mm: stop kswapd's infinite loop at high order allocation
KOSAKI Motohiro [Tue, 6 Jan 2009 22:40:33 +0000 (14:40 -0800)]
Wassim Dagash reported following kswapd infinite loop problem.

  kswapd runs in some infinite loop trying to swap until order 10 of zone
  highmem is OK.... kswapd will continue to try to balance order 10 of zone
  highmem forever (or until someone release a very large chunk of highmem).

For non order-0 allocations, the system may never be balanced due to
fragmentation but kswapd should not infinitely loop as a result.

Instead, recheck all watermarks at order-0 as they are the most important.
If watermarks are ok, kswapd will go back to sleep.

[akpm@linux-foundation.org: fix comment]
Reported-by: wassim dagash <wassim.dagash@gmail.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/vmscan.c

index 5daf606..b07c48b 100644 (file)
@@ -1867,6 +1867,23 @@ out:
 
                try_to_freeze();
 
+               /*
+                * Fragmentation may mean that the system cannot be
+                * rebalanced for high-order allocations in all zones.
+                * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
+                * it means the zones have been fully scanned and are still
+                * not balanced. For high-order allocations, there is
+                * little point trying all over again as kswapd may
+                * infinite loop.
+                *
+                * Instead, recheck all watermarks at order-0 as they
+                * are the most important. If watermarks are ok, kswapd will go
+                * back to sleep. High-order users can still perform direct
+                * reclaim if they wish.
+                */
+               if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
+                       order = sc.order = 0;
+
                goto loop_again;
        }