memcg, shmem: fix shmem migration to use lrucare

It has been reported that 965GM might trigger

  VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage)

in mem_cgroup_migrate when shmem wants to replace a swap cache page
because of shmem_should_replace_page (the page is allocated from an
inappropriate zone).  shmem_replace_page expects that the oldpage is not
on LRU list and calls mem_cgroup_migrate without lrucare.  This is
obviously incorrect because swapcache pages might be on the LRU list
(e.g. swapin readahead page).

Fix this by enabling lrucare for the migration in shmem_replace_page.
Also clarify that lrucare should be used even if one of the pages might
be on LRU list.

The BUG_ON will trigger only when CONFIG_DEBUG_VM is enabled but even
without that the migration code might leave the old page on an
inappropriate memcg' LRU which is not that critical because the page
would get removed with its last reference but it is still confusing.

Fixes: 0a31bc9 ("mm: memcontrol: rewrite uncharge API")
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reported-by: Chris Wilson <chris@chris-wilson.co.uk>
Reported-by: Dave Airlie <airlied@gmail.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org>	[3.17+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memcontrol.c

@@ -5773,7 +5773,7 @@ void mem_cgroup_uncharge_list(struct list_head *page_list)
  * mem_cgroup_migrate - migrate a charge to another page
  * @oldpage: currently charged page
  * @newpage: page to transfer the charge to
- * @lrucare: both pages might be on the LRU already
+ * @lrucare: either or both pages might be on the LRU already
  *
  * Migrate the charge from @oldpage to @newpage.
  *

mm/shmem.c

@@ -1013,7 +1013,7 @@ static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 		 */
 		oldpage = newpage;
 	} else {
-		mem_cgroup_migrate(oldpage, newpage, false);
+		mem_cgroup_migrate(oldpage, newpage, true);
 		lru_cache_add_anon(newpage);
 		*pagep = newpage;
 	}