[PATCH 2/3] slub: fallback to node_to_mem_node() node if allocating on memoryless node

Nishanth Aravamudan nacc at linux.vnet.ibm.com
Wed Sep 10 10:55:42 EST 2014


On 09.09.2014 [17:11:25 -0700], Andrew Morton wrote:
> On Tue, 9 Sep 2014 12:05:14 -0700 Nishanth Aravamudan <nacc at linux.vnet.ibm.com> wrote:
> 
> > From: Joonsoo Kim <iamjoonsoo.kim at lge.com>
> > 
> > Update the SLUB code to search for partial slabs on the nearest node
> > with memory in the presence of memoryless nodes. Additionally, do not
> > consider it to be an ALLOC_NODE_MISMATCH (and deactivate the slab) when
> > a memoryless-node specified allocation goes off-node.
> > 
> > ...
> >
> > --- a/mm/slub.c
> > +++ b/mm/slub.c
> > @@ -1699,7 +1699,12 @@ static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
> >  		struct kmem_cache_cpu *c)
> >  {
> >  	void *object;
> > -	int searchnode = (node == NUMA_NO_NODE) ? numa_mem_id() : node;
> > +	int searchnode = node;
> > +
> > +	if (node == NUMA_NO_NODE)
> > +		searchnode = numa_mem_id();
> > +	else if (!node_present_pages(node))
> > +		searchnode = node_to_mem_node(node);
> 
> I expect a call to node_to_mem_node() will always be preceded by a test
> of node_present_pages().  Perhaps node_to_mem_node() should just do the
> node_present_pages() call itself?

Really, we don't need that test here. We could always use the result of
node_to_mem_node() in the else. If memoryless nodes are not supported
(off in .config), then node_to_mem_node() trivially returns. If they are
supported, it returns the correct value for all nodes.
 
It's just an optimization (premature?) since we can avoid worrying (in
this path) about memoryless nodes if the node in question has memory.

And, in fact, in __slab_alloc(), we could do the following:

...
	int searchnode = node;

	if (node != NUMA_NO_NODE)
		searchnode = node_to_mem_node(node);

	if (node != searchnode &&
		unlikely(!node_match(page, searchnode))) {

...

which would minimize the impact to non-memoryless node NUMA configs.

Does that seem better to you? I can add comments to this patch as well.

Thanks,
Nish



More information about the Linuxppc-dev mailing list