Re: [patch] mm, page_alloc: allow __GFP_NOFAIL to allocate below watermarks after reclaim

[Mel Gorman <mgorman@suse.de>]

On Tue, Dec 10, 2013 at 03:03:39PM -0800, David Rientjes wrote:
> On Tue, 10 Dec 2013, Mel Gorman wrote:
> 
> > > If direct reclaim has failed to free memory, __GFP_NOFAIL allocations
> > > can potentially loop forever in the page allocator.  In this case, it's
> > > better to give them the ability to access below watermarks so that they
> > > may allocate similar to the same privilege given to GFP_ATOMIC
> > > allocations.
> > > 
> > > We're careful to ensure this is only done after direct reclaim has had
> > > the chance to free memory, however.
> > > 
> > > Signed-off-by: David Rientjes <rientjes@google.com>
> > 
> > The main problem with doing something like this is that it just smacks
> > into the adjusted watermark if there are a number of __GFP_NOFAIL. Who
> > was the user of __GFP_NOFAIL that was fixed by this patch?
> > 
> 
> Nobody, it comes out of a memcg discussion where __GFP_NOFAIL were 
> recently given the ability to bypass charges to the root memcg when the 
> memcg has hit its limit since we disallow the oom killer to kill a process 
> (for the same reason that the vast majority of __GFP_NOFAIL users, those 
> that do GFP_NOFS | __GFP_NOFAIL, disallow the oom killer in the page 
> allocator).
> 
> Without some other thread freeing memory, these allocations simply loop 
> forever.  We probably don't want to reconsider the choice that prevents 
> calling the oom killer in !__GFP_FS contexts since it will allow 
> unnecessary oom killing when memory can actually be freed by another 
> thread.
> 
> Since there are comments in both gfp.h and page_alloc.c that say no new 
> users will be added, it seems legitimate to ensure that the allocation 
> will at least have a chance of succeeding, but not the point of depleting 
> memory reserves entirely.
> 

Which __GFP_NOFAIL on its own does not guarantee if they just smack into
that barrier and cannot do anything. It changes the timing, not fixes
the problem.

> > There are enough bad users of __GFP_NOFAIL that I really question how
> > good an idea it is to allow emergency reserves to be used when they are
> > potentially leaked to other !__GFP_NOFAIL users via the slab allocator
> > shortly afterwards.
> > 
> 
> You could make the same argument for GFP_ATOMIC which can also allow 
> access to memory reserves.

The critical difference being that GFP_ATOMIC callers typically can handle
NULL being returned to them. GFP_ATOMIC storms may starve !GFP_ATOMIC
requests but it does not cause the same types of problems that
__GFP_NOFAIL using reserves would.

-- 
Mel Gorman
SUSE Labs

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>