vmscan: let GFP_NOFS go to swap again
Hugh Dickins [Wed, 19 Nov 2008 23:36:37 +0000 (15:36 -0800)]
In the past, GFP_NOFS (but of course not GFP_NOIO) was allowed to reclaim
by writing to swap.  That got partially broken in 2.6.23, when may_enter_fs
initialization was moved up before the allocation of swap, so its
PageSwapCache test was failing the first time around,

Fix it by setting may_enter_fs when add_to_swap() succeeds with
__GFP_IO.  In fact, check __GFP_IO before calling add_to_swap():
allocating swap we're not ready to use just increases disk seeking.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/vmscan.c

index c141b3e..f83a7ed 100644 (file)
@@ -623,6 +623,8 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                 * Try to allocate it some swap space here.
                 */
                if (PageAnon(page) && !PageSwapCache(page)) {
+                       if (!(sc->gfp_mask & __GFP_IO))
+                               goto keep_locked;
                        switch (try_to_munlock(page)) {
                        case SWAP_FAIL:         /* shouldn't happen */
                        case SWAP_AGAIN:
@@ -634,6 +636,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                        }
                        if (!add_to_swap(page, GFP_ATOMIC))
                                goto activate_locked;
+                       may_enter_fs = 1;
                }
 #endif /* CONFIG_SWAP */