On Thu, Jun 4, 2020 at 1:17 AM Greg Thelen <gthelen@xxxxxxxxxx> wrote:
>
> Yang Shi <shy828301@xxxxxxxxx> wrote:
>
> > On Sun, May 31, 2020 at 8:22 PM Greg Thelen <gthelen@xxxxxxxxxx> wrote:
> >>
> >> Since v4.19 commit b0dedc49a2da ("mm/vmscan.c: iterate only over charged
> >> shrinkers during memcg shrink_slab()") a memcg aware shrinker is only
> >> called when the per-memcg per-node shrinker_map indicates that the
> >> shrinker may have objects to release to the memcg and node.
> >>
> >> shmem_unused_huge_count and shmem_unused_huge_scan support the per-tmpfs
> >> shrinker which advertises per memcg and numa awareness. The shmem
> >> shrinker releases memory by splitting hugepages that extend beyond
> >> i_size.
> >>
> >> Shmem does not currently set bits in shrinker_map. So, starting with
> >> b0dedc49a2da, memcg reclaim avoids calling the shmem shrinker under
> >> pressure. This leads to undeserved memcg OOM kills.
> >> Example that reliably sees memcg OOM kill in unpatched kernel:
> >> FS=/tmp/fs
> >> CONTAINER=/cgroup/memory/tmpfs_shrinker
> >> mkdir -p $FS
> >> mount -t tmpfs -o huge=always nodev $FS
> >> # Create 1000 MB container, which shouldn't suffer OOM.
> >> mkdir $CONTAINER
> >> echo 1000M > $CONTAINER/memory.limit_in_bytes
> >> echo $BASHPID >> $CONTAINER/cgroup.procs
> >> # Create 4000 files. Ideally each file uses 4k data page + a little
> >> # metadata. Assume 8k total per-file, 32MB (4000*8k) should easily
> >> # fit within container's 1000 MB. But if data pages use 2MB
> >> # hugepages (due to aggressive huge=always) then files consume 8GB,
> >> # which hits memcg 1000 MB limit.
> >> for i in {1..4000}; do
> >> echo . > $FS/$i
> >> done
> >
> > It looks all the inodes which have tail THP beyond i_size are on one
> > single list, then the shrinker actually just splits the first
> > nr_to_scan inodes. But since the list is not memcg aware, so it seems
> > it may split the THPs which are not charged to the victim memcg and
> > the victim memcg still may suffer from pre-mature oom, right?
>
> Correct. shmem_unused_huge_shrink() is not memcg aware. In response to
> memcg pressure it will split the post-i_size tails of nr_to_scan tmpfs
> inodes regardless of if they're charged to the under-pressure memcg.
> do_shrink_slab() looks like it'll repeatedly call
> shmem_unused_huge_shrink(). So it will split tails of many inodes. So
> I think it'll avoid the oom by over shrinking. This is not ideal. But
> it seems better than undeserved oom kill.
>
> I think the solution (as Kirill Tkhai suggested) a memcg-aware index
> would solve both:
> 1) avoid premature oom by registering shrinker to responding to memcg
> pressure
> 2) avoid shrinking/splitting inodes unrelated to the under-pressure
> memcg
I do agree with Kirill. Using list_lru sounds optimal. But, it looks
the memcg index is tricky. The index of memcg which the beyond i_size
THP is charged to should be used instead of the inode's memcg which
may charge to different memcg.
>
> I can certainly look into that (thanks Kirill for the pointers). In the
> short term I'm still interested in avoiding premature OOMs with the
> original thread (i.e. restore pre-4.19 behavior to shmem shrinker for
> memcg pressure). I plan to test and repost v2.
