and starts over again) running in
   parallel for 10s to build up a lot of dirty pages when 100 parallel
   mem_eaters (anon private populated mmap which waits until it gets
   signal) with 80M each.

   This causes an OOM flood of course and I have compared both patched
   and unpatched kernels. The test is considered finished after there
   are no OOM conditions detected. This should tell us whether there are
   any excessive kills or some of them premature (e.g. due to dirty pages):

I have performed two runs this time each after a fresh boot.

* base kernel
$ grep "Out of memory:" base-oom-run1.log | wc -l
78
$ grep "Out of memory:" base-oom-run2.log | wc -l
78

$ grep "Kill process" base-oom-run1.log | tail -n1
[   91.391203] Out of memory: Kill process 3061 (mem_eater) score 39 or sacrifice child
$ grep "Kill process" base-oom-run2.log | tail -n1
[   82.141919] Out of memory: Kill process 3086 (mem_eater) score 39 or sacrifice child

$ grep "DMA32 free:" base-oom-run1.log | sed 's@.*free:\([0-9]*\)kB.*@\1@' | calc_min_max.awk 
min: 5376.00 max: 6776.00 avg: 5530.75 std: 166.50 nr: 61
$ grep "DMA32 free:" base-oom-run2.log | sed 's@.*free:\([0-9]*\)kB.*@\1@' | calc_min_max.awk 
min: 5416.00 max: 5608.00 avg: 5514.15 std: 42.94 nr: 52

$ grep "DMA32.*all_unreclaimable? no" base-oom-run1.log | wc -l
1
$ grep "DMA32.*all_unreclaimable? no" base-oom-run2.log | wc -l
3

* patched kernel
$ grep "Out of memory:" patched-oom-run1.log | wc -l
78
miso@tiehlicka /mnt/share/devel/miso/kvm $ grep "Out of memory:" patched-oom-run2.log | wc -l
77

e grep "Kill process" patched-oom-run1.log | tail -n1
[  497.317732] Out of memory: Kill process 3108 (mem_eater) score 39 or sacrifice child
$ grep "Kill process" patched-oom-run2.log | tail -n1
[  316.169920] Out of memory: Kill process 3093 (mem_eater) score 39 or sacrifice child

$ grep "DMA32 free:" patched-oom-run1.log | sed 's@.*free:\([0-9]*\)kB.*@\1@' | calc_min_max.awk 
min: 5420.00 max: 5808.00 avg: 5513.90 std: 60.45 nr: 78
$ grep "DMA32 free:" patched-oom-run2.log | sed 's@.*free:\([0-9]*\)kB.*@\1@' | calc_min_max.awk 
min: 5380.00 max: 6384.00 avg: 5520.94 std: 136.84 nr: 77

e grep "DMA32.*all_unreclaimable? no" patched-oom-run1.log | wc -l
2
$ grep "DMA32.*all_unreclaimable? no" patched-oom-run2.log | wc -l
3

The patched kernel run noticeably longer while invoking OOM killer same
number of times. This means that the original implementation is much
more aggressive and triggers the OOM killer sooner. free pages stats
show that neither kernels went OOM too early most of the time, though. I
guess the difference is in the backoff when retries without any progress
do sleep for a while if there is memory under writeback or dirty which
is highly likely considering the parallel IO.
Both kernels have seen races where zone wasn't marked unreclaimable
and we still hit the OOM killer. This is most likely a race where
a task managed to exit between the last allocation attempt and the oom
killer invocation.

2) 2 writers again with 10s of run and then 10 mem_eaters to consume as much
   memory as possible without triggering the OOM killer. This required a lot
   of tuning but I've considered 3 consecutive runs in three different boots
   without OOM as a success.

* base kernel
size=$(awk '/MemFree/{printf "%dK", ($2/10)-(16*1024)}' /proc/meminfo)

* patched kernel
size=$(awk '/MemFree/{printf "%dK", ($2/10)-(12*1024)}' /proc/meminfo)

That means 40M more memory was usable without triggering OOM killer. The
base kernel sometimes managed to handle the same as patched but it
wasn't consistent and failed in at least on of the 3 runs. This seems
like a minor improvement.

I was testing also GPF_REPEAT costly requests (hughetlb) with fragmented
memory and under memory pressure. The results are in patch 11 where the
logic is implemented. In short I can see huge improvement there.

I am certainly interested in other usecases as well as well as any
feedback. Especially those which require higher order requests.

* Changes since v4
- dropped __GFP_REPEAT for costly allocation as it is now replaced by
  the compaction based feedback logic
- !costly high order requests are retried based on the compaction feedback
- compaction feedback has been tweaked to give us an useful information
  to make decisions in the page allocator
- rebased on the current mmotm-2016-04-01-16-24 with the previous version
  of the rework reverted

* Changes since v3
- factor out the new heuristic into its own function as suggested by
  Johannes (no functional changes)

* Changes since v2
- rebased on top of mmotm-2015-11-25-17-08 which includes
  wait_iff_congested related changes which needed refresh in
  patch#1 and patch#2
- use zone_page_state_snapshot for NR_FREE_PAGES per David
- shrink_zones doesn't need to return anything per David
- retested because the major kernel version has changed since
  the last time (4.2 -> 4.3 based kernel + mmotm patches)

* Changes since v1
- backoff calculation was de-obfuscated by using DIV_ROUND_UP
- __GFP_NOFAIL high order migh fail fixed - theoretical bug

[1] http://lkml.kernel.org/r/1450203586-10959-1-git-send-email-mhocko@xxxxxxxxxx
[2] http://lkml.kernel.org/r/alpine.LSU.2.11.1602241832160.15564@xxxxxxxxxxxx
[3] http://lkml.kernel.org/r/CA+55aFwapaED7JV6zm-NVkP-jKie+eQ1vDXWrKD=SkbshZSgmw@xxxxxxxxxxxxxx
[4] http://lkml.kernel.org/r/CA+55aFxwg=vS2nrXsQhAUzPQDGb8aQpZi0M7UUh21ftBo-z46Q@xxxxxxxxxxxxxx