Re: [patch] mm, memcg: add oom killer delay

[Johannes Weiner <hannes-druUgvl0LCNAfugRpC6u6w@public.gmane.org>]

On Mon, Jun 03, 2013 at 06:31:34PM +0200, Michal Hocko wrote:
> On Sat 01-06-13 02:11:51, Johannes Weiner wrote:
> > @@ -2076,6 +2077,7 @@ static void memcg_wakeup_oom(struct mem_cgroup *memcg)
> >  {
> >  	/* for filtering, pass "memcg" as argument. */
> >  	__wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
> > +	atomic_inc(&memcg->oom_wakeups);
> >  }
> >  
> >  static void memcg_oom_recover(struct mem_cgroup *memcg)
> [...]
> > +	prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
> > +	/* Only sleep if we didn't miss any wakeups since OOM */
> > +	if (atomic_read(&memcg->oom_wakeups) == current->memcg_oom.wakeups)
> > +		schedule();
> 
> On the way home it occured to me that the ordering might be wrong here.
> The wake up can be lost here.
> 					__wake_up(memcg_oom_waitq)
> 					<preempted>
> prepare_to_wait
> atomic_read(&memcg->oom_wakeups)
> 					atomic_inc(oom_wakeups)
> 
> I guess we want atomic_inc before __wake_up, right?

I think you are right, thanks for spotting this.  Will be fixed in
version 2.