[PATCH] mm: make nr_scan_try_batch() more safe on races

[Wu Fengguang <fengguang.wu@intel.com>]

nr_scan_try_batch() can be called concurrently on the same zone
and the non-atomic calculations can go wrong. This is not a big
problem as long as the errors are small and won't impact the
balanced zone aging noticeably.

@nr_to_scan could be much larger values than @swap_cluster_max.
So don't store such large values to *nr_saved_scan directly,
which helps reducing possible errors on races.

CC: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
---
 mm/vmscan.c |    9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

--- linux.orig/mm/vmscan.c	2009-08-20 10:46:30.000000000 +0800
+++ linux/mm/vmscan.c	2009-08-20 10:49:36.000000000 +0800
@@ -1496,15 +1496,14 @@ static unsigned long nr_scan_try_batch(u
 				       unsigned long *nr_saved_scan,
 				       unsigned long swap_cluster_max)
 {
-	unsigned long nr;
-
-	*nr_saved_scan += nr_to_scan;
-	nr = *nr_saved_scan;
+	unsigned long nr = *nr_saved_scan + nr_to_scan;
 
 	if (nr >= swap_cluster_max)
 		*nr_saved_scan = 0;
-	else
+	else {
+		*nr_saved_scan = nr;
 		nr = 0;
+	}
 
 	return nr;
 }