Re: [patch] mm, oom: prevent additional oom kills before memory is freed

[Michal Hocko]

On Thu 15-06-17 22:01:33, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Thu 15-06-17 14:03:35, Michal Hocko wrote:
> > > On Thu 15-06-17 20:32:39, Tetsuo Handa wrote:
> > > > @@ -556,25 +553,21 @@ static void oom_reap_task(struct task_struct *tsk)
> > > >  	struct mm_struct *mm = tsk->signal->oom_mm;
> > > >  
> > > >  	/* Retry the down_read_trylock(mmap_sem) a few times */
> > > > -	while (attempts++ < MAX_OOM_REAP_RETRIES && !__oom_reap_task_mm(tsk, mm))
> > > > +	while (__oom_reap_task_mm(tsk, mm), !test_bit(MMF_OOM_SKIP, &mm->flags)
> > > > +	       && attempts++ < MAX_OOM_REAP_RETRIES)
> > > >  		schedule_timeout_idle(HZ/10);
> > > >  
> > > > -	if (attempts <= MAX_OOM_REAP_RETRIES)
> > > > -		goto done;
> > > > -
> > > > -
> > > > -	pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
> > > > -		task_pid_nr(tsk), tsk->comm);
> > > > -	debug_show_all_locks();
> > > > -
> > > > -done:
> > > > -	tsk->oom_reaper_list = NULL;
> > > > -
> > > >  	/*
> > > >  	 * Hide this mm from OOM killer because it has been either reaped or
> > > >  	 * somebody can't call up_write(mmap_sem).
> > > >  	 */
> > > > -	set_bit(MMF_OOM_SKIP, &mm->flags);
> > > > +	if (!test_and_set_bit(MMF_OOM_SKIP, &mm->flags)) {
> > > > +		pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
> > > > +			task_pid_nr(tsk), tsk->comm);
> > > > +		debug_show_all_locks();
> > > > +	}
> > > > +
> > > 
> > > How does this _solve_ anything? Why would you even retry when you
> > > _know_ that the reference count dropped to zero. It will never
> > > increment. So the above is basically just schedule_timeout_idle(HZ/10) *
> > > MAX_OOM_REAP_RETRIES before we set MMF_OOM_SKIP.
> 
> If the OOM reaper knows that mm->users == 0, it gives __mmput() some time
> to "complete exit_mmap() etc. and set MMF_OOM_SKIP". If __mmput() released
> some memory, subsequent OOM killer invocation is automatically avoided.
> If __mmput() did not release some memory, let the OOM killer invoke again.
> 
> > 
> > Just to make myself more clear. The above assumes that the victim hasn't
> > passed exit_mmap and MMF_OOM_SKIP in __mmput. Which is the case we want to
> > address here.
> 
> David is trying to avoid setting MMF_OOM_SKIP when the OOM reaper found that
> mm->users == 0. But we must not wait forever because __mmput() might fail to
> release some memory immediately. If __mmput() did not release some memory within
> schedule_timeout_idle(HZ/10) * MAX_OOM_REAP_RETRIES sleep, let the OOM killer
> invoke again. So, this is the case we want to address here, isn't it?

And we are back with a timeout based approach... Sigh. Just imagine that
you have a really large process which will take some time to tear down.
While it frees memory that might be in a different oom domain. Now you
pretend to keep retrying and eventually give up to allow a new oom
victim from that oom domain.

If we want to handle oom victims with mm_users == 0 then let's do it
properly, please.

-- 
Michal Hocko
SUSE Labs