On Thu, May 05, 2022 at 10:00:07AM -0700, Mike Kravetz wrote: > On 5/4/22 23:48, Minchan Kim wrote: > > On Wed, May 04, 2022 at 03:48:54PM -0700, Minchan Kim wrote: > >> On Tue, May 03, 2022 at 06:02:33PM +0200, David Hildenbrand wrote: > >>> On 03.05.22 17:26, Minchan Kim wrote: > >>>> On Tue, May 03, 2022 at 03:15:24AM +0200, David Hildenbrand wrote: > >>>>> > >>>>>>>>> However, I assume we have the same issue right now already with > >>>>>>> ZONE_MOVABLE and MIGRATE_CMA when trying to pin a page residing on these > >>>>>> > >>>>>> ZONE_MOVALBE is also changed dynamically? > >>>>>> > >>>>> > >>>>> Sorry, with "same issue" I meant failing to pin if having to migrate and > >>>>> the page is temporarily unmovable. > >>>>> > >>>>>>> there are temporarily unmovable and we fail to migrate. But it would now > >>>>>>> apply even without ZONE_MOVABLE or MIGRATE_CMA. Hm... > >>>>>> > >>>>>> Didn't parse your last mention. > >>>>> > >>>>> On a system that neither uses ZONE_MOVABLE nor MIGRATE_CMA we might have > >>>>> to migrate now when pinning. > >>>> > >>>> I don't understand your point. My problem is pin_user_pages with > >>>> FOLL_LONGTERM. It shouldn't pin a page from ZONE_MOVABLE and cma area > >>>> without migrating page out of movable zone or CMA area. > >>>> That's why try_grab_folio checks whether target page stays in those > >>>> movable areas. However, to check CMA area, is_migrate_cma_page is > >>>> racy so the FOLL_LONGTERM flag semantic is broken right now. > >>>> > >>>> Do you see any problem of the fix? > >>> > >>> My point is that you might decide to migrate a page because you stumble > >>> over MIGRATE_ISOLATE, although there is no need to reject long-term > >>> pinning and to trigger page migration. > >>> > >>> Assume a system without ZONE_MOVABLE and without MIGRATE_CMA. Assume > >>> someone reserves gigantic pages (alloc_contig_range()) and you have > >>> concurrent long-term pinning on a page that is no MIGRATE_ISOLATE. > >>> > >>> GUP would see MIGRATE_ISOLATE and would reject pinning. The page has to > >>> be migrated, which can fail if the page is temporarily unmovable. > >> > >> A dump question since I'm not familiar with hugetlb. > >> > >> Is above reasonable scenario? > >> > >> The gigantic page is about to be created using alloc_contig_range so > >> they has MIGRATE_ISOLATE as temporal state. It means no one uses the > >> page yet so I guess the page is not mapped at userspace but other is > >> trying to access the page using pin_user_pages? > >> > > > > Too dump question. Never mind. > > Posted v2 - https://lore.kernel.org/all/20220505064429.2818496-1-minchan@xxxxxxxxxx/T/#u > > > > Well your question mentioned hugetlb so my mail filters caught it :) > > Your question caused me to think of the following. No need for any immediate > change: I think. Just wanted to share. > > Suppose someone has reserved CMA for gigantic hugetlb allocations. And, > suppose FOLL_LONGTERM is attempted on such a page (it would be in use). The > desired action would be to migrate the page out of CMA. Correct? > > Gigantic pages can only be migrated IF there is another (already allocated) > gigantic page available. The routine to try and allocate a page 'on the fly' > for migration will fail if passed a gigantic size. There 'might' be a free > pre-allocated gigantic page. However, if the user set up CMA reserves for > gigantic page allocations it is likely the free gigantic page is also in CMA. > Therefore, it can not be used for this migration. So, unless my reasoning > is wrong, FOLL_LONGTERM would almost always fail for gigantic pages in CMA. FOLL_LONGTERM on CMA-backed gigantic page would already fail, Thanks for sharing! Anyway, David's concern was non-CMA-backed gigantic page. The alloc_contig_range with MIGRATE_ISOLATE runs with concurrent FOLL_LONGTERM pinning, which could trigger page migration we didn't have before so it might increase FOLL_LONGTERM GUP failure rate.
