Re: [PATCH v2] mm/hugetlb: fix set_max_huge_pages() when there are surplus pages

[Oscar Salvador <osalvador@suse.de>]

On Mon, Apr 07, 2025 at 11:35:27AM +0200, Oscar Salvador wrote:
> On Mon, Apr 07, 2025 at 03:23:08PM +0800, Jinjiang Tu wrote:
> > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > index 39f92aad7bd1..4afcd5ce417c 100644
> > --- a/mm/hugetlb.c
> > +++ b/mm/hugetlb.c
> > @@ -3826,6 +3826,7 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> >                               nodemask_t *nodes_allowed)
> >  {
> >         unsigned long min_count;
> > +       unsigned long persistent_free_count;
> >         unsigned long allocated;
> >         struct folio *folio;
> >         LIST_HEAD(page_list);
> > @@ -3960,7 +3961,14 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> >          * and won't grow the pool anywhere else. Not until one of the
> >          * sysctls are changed, or the surplus pages go out of use.
> >          */
> > -       min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
> > +       persistent_free_count = h->free_huge_pages;
> > +       if (h->free_huge_pages > persistent_huge_pages(h)) {
> > +               if (h->free_huge_pages > h->surplus_huge_pages)
> > +                       persistent_free_count -= h->surplus_huge_pages;
> > +               else
> > +                       persistent_free_count = 0;
> > +       }
> > +       min_count = h->resv_huge_pages + persistent_huge_pages(h) - persistent_free_count;

I forgot to mention that before this chunk, we might still want
something like your remove_surplus_pool_hugetlb_folio(), but that goes
through all nodes, not only through the nodes_allowed nodemask and
checks for free surplus pages due to hvo, and tries to get rid of them.

But on the oher hand, we can leave as is, since "meh, we failed to restore
those pages so though luck", which is the status quo right now, and I
am fine with it.

In the end, the situation can get fixed by itself as those pages might be
reallocated again and then we have a second chance to restore vmemmap when
they get freed again.

 

-- 
Oscar Salvador
SUSE Labs