mm: fix race between mremap and removing migration entry

I don't usually pay much attention to the stale "? " addresses in
stack backtraces, but this lucky report from Pawel Sikora hints that
mremap's move_ptes() has inadequate locking against page migration.

 3.0 BUG_ON(!PageLocked(p)) in migration_entry_to_page():
 kernel BUG at include/linux/swapops.h:105!
 RIP: 0010:[<ffffffff81127b76>]  [<ffffffff81127b76>]
                       migration_entry_wait+0x156/0x160
  [<ffffffff811016a1>] handle_pte_fault+0xae1/0xaf0
  [<ffffffff810feee2>] ? __pte_alloc+0x42/0x120
  [<ffffffff8112c26b>] ? do_huge_pmd_anonymous_page+0xab/0x310
  [<ffffffff81102a31>] handle_mm_fault+0x181/0x310
  [<ffffffff81106097>] ? vma_adjust+0x537/0x570
  [<ffffffff81424bed>] do_page_fault+0x11d/0x4e0
  [<ffffffff81109a05>] ? do_mremap+0x2d5/0x570
  [<ffffffff81421d5f>] page_fault+0x1f/0x30

mremap's down_write of mmap_sem, together with i_mmap_mutex or lock,
and pagetable locks, were good enough before page migration (with its
requirement that every migration entry be found) came in, and enough
while migration always held mmap_sem; but not enough nowadays, when
there's memory hotremove and compaction.

The danger is that move_ptes() lets a migration entry dodge around
behind remove_migration_pte()'s back, so it's in the old location when
looking at the new, then in the new location when looking at the old.

Either mremap's move_ptes() must additionally take anon_vma lock(), or
migration's remove_migration_pte() must stop peeking for is_swap_entry()
before it takes pagetable lock.

Consensus chooses the latter: we prefer to add overhead to migration
than to mremapping, which gets used by JVMs and by exec stack setup.

Reported-and-tested-by: Paweł Sikora <[email protected]>
Signed-off-by: Hugh Dickins <[email protected]>
Acked-by: Andrea Arcangeli <[email protected]>
Acked-by: Mel Gorman <[email protected]>
Cc: [email protected]
Signed-off-by: Linus Torvalds <[email protected]>

Author: Hugh Dickins

mm/migrate.c

@@ -120,10 +120,10 @@ static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
 
 		ptep = pte_offset_map(pmd, addr);
 
-		if (!is_swap_pte(*ptep)) {
-			pte_unmap(ptep);
-			goto out;
-		}
+		/*
+		 * Peek to check is_swap_pte() before taking ptlock?  No, we
+		 * can race mremap's move_ptes(), which skips anon_vma lock.
+		 */
 
 		ptl = pte_lockptr(mm, pmd);
 	}