Re: [PATCH] mm: fix the incorrect hugepages count

[Naoya Horiguchi]

On Tue, Aug 09, 2016 at 06:32:39PM +0800, zhong jiang wrote:
> On 2016/8/9 1:14, Mike Kravetz wrote:
> > On 08/07/2016 07:49 PM, zhongjiang wrote:
> >> From: zhong jiang <zhongjiang@xxxxxxxxxx>
> >>
> >> when memory hotplug enable, free hugepages will be freed if movable node offline.
> >> therefore, /proc/sys/vm/nr_hugepages will be incorrect.

This sounds a bit odd to me because /proc/sys/vm/nr_hugepages returns
h->nr_huge_pages or h->nr_huge_pages_node[nid], which is already
considered in dissolve_free_huge_page (via update_and_free_page).

I think that h->max_huge_pages effectively means the pool size, and
h->nr_huge_pages means total hugepage number (which can be greater than
the pool size when there's overcommiting/surplus.)

dissolve_free_huge_page intends to break a hugepage into buddy, and
the destination hugepage is supposed to be allocated from the pool of
the destination node, so the system-wide pool size is reduced.
So adding h->max_huge_pages-- makes sense to me.

Acked-by: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx>

> >>
> >> The patch fix it by reduce the max_huge_pages when the node offline.
> >>
> >> Signed-off-by: zhong jiang <zhongjiang@xxxxxxxxxx>
> >> ---
> >>  mm/hugetlb.c | 1 +
> >>  1 file changed, 1 insertion(+)
> >>
> >> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> >> index f904246..3356e3a 100644
> >> --- a/mm/hugetlb.c
> >> +++ b/mm/hugetlb.c
> >> @@ -1448,6 +1448,7 @@ static void dissolve_free_huge_page(struct page *page)
> >>  		list_del(&page->lru);
> >>  		h->free_huge_pages--;
> >>  		h->free_huge_pages_node[nid]--;
> >> +		h->max_huge_pages--;
> >>  		update_and_free_page(h, page);
> >>  	}
> >>  	spin_unlock(&hugetlb_lock);
> >>
> > Adding Naoya as he was the original author of this code.
> >
> > >From quick look it appears that the huge page will be migrated (allocated
> > on another node).  If my understanding is correct, then max_huge_pages
> > should not be adjusted here.
> >
>   we need to take free hugetlb pages into account.  of course, the allocated huge pages is no
>   need to reduce.  The patch just reduce the free hugetlb pages count.

I