Re: [PATCH 2/2] mm: replace TIF_MEMDIE checks by tsk_is_oom_victim

[Roman Gushchin]

On Thu, Jul 27, 2017 at 04:45:44PM +0200, Michal Hocko wrote:
> On Thu 27-07-17 23:01:05, Tetsuo Handa wrote:
> > Michal Hocko wrote:
> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index 544d47e5cbbd..86a48affb938 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -1896,7 +1896,7 @@ static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
> > >  	 * bypass the last charges so that they can exit quickly and
> > >  	 * free their memory.
> > >  	 */
> > > -	if (unlikely(test_thread_flag(TIF_MEMDIE) ||
> > > +	if (unlikely(tsk_is_oom_victim(current) ||
> > >  		     fatal_signal_pending(current) ||
> > >  		     current->flags & PF_EXITING))
> > >  		goto force;
> > 
> > Did we check http://lkml.kernel.org/r/20160909140508.GO4844@xxxxxxxxxxxxxx ?
> 
> OK, so your concern was
> 
> > Does this test_thread_flag(TIF_MEMDIE) (or tsk_is_oom_victim(current)) make sense?
> > 
> > If current thread is OOM-killed, SIGKILL must be pending before arriving at
> > do_exit() and PF_EXITING must be set after arriving at do_exit().
> 
> > But I can't find locations which do memory allocation between clearing
> > SIGKILL and setting PF_EXITING.
> 
> I can't find them either and maybe there are none. But why do we care
> in this particular patch which merely replaces TIF_MEMDIE check by
> tsk_is_oom_victim? The code will surely not become less valid. If
> you believe this check is redundant then send a patch with the clear
> justification. But I would say, at least from the robustness point of
> view I would just keep it there. We do not really have any control on
> what happens between clearing signals and setting PF_EXITING.

I agree, this check is probably redundant, but it really makes no difference,
let's keep it bullet-proof. If we care about performance here, we can rearrange
the checks:
  if (unlikely(fatal_signal_pending(current) ||
  	     current->flags & PF_EXITING) ||
  	     tsk_is_oom_victim(current))
  	goto force;

Roman