Re: [RFC PATCH] mm, oom: move GFP_NOFS check to out_of_memory

[Michal Hocko]

On Wed 30-03-16 20:46:48, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Tue 29-03-16 15:13:54, David Rientjes wrote:
> > > On Tue, 29 Mar 2016, Michal Hocko wrote:
> > > 
> > > > diff --git a/mm/oom_kill.c b/mm/oom_kill.c
> > > > index 86349586eacb..1c2b7a82f0c4 100644
> > > > --- a/mm/oom_kill.c
> > > > +++ b/mm/oom_kill.c
> > > > @@ -876,6 +876,10 @@ bool out_of_memory(struct oom_control *oc)
> > > >  		return true;
> > > >  	}
> > > >  
> > > > +	/* The OOM killer does not compensate for IO-less reclaim. */
> > > > +	if (!(oc->gfp_mask & __GFP_FS))
> > > > +		return true;
> > > > +
> 
> This patch will disable pagefault_out_of_memory() because currently
> pagefault_out_of_memory() is passing oc->gfp_mask == 0.
> 
> Because of current behavior, calling oom notifiers from !__GFP_FS seems
> to be safe.

You are right! I have completely missed that and thought we were
providing GFP_KERNEL there. So we have two choices. Either we do
use GFP_KERNEL (same as we do for sysrq+f) or we special case
pagefault_out_of_memory in some way. The second option seems to be safer
because the gfp_mask has to contain at least ___GFP_DIRECT_RECLAIM to
trigger the OOM path.

> > > >  	/*
> > > >  	 * Check if there were limitations on the allocation (only relevant for
> > > >  	 * NUMA) that may require different handling.
> > > 
> > > I don't object to this necessarily, but I think we need input from those 
> > > that have taken the time to implement their own oom notifier to see if 
> > > they agree.  In the past, they would only be called if reclaim has 
> > > completely failed; now, they can be called in low memory situations when 
> > > reclaim has had very little chance to be successful.  Getting an ack from 
> > > them would be helpful.
> > 
> > I will make sure to put them on the CC and mention this in the changelog
> > when I post this next time. I personally think that this shouldn't make
> > much difference in the real life because GFP_NOFS only loads are rare
> 
> GFP_NOFS only loads are rare. But some GFP_KERNEL load which got TIF_MEMDIE
> might be waiting for GFP_NOFS or GFP_NOIO loads to make progress.

How would that matter to oom notifiers?

> I think we are not ready to handle situations where out_of_memory() is called
> again after current thread got TIF_MEMDIE due to __GFP_NOFAIL allocation
> request when we ran out of memory reserves. We should not assume that the
> victim target thread does not have TIF_MEMDIE yet. I think we can handle it
> by making mark_oom_victim() return a bool and return via shortcut only if
> mark_oom_victim() successfully set TIF_MEMDIE. Though I don't like the
> shortcut approach that lacks a guaranteed unlocking mechanism.

That would lead to premature follow up OOM when TIF_MEMDIE makes some
progress just not in time.
 
> > and we should rather help by releasing memory when it is available
> > rather than rely on something else to do it for us. Waiting for Godot is
> > never a good strategy.
> > 
> > > I also think we have discussed this before, but I think the oom notifier 
> > > handling should be in done in the page allocator proper, i.e. in 
> > > __alloc_pages_may_oom().  We can leave out_of_memory() for a clear defined 
> > > purpose: to kill a process when all reclaim has failed.
> > 
> > I vaguely remember there was some issue with that the last time we have
> > discussed that. It was the duplication from the page fault and allocator
> > paths AFAIR. Nothing that cannot be handled though but the OOM notifier
> > API is just too ugly to spread outside OOM proper I guess. Why we cannot
> > move those users to use proper shrinkers interface (after it gets
> > extended by a priority of some sort and release some objects only after
> > we are really in troubles)? Something for a separate discussion,
> > though...
> 
> Calling oom notifiers from SysRq-f is what we want?

I am not really sure about that to be honest. The semantic is really
weak but what would be a downside? This operation shouldn't be fatal
and dropped object can be reconstructed.
-- 
Michal Hocko
SUSE Labs