mm: fix s390 BUG by __set_page_dirty_no_writeback on swap

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
radix_tree.

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 <[email protected]>
Signed-off-by: Hugh Dickins <[email protected]>
Acked-by: Mel Gorman <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: Martin Schwidefsky <[email protected]>
Cc: Heiko Carstens <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Ken Chen <[email protected]>
Cc: [email protected]
Signed-off-by: Linus Torvalds <[email protected]>

mm/swap_state.c

@@ -26,7 +26,7 @@
  */
 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,
 };