lkml.org 
[lkml]   [2018]   [Apr]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v1] mm: consider non-anonymous thp as unmovable page
On Thu, Apr 05, 2018 at 10:59:27AM +0200, Michal Hocko wrote:
> On Tue 03-04-18 13:54:11, Kirill A. Shutemov wrote:
> > On Tue, Apr 03, 2018 at 10:34:51AM +0200, Michal Hocko wrote:
> > > On Tue 03-04-18 08:24:06, Naoya Horiguchi wrote:
> > > > On Tue, Apr 03, 2018 at 09:59:28AM +0200, Michal Hocko wrote:
> > > > > On Tue 03-04-18 13:46:28, Naoya Horiguchi wrote:
> > > > > > My testing for the latest kernel supporting thp migration found out an
> > > > > > infinite loop in offlining the memory block that is filled with shmem
> > > > > > thps. We can get out of the loop with a signal, but kernel should
> > > > > > return with failure in this case.
> > > > > >
> > > > > > What happens in the loop is that scan_movable_pages() repeats returning
> > > > > > the same pfn without any progress. That's because page migration always
> > > > > > fails for shmem thps.
> > > > >
> > > > > Why does it fail? Shmem pages should be movable without any issues.
> > > >
> > > > .. because try_to_unmap_one() explicitly skips unmapping for migration.
> > > >
> > > > #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
> > > > /* PMD-mapped THP migration entry */
> > > > if (!pvmw.pte && (flags & TTU_MIGRATION)) {
> > > > VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
> > > >
> > > > if (!PageAnon(page))
> > > > continue;
> > > >
> > > > set_pmd_migration_entry(&pvmw, page);
> > > > continue;
> > > > }
> > > > #endif
> > > >
> > > > When I implemented this code, I felt hard to work on both of anon thp
> > > > and shmem thp at one time, so I separated the proposal into smaller steps.
> > > > Shmem uses pagecache so we need some non-trivial effort (including testing)
> > > > to extend thp migration for shmem. But I think it's a reasonable next step.
> > >
> > > OK, I see. I have forgot about this part. Please be explicit about that
> > > in the changelog. Also the proper fix is to not use movable zone for
> > > shmem page THP rather than hack around it in the hotplug specific code
> > > IMHO.
> >
> > No. We should just split the page before running
> > try_to_unmap(TTU_MIGRATION) on the page.
>
> Something like this or it is completely broken. I completely forgot the
> whole page_vma_mapped_walk business.

No, this wouldn't work. We need to split page, not pmd to make migration
work.

>
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 9eaa6354fe70..cbbfbcb08b83 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -1356,6 +1356,7 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
> return true;
>
> if (flags & TTU_SPLIT_HUGE_PMD) {
> +split:
> split_huge_pmd_address(vma, address,
> flags & TTU_SPLIT_FREEZE, page);
> }
> @@ -1375,7 +1376,7 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
> VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
>
> if (!PageAnon(page))
> - continue;
> + goto split;
>
> set_pmd_migration_entry(&pvmw, page);
> continue;
> --
> Michal Hocko
> SUSE Labs

--
Kirill A. Shutemov

\
 
 \ /
  Last update: 2018-04-05 14:30    [W:0.102 / U:0.216 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site