Re: [PATCH v2] mm,oom: Exclude TIF_MEMDIE processes from candidates.

[David Rientjes]

On Thu, 14 Jan 2016, Tetsuo Handa wrote:

> > > > > @@ -171,7 +195,7 @@ unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
> > > > >  	if (oom_unkillable_task(p, memcg, nodemask))
> > > > >  		return 0;
> > > > > 
> > > > > -	p = find_lock_task_mm(p);
> > > > > +	p = find_lock_non_victim_task_mm(p);
> > > > >  	if (!p)
> > > > >  		return 0;
> > > > > 
> > > > 
> > > > I understand how this may make your test case pass, but I simply don't 
> > > > understand how this could possibly be the correct thing to do.  This would 
> > > > cause oom_badness() to return 0 for any process where a thread has 
> > > > TIF_MEMDIE set.  If the oom killer is called from the page allocator, 
> > > > kills a thread, and it is recalled before that thread may exit, then this 
> > > > will panic the system if there are no other eligible processes to kill.
> > > > 
> > > Why? oom_badness() is called after oom_scan_process_thread() returned OOM_SCAN_OK.
> > > oom_scan_process_thread() returns OOM_SCAN_ABORT if a thread has TIF_MEMDIE set.
> > > 
> > 
> > oom_scan_process_thread() checks for TIF_MEMDIE on p, not on p's threads.
> > If one of p's threads has TIF_MEMDIE set and p does not, we actually want 
> > to set TIF_MEMDIE for p.  That's the current behavior since it will lead 
> > to p->mm memory freeing.  Your patch is excluding such processes entirely 
> > and selecting another process to kill unnecessarily.
> > 
> 
> I think p's threads are checked by oom_scan_process_thread() for TIF_MEMDIE
> even if p does not have TIF_MEMDIE. What am I misunderstanding about what
> for_each_process_thread(g, p) is doing?
> 
>   #define for_each_process_thread(p, t) for_each_process(p) for_each_thread(p, t)
> 
>   select_bad_process() {
>     for_each_process_thread(g, p) {
>       oom_scan_process_thread(oc, p, totalpages));
>       oom_badness(p);
>     }
>   }
> 

Yes, select_bad_process() iterates over threads, that is obvious.  The 
point is that today it can select a thread independent of whether any of 
its other threads have TIF_MEMDIE set, which is the desired behavior per 
the above.  With your change, that is no longer possible because we 
disregard _all_ threads if one of them has TIF_MEMDIE set.