On Mon 05-08-19 23:00:12, Tetsuo Handa wrote: > On 2019/08/05 20:44, Michal Hocko wrote: > >> Allowing forced charge due to being unable to invoke memcg OOM killer > >> will lead to global OOM situation, and just returning -ENOMEM will not > >> solve memcg OOM situation. > > > > Returning -ENOMEM would effectivelly lead to triggering the oom killer > > from the page fault bail out path. So effectively get us back to before > > 29ef680ae7c21110. But it is true that this is riskier from the > > observability POV when a) the OOM path wouldn't point to the culprit and > > b) it would leak ENOMEM from g-u-p path. > > > > Excuse me? But according to my experiment, below code showed flood of > "Returning -ENOMEM" message instead of invoking the OOM killer. > I didn't find it gets us back to before 29ef680ae7c21110... You would need to declare OOM_ASYNC to return ENOMEM properly from the charge (which is effectivelly a revert of 29ef680ae7c21110 for NOFS allocations). Something like the following diff --git a/mm/memcontrol.c b/mm/memcontrol.c index ba9138a4a1de..cc34ff0932ce 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -1797,7 +1797,7 @@ static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int * Please note that mem_cgroup_out_of_memory might fail to find a * victim and then we have to bail out from the charge path. */ - if (memcg->oom_kill_disable) { + if (memcg->oom_kill_disable || !(mask & __GFP_FS)) { if (!current->in_user_fault) return OOM_SKIPPED; css_get(&memcg->css); I am quite surprised that your patch didn't trigger the global OOM though. It might mean that ENOMEM doesn't propagate all the way down to the #PF handler for this path for some reason. Anyway what I meant to say is that returning ENOMEM has the observable issues as well. -- Michal Hocko SUSE Labs