Re: [PATCH] mm,oom: Exclude TIF_MEMDIE processes from candidates.

[Michal Hocko]

On Fri 08-01-16 22:14:54, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Thu 07-01-16 11:28:15, Johannes Weiner wrote:
> > > On Tue, Dec 29, 2015 at 10:58:22PM +0900, Tetsuo Handa wrote:
> > > > >From 8bb9e36891a803e82c589ef78077838026ce0f7d Mon Sep 17 00:00:00 2001
> > > > From: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
> > > > Date: Tue, 29 Dec 2015 22:20:58 +0900
> > > > Subject: [PATCH] mm,oom: Exclude TIF_MEMDIE processes from candidates.
> > > > 
> > > > The OOM reaper kernel thread can reclaim OOM victim's memory before the victim
> > > > terminates. But since oom_kill_process() tries to kill children of the memory
> > > > hog process first, the OOM reaper can not reclaim enough memory for terminating
> > > > the victim if the victim is consuming little memory. The result is OOM livelock
> > > > as usual, for timeout based next OOM victim selection is not implemented.
> > > 
> > > What we should be doing is have the OOM reaper clear TIF_MEMDIE after
> > > it's done. There is no reason to wait for and prioritize the exit of a
> > > task that doesn't even have memory anymore. Once a task's memory has
> > > been reaped, subsequent OOM invocations should evaluate anew the most
> > > desirable OOM victim.
> > 
> > This is an interesting idea. It definitely sounds better than timeout
> > based solutions. I will cook up a patch for this. The API between oom
> > killer and the reaper has to change slightly but that shouldn't be a big
> > deal.
> 
> That is part of what I suggested at
> http://lkml.kernel.org/r/201512052133.IAE00551.LSOQFtMFFVOHOJ@xxxxxxxxxxxxxxxxxxx .
> | What about marking current OOM victim unkillable by updating
> | victim->signal->oom_score_adj to OOM_SCORE_ADJ_MIN and clearing victim's
> | TIF_MEMDIE flag when the victim is still alive for a second after
> | oom_reap_vmas() completed?

Sorry, I must have missed this part. I have added your Suggested-by to the
patch description.

> Can we update victim's oom_score_adj as well? Otherwise, the OOM killer
> might choose the same victim if victim's oom_score_adj was set to 1000.

Yes I've done that in the patch I am testing ATM.

-- 
Michal Hocko
SUSE Labs