Re: [PATCH] mm/hugetlb: try preferred node first when alloc gigantic page from cma

[Michal Hocko <mhocko@suse.com>]

On Mon 31-08-20 14:44:40, Mike Kravetz wrote:
> On 8/30/20 7:04 AM, Li Xinhai wrote:
> > Since commit cf11e85fc08cc6a4 ("mm: hugetlb: optionally allocate gigantic
> > hugepages using cma"), the gigantic page would be allocated from node
> > which is not the preferred node, although there are pages available from
> > that node. The reason is that the nid parameter has been ignored in
> > alloc_gigantic_page().
> > 
> > After this patch, the preferred node is tried first before other allowed
> > nodes.
> 
> Thank you!
> This is an issue that needs to be fixed.
> 
> > Fixes: cf11e85fc08cc6a4 ("mm: hugetlb: optionally allocate gigantic hugepages using cma")
> > Cc: Roman Gushchin <guro@fb.com>
> > Cc: Mike Kravetz <mike.kravetz@oracle.com>
> > Cc: Michal Hocko <mhocko@kernel.org>
> > Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
> > ---
> >  mm/hugetlb.c | 9 ++++++++-
> >  1 file changed, 8 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > index a301c2d672bf..4a28b8853d47 100644
> > --- a/mm/hugetlb.c
> > +++ b/mm/hugetlb.c
> > @@ -1256,8 +1256,15 @@ static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
> >  		struct page *page;
> >  		int node;
> >  
> > +		if (hugetlb_cma[nid]) {
> > +			page = cma_alloc(hugetlb_cma[nid], nr_pages,
> > +					huge_page_order(h), true);
> > +			if (page)
> > +				return page;
> > +		}
> > +
> 
> When looking at your changes, I noticed that this code for allocation
> from CMA does not take gfp_mask into account.  The 'normal' use case
> is to allocate pool pages with something similar to:
> 
> echo 16 > /sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages
> 
> The routine alloc_pool_huge_page will try to interleave pages among nodes:
> 
> 	...
>         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
> 
>         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
> 	...
> 
> which will eventually call alloc_gigantic_page.  If __GFP_THISNODE is
> set we really do not want to execute the below for loop in alloc_gigantic_page.

Yes, this is the case indeed.
 
> I think the convention in the mm code is that only the lowest level
> allocation routines should interpret the GFP flags.  We may need to make
> an exception here and check for __GFP_THISNODE.

Yes this is true, But alloc_gigantic_page is actually low level
allocation routine in fact.

I would go with the following
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index a301c2d672bf..124754240b56 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -1256,6 +1256,16 @@ static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
 		struct page *page;
 		int node;
 
+		if (nid != NUMA_NO_NODE && hugetlb_cma[nid]) {
+			page = cma_alloc(hugetlb_cma[nid], nr_pages,
+					 huge_page_order(h), true);
+			if (page)
+				return page;
+		}
+
+		if (gfp_mask & __GFP_THISNODE)
+			return NULL;
+
 		for_each_node_mask(node, *nodemask) {
 			if (!hugetlb_cma[node])
 				continue;

I do not think we actually do have an explicit NUMA_NO_NODE user but it
is safer to not asume that here.
-- 
Michal Hocko
SUSE Labs