mm, thp: fix migration of PTE-mapped transparent huge pages
authorKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Wed, 9 Mar 2016 22:08:30 +0000 (14:08 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 9 Mar 2016 23:43:42 +0000 (15:43 -0800)
We don't have native support of THP migration, so we have to split huge
page into small pages in order to migrate it to different node.  This
includes PTE-mapped huge pages.

I made mistake in refcounting patchset: we don't actually split
PTE-mapped huge page in queue_pages_pte_range(), if we step on head
page.

The result is that the head page is queued for migration, but none of
tail pages: putting head page on queue takes pin on the page and any
subsequent attempts of split_huge_pages() would fail and we skip queuing
tail pages.

unmap_and_move_huge_page() will eventually split the huge pages, but
only one of 512 pages would get migrated.

Let's fix the situation.

Fixes: 248db92da13f2507 ("migrate_pages: try to split pages on queuing")
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/mempolicy.c

index 4c4187c0e1deeb25bdbf5eb383132d27feb9be90..9a3f6b90e6283b24bff091213768612b04b6d928 100644 (file)
@@ -532,7 +532,7 @@ retry:
                nid = page_to_nid(page);
                if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
                        continue;
-               if (PageTail(page) && PageAnon(page)) {
+               if (PageTransCompound(page) && PageAnon(page)) {
                        get_page(page);
                        pte_unmap_unlock(pte, ptl);
                        lock_page(page);