One of our services observed a high rate of cgroup OOM kills in the
presence of large amounts of clean cache.  Debugging showed that the
culprit is the shared cgroup iteration in page reclaim.

Under high allocation concurrency, multiple threads enter reclaim at the
same time.  Fearing overreclaim when we first switched from the single
global LRU to cgrouped LRU lists, we introduced a shared iteration state
for reclaim invocations - whether 1 or 20 reclaimers are active
concurrently, we only walk the cgroup tree once: the 1st reclaimer
reclaims the first cgroup, the second the second one etc.  With more
reclaimers than cgroups, we start another walk from the top.

This sounded reasonable at the time, but the problem is that reclaim
concurrency doesn't scale with allocation concurrency.  As reclaim
concurrency increases, the amount of memory individual reclaimers get to
scan gets smaller and smaller.  Individual reclaimers may only see one
cgroup per cycle, and that may not have much reclaimable memory.  We see
individual reclaimers declare OOM when there is plenty of reclaimable
memory available in cgroups they didn't visit.

This patch does away with the shared iterator, and every reclaimer is
allowed to scan the full cgroup tree and see all of reclaimable memory,
just like it would on a non-cgrouped system.  This way, when OOM is
declared, we know that the reclaimer actually had a chance.

To still maintain fairness in reclaim pressure, disallow cgroup reclaim
from bailing out of the tree walk early.  Kswapd and regular direct
reclaim already don't bail, so it's not clear why limit reclaim would have
to, especially since it only walks subtrees to begin with.

This change completely eliminates the OOM kills on our service, while
showing no signs of overreclaim - no increased scan rates, %sys time, or
abrupt free memory spikes.  I tested across 100 machines that have 64G of
RAM and host about 300 cgroups each.

[ It's possible overreclaim never was a *practical* issue to begin
  with - it was simply a concern we had on the mailing lists at the
  time, with no real data to back it up. But we have also added more
  bail-out conditions deeper inside reclaim (e.g. the proportional
  exit in shrink_node_memcg) since. Regardless, now we have data that
  suggests full walks are more reliable and scale just fine. ]

Link: http://lkml.kernel.org/r/20190812192316.13615-1-hannes@cmpxchg.orgSigned-off-by: NJohannes Weiner <hannes@cmpxchg.org>
Reviewed-by: NRoman Gushchin <guro@fb.com>
Acked-by: NMichal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>