On Wed 17-06-20 21:23:05, Naresh Kamboju wrote: > On Wed, 17 Jun 2020 at 19:41, Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > > > [Our emails have crossed] > > > > On Wed 17-06-20 14:57:58, Chris Down wrote: > > > Naresh Kamboju writes: > > > > mkfs -t ext4 /dev/disk/by-id/ata-TOSHIBA_MG04ACA100N_Y8RQK14KF6XF > > > > mke2fs 1.43.8 (1-Jan-2018) > > > > Creating filesystem with 244190646 4k blocks and 61054976 inodes > > > > Filesystem UUID: 7c380766-0ed8-41ba-a0de-3c08e78f1891 > > > > Superblock backups stored on blocks: > > > > 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632, 2654208, > > > > 4096000, 7962624, 11239424, 20480000, 23887872, 71663616, 78675968, > > > > 102400000, 214990848 > > > > Allocating group tables: 0/7453 done > > > > Writing inode tables: 0/7453 done > > > > Creating journal (262144 blocks): [ 51.544525] under min:0 emin:0 > > > > [ 51.845304] under min:0 emin:0 > > > > [ 51.848738] under min:0 emin:0 > > > > [ 51.858147] under min:0 emin:0 > > > > [ 51.861333] under min:0 emin:0 > > > > [ 51.862034] under min:0 emin:0 > > > > [ 51.862442] under min:0 emin:0 > > > > [ 51.862763] under min:0 emin:0 > > > > > > Thanks, this helps a lot. Somehow we're entering mem_cgroup_below_min even > > > when min/emin is 0 (which should indeed be the case if you haven't set them > > > in the hierarchy). > > > > > > My guess is that page_counter_read(&memcg->memory) is 0, which means > > > mem_cgroup_below_min will return 1. > > > > Yes this is the case because this is likely the root memcg which skips > > all charges. > > > > > However, I don't know for sure why that should then result in the OOM killer > > > coming along. My guess is that since this memcg has 0 pages to scan anyway, > > > we enter premature OOM under some conditions. I don't know why we wouldn't > > > have hit that with the old version of mem_cgroup_protected that returned > > > MEMCG_PROT_* members, though. > > > > Not really. There is likely no other memcg to reclaim from and assuming > > min limit protection will result in no reclaimable memory and thus the > > OOM killer. > > > > > Can you please try the patch with the `>=` checks in mem_cgroup_below_min > > > and mem_cgroup_below_low changed to `>`? If that fixes it, then that gives a > > > strong hint about what's going on here. > > > > This would work but I believe an explicit check for the root memcg would > > be easier to spot the reasoning. > > May I request you to send debugging or proposed fix patches here. > I am happy to do more testing. Sure, here is the diff to test. diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h index c74a8f2323f1..6b5a31672fbe 100644 --- a/include/linux/memcontrol.h +++ b/include/linux/memcontrol.h @@ -392,6 +392,13 @@ static inline bool mem_cgroup_below_low(struct mem_cgroup *memcg) if (mem_cgroup_disabled()) return false; + /* + * Root memcg doesn't account charges and doesn't support + * protection + */ + if (mem_cgroup_is_root(memcg)) + return false; + return READ_ONCE(memcg->memory.elow) >= page_counter_read(&memcg->memory); } @@ -401,6 +408,13 @@ static inline bool mem_cgroup_below_min(struct mem_cgroup *memcg) if (mem_cgroup_disabled()) return false; + /* + * Root memcg doesn't account charges and doesn't support + * protection + */ + if (mem_cgroup_is_root(memcg)) + return false; + return READ_ONCE(memcg->memory.emin) >= page_counter_read(&memcg->memory); } -- Michal Hocko SUSE Labs
