On Wed, Feb 10, 2021 at 09:56:37AM +0100, David Hildenbrand wrote:
> On 08.02.21 11:38, Oscar Salvador wrote:
> > alloc_contig_range is not prepared to handle hugetlb pages and will
> > fail if it ever sees one, but since they can be migrated as any other
> > page (LRU and Movable), it makes sense to also handle them.
> >
> > For now, do it only when coming from alloc_contig_range.
> >
> > Signed-off-by: Oscar Salvador <osalvador@xxxxxxx>
> > ---
> > mm/compaction.c | 17 +++++++++++++++++
> > mm/vmscan.c | 5 +++--
> > 2 files changed, 20 insertions(+), 2 deletions(-)
> >
> > diff --git a/mm/compaction.c b/mm/compaction.c
> > index e5acb9714436..89cd2e60da29 100644
> > --- a/mm/compaction.c
> > +++ b/mm/compaction.c
> > @@ -940,6 +940,22 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
> > goto isolate_fail;
> > }
> > + /*
> > + * Handle hugetlb pages only when coming from alloc_contig
> > + */
> > + if (PageHuge(page) && cc->alloc_contig) {
> > + if (page_count(page)) {
>
> I wonder if we should care about races here. What if someone concurrently
> allocates/frees?
>
> Note that PageHuge() succeeds on tail pages, isolate_huge_page() not, i
> assume we'll have to handle that as well.
>
> I wonder if it would make sense to move some of the magic to hugetlb code
> and handle it there with less chances for races (isolate if used,
> alloc-and-dissolve if not).
Yes, it makes sense to keep the magic in hugetlb code.
Note, though, that removing all races might be tricky.
isolate_huge_page() checks for PageHuge under hugetlb_lock,
so there is a race between a call to PageHuge(x) and a subsequent
call to isolate_huge_page().
But we should be fine as isolate_huge_page will fail in case the page is
no longer HugeTLB.
Also, since isolate_migratepages_block() gets called with ranges
pageblock aligned, we should never be handling tail pages in the core
of the function. E.g: the same way we handle THP:
/* The whole page is taken off the LRU; skip the tail pages. */
if (PageCompound(page))
low_pfn += compound_nr(page) - 1;
But all in all, the code has to be more bullet-proof. This RFC was more
like a PoC to see whether something crazy was done.
And as I said, moving the handling of hugetlb pages to hugetlb.c might
help towards a better error-race-handling.
Thanks for having a look ;-)
--
Oscar Salvador
SUSE L3
