Re: [PATCH] mm,oom: Don't call schedule_timeout_killable() with oom_lock held.

[Michal Hocko <mhocko@kernel.org>]

On Tue 29-05-18 19:18:33, Michal Hocko wrote:
> On Tue 29-05-18 23:33:13, Tetsuo Handa wrote:
> > On 2018/05/29 17:16, Michal Hocko wrote:
> > > With the full changelog. This can be either folded into the respective
> > > patch or applied on top.
> > > 
> > >>From 0bd619e7a68337c97bdaed288e813e96a14ba339 Mon Sep 17 00:00:00 2001
> > > From: Michal Hocko <mhocko@suse.com>
> > > Date: Tue, 29 May 2018 10:09:33 +0200
> > > Subject: [PATCH] mm, memcg, oom: fix pre-mature allocation failures
> > > 
> > > Tetsuo has noticed that "mm, oom: cgroup-aware OOM killer" can lead to a
> > > pre-mature allocation failure if the cgroup aware oom killer is enabled
> > > and select_victim_memcg doesn't pick up any memcg to kill because there
> > > is a memcg already being killed. oc->chosen_memcg will become INFLIGHT_VICTIM
> > > and oom_kill_memcg_victim will bail out early. oc->chosen_task will
> > > stay NULL, however, and out_of_memory will therefore return false which
> > > forces __alloc_pages_may_oom to not set did_some_progress and the page
> > > allocator backs out and fails the allocation.
> > > U
> > > Fix this by checking both chosen_task and chosen_memcg in out_of_memory
> > > and return false only when _both_ are NULL.
> > 
> > I don't like this patch. It is not easy to understand and is fragile to
> > future changes. Currently the only case !!oc->chosen can become false is that
> > there was no eligible tasks when SysRq-f was requested or memcg OOM occurred.
> 
> Well, the current contract is not easy unfortunatelly. We have two
> different modes of operation. We are either killing whole cgroups or a
> task from a cgroup. In any case, the contract says that if we have any
> killable entity then at least one of chosen* is set to INFLIGHT_VICTIM.
> Other than that one of them has to be !NULL or we have no eligible
> killable entity. The return value reflects all these cases.

Btw. if your concern is the readability then we can add a helper and
decsribe all the above in the comment.
-- 
Michal Hocko
SUSE Labs