Re: [PATCH -mm] slab: fix cpuset check in fallback_alloc

[David Rientjes <rientjes@google.com>]

On Mon, 11 Aug 2014, Vladimir Davydov wrote:

> > diff --git a/mm/slab.c b/mm/slab.c
> > --- a/mm/slab.c
> > +++ b/mm/slab.c
> > @@ -3047,16 +3047,19 @@ retry:
> >  	 * from existing per node queues.
> >  	 */
> >  	for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
> > -		nid = zone_to_nid(zone);
> > +		struct kmem_cache_node *n;
> >  
> > -		if (cpuset_zone_allowed_hardwall(zone, flags) &&
> > -			get_node(cache, nid) &&
> > -			get_node(cache, nid)->free_objects) {
> > -				obj = ____cache_alloc_node(cache,
> > -					flags | GFP_THISNODE, nid);
> > -				if (obj)
> > -					break;
> > -		}
> > +		nid = zone_to_nid(zone);
> > +		if (!cpuset_zone_allowed(zone, flags | __GFP_HARDWALL))
> 
> We must use softwall check here, otherwise we will proceed to
> alloc_pages even if there are lots of free slabs on other nodes.
> alloc_pages, in turn, may allocate from other nodes in case
> cpuset.mem_hardwall=0, because it uses softwall check, so it may add yet
> another free slab to another node's list even if it isn't empty. As a
> result, we may get free list bloating on other nodes. I've seen a
> machine with one of its nodes almost completely filled with inactive
> slabs for buffer_heads (dozens of GBs) w/o any chance to drop them. So,
> this is a bug that must be fixed.
> 

Right, I understand, and my patch makes no attempt to fix that issue, it's 
simply collapsing the code down into a single cpuset_zone_allowed() 
function and the context for the allocation is controlled by the gfp 
flags (and hardwall is controlled by setting __GFP_HARDWALL) as it should 
be.  I understand the issue you face, but I can't combine a cleanup with a 
fix and I would prefer to have your patch keep your commit description.  

The diffstat for my proposal removes many more lines than it adds and I 
think it will avoid this type of issue in the future for new callers.  
Your patch could then be based on the single cpuset_zone_allowed() 
function where you would simply have to remove the __GFP_HARDWALL above.  
Or, your patch could be merged first and then my cleanup on top, but it 
seems like your one-liner would be more clear if it is based on mine.

--
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>