mm: fix s390 BUG by __set_page_dirty_no_writeback on swap
authorHugh Dickins <>
Mon, 23 Apr 2012 18:14:50 +0000 (11:14 -0700)
committerBen Hutchings <>
Fri, 11 May 2012 12:13:59 +0000 (13:13 +0100)
commit aca50bd3b4c4bb5528a1878158ba7abce41de534 upstream.

Mel reports a BUG_ON(slot == NULL) in radix_tree_tag_set() on s390
3.0.13: called from __set_page_dirty_nobuffers() when page_remove_rmap()
tries to transfer dirty flag from s390 storage key to struct page and

That would be because of reclaim's shrink_page_list() calling
add_to_swap() on this page at the same time: first PageSwapCache is set
(causing page_mapping(page) to appear as &swapper_space), then
page->private set, then tree_lock taken, then page inserted into
radix_tree - so there's an interval before taking the lock when the
radix_tree slot is empty.

We could fix this by moving __add_to_swap_cache()'s spin_lock_irq up
before the SetPageSwapCache.  But a better fix is simply to do what's
five years overdue: Ken Chen introduced __set_page_dirty_no_writeback()
(if !PageDirty TestSetPageDirty) for tmpfs to skip all the radix_tree
overhead, and swap is just the same - it ignores the radix_tree tag, and
does not participate in dirty page accounting, so should be using
__set_page_dirty_no_writeback() too.

s390 testing now confirms that this does indeed fix the problem.

Reported-by: Mel Gorman <>
Signed-off-by: Hugh Dickins <>
Acked-by: Mel Gorman <>
Cc: Andrew Morton <>
Cc: Martin Schwidefsky <>
Cc: Heiko Carstens <>
Cc: Rik van Riel <>
Cc: Ken Chen <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>

index 78cc4d1..7704d9c 100644 (file)
@@ -27,7 +27,7 @@
 static const struct address_space_operations swap_aops = {
        .writepage      = swap_writepage,
-       .set_page_dirty = __set_page_dirty_nobuffers,
+       .set_page_dirty = __set_page_dirty_no_writeback,
        .migratepage    = migrate_page,