From mboxrd@z Thu Jan 1 00:00:00 1970 From: Michal Hocko Subject: Re: [PATCH memcg 3/3] memcg: handle memcg oom failures Date: Thu, 21 Oct 2021 18:47:29 +0200 Message-ID: References: Mime-Version: 1.0 Return-path: DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=suse.com; s=susede1; t=1634834851; h=from:from:reply-to:date:date:message-id:message-id:to:to:cc:cc: mime-version:mime-version:content-type:content-type: in-reply-to:in-reply-to:references:references; bh=++iA8+uhip7K2PBGESkJUKSkd2b03fx9ioYV7M+4ADc=; b=hW3ymvnHy/zIjqEGz83CLNjmBv0TkMYAecEBj9ZicauP81r1HFWZ1zXmQBDdateXmP6kUp mtbGbyRFsaIHg764GPo8VMrUN+e05/45ZU7+3dDBD60xJXN36r8xEaTA6nM84CnyhJB6Tt gEEGzt8067R6EwrCFARVP0xMDJPvyPQ= Content-Disposition: inline In-Reply-To: List-ID: Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Vasily Averin Cc: Johannes Weiner , Vladimir Davydov , Andrew Morton , Roman Gushchin , Uladzislau Rezki , Vlastimil Babka , Shakeel Butt , Mel Gorman , Tetsuo Handa , cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-mm-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org, linux-kernel-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, kernel-GEFAQzZX7r8dnm+yROfE0A@public.gmane.org On Thu 21-10-21 18:05:28, Vasily Averin wrote: > On 21.10.2021 14:49, Michal Hocko wrote: > > I do understand that handling a very specific case sounds easier but it > > would be better to have a robust fix even if that requires some more > > head scratching. So far we have collected several reasons why the it is > > bad to trigger oom killer from the #PF path. There is no single argument > > to keep it so it sounds like a viable path to pursue. Maybe there are > > some very well hidden reasons but those should be documented and this is > > a great opportunity to do either of the step. > > > > Moreover if it turns out that there is a regression then this can be > > easily reverted and a different, maybe memcg specific, solution can be > > implemented. > > Now I'm agree, > however I still have a few open questions. > > 1) VM_FAULT_OOM may be triggered w/o execution of out_of_memory() > for exampel it can be caused by incorrect vm fault operations, > (a) which can return this error without calling allocator at all. I would argue this to be a bug. How can that particular code tell whether the system is OOM and the oom killer is the a reasonable measure to take? > (b) or which can provide incorrect gfp flags and allocator can fail without execution of out_of_memory. I am not sure I can see any sensible scenario where pagefault oom killer would be an appropriate fix for that. > (c) This may happen on stable/LTS kernels when successful allocation was failed by hit into limit of legacy memcg-kmem contoller. > We'll drop it in upstream kernels, however how to handle it in old kenrels? Triggering the global oom killer for legacy kmem charge failure is clearly wrong. Removing oom killer from #PF would fix that problem. > We can make sure that out_of_memory or alocator was called by set of some per-task flags. I am not sure I see how that would be useful other than reporting a dubious VM_FAULT_OOM usage. I am also not sure how that would be implemented as allocator can be called several times not to mention that the allocation itself could have been done from a different context - e.g. WQ. > Can pagefault_out_of_memory() send itself a SIGKILL in all these cases? In principle it can as sending signal is not prohibited. I would argue it should not though because it is just wrong thing to do in all those cases. > If not -- task will be looped. Yes, but it will be killable from userspace. So this is not an unrecoverable situation. > It is much better than execution of global OOM, however it would be even better to avoid it somehow. How? > You said: "We cannot really kill the task if we could we would have done it by the oom killer already". > However what to do if we even not tried to use oom-killer? (see (b) and (c)) > or if we did not used the allocator at all (see (a)) See above > 2) in your patch we just exit from pagefault_out_of_memory(). and restart new #PF. > We can call schedule_timeout() and wait some time before a new #PF restart. > Additionally we can increase this delay in each new cycle. > It helps to save CPU time for other tasks. > What do you think about? I do not have a strong opinion on this. A short sleep makes sense. I am not sure a more complex implementation is really needed. -- Michal Hocko SUSE Labs