On 2/3/21 2:53 PM, Jason Gunthorpe wrote:

On Wed, Feb 03, 2021 at 01:22:18PM +0000, Joao Martins wrote:

quoted

With this, longterm gup will 'regress' for hugetlbfs e.g. from ~6k -> ~32k usecs when
pinning a 16G hugetlb file.

Yes, but correctness demands it.

Hence the quotes around the word regress.

But still, we are adding unnecessary overhead (previously nonexistent) to compound pages
even those where the issue doesn't exist (hugetlb).

The solution is to track these pages as we discover them so we know if
a PMD/PUD points and can directly skip the duplicated work

That looks better. It would require moving check_and_migrate_movable_pages() elsewhere,
and/or possibly reworking the entire series?

quoted

Splitting can only occur on THP right? If so, perhaps we could
retain the @step increment for compound pages but when
!is_transparent_hugepage(head) or just PageHuge(head) like:

Honestly I'd rather see it fixed properly which will give even bigger
performance gains - avoiding the entire rescan of the page list will
be a win

If check_and_migrate_movable_pages() is meant to migrate unpinned pages, then rather than
pinning+unpinning+moving, perhaps it would be called in __get_user_pages() place where we
are walking page tables and know if it's a PUD/PMD and can skip all the subpages and just
record and migrate those instead. Was that your thinking?

	Joao