On Thu, 3 Aug 2023, Kefeng Wang wrote: > On 2023/8/2 20:30, Matthew Wilcox wrote: > > On Wed, Aug 02, 2023 at 05:53:44PM +0800, Kefeng Wang wrote: > >> /* Do not migrate THP mapped by multiple processes */ > >> - if (PageTransHuge(page) && total_mapcount(page) > 1) > >> + if (folio_test_pmd_mappable(folio) && folio_estimated_sharers(folio) > > >> 1) > >> return 0; > > > > I don't know if this is the right logic. We've willing to move folios > > mapped by multiple processes, as long as they're smaller than PMD size, > > but once they get to PMD size they're magical and can't be moved? > > It seems that the logical is introduced by commit 04fa5d6a6547 ("mm: > migrate: check page_count of THP before migrating") and refactor by > 340ef3902cf2 ("mm: numa: cleanup flow of transhuge page migration"), > > > "Hugh Dickins pointed out that migrate_misplaced_transhuge_page() does > not check page_count before migrating like base page migration and > khugepage. He could not see why this was safe and he is right." > > For now, there is no migrate_misplaced_transhuge_page() and base/thp > page migrate's path is unified, there is a check(for old/new kernel) in > migrate_misplaced_page(), Right, Mel's comment on safety above comes from a time when migrate_misplaced_transhuge_page() went its own way, and so did not reach the careful reference count checking found in (the now named) folio_migrate_mapping(). If migrate_misplaced_page() is now using the standard migrate_pages() for small pages and THPs, then there should no longer be safety concerns. > > "Don't migrate file pages that are mapped in multiple processes > with execute permissions as they are probably shared libraries." > > We could drop the above check in numamigrate_isolate_page(), but > according to 04fa5d6a6547, maybe disable migrate page shared by > multi-process during numa balance for both base/thp page. I'm out of touch with the NUMA balancing preferences at present, but think that you're probably right that it should aim to treat folio mapped into multiple processes the same way for small and large and THP (except, of course, that large and THP, with multiple PTEs in the same process, can present more challenges to how to go about that). Hugh