vmscan: shrinker->nr updates race and go wrong
authorDave Chinner <dchinner@redhat.com>
Fri, 8 Jul 2011 04:14:35 +0000 (14:14 +1000)
committerAl Viro <viro@zeniv.linux.org.uk>
Wed, 20 Jul 2011 05:44:29 +0000 (01:44 -0400)
shrink_slab() allows shrinkers to be called in parallel so the
struct shrinker can be updated concurrently. It does not provide any
exclusio for such updates, so we can get the shrinker->nr value
increasing or decreasing incorrectly.

As a result, when a shrinker repeatedly returns a value of -1 (e.g.
a VFS shrinker called w/ GFP_NOFS), the shrinker->nr goes haywire,
sometimes updating with the scan count that wasn't used, sometimes
losing it altogether. Worse is when a shrinker does work and that
update is lost due to racy updates, which means the shrinker will do
the work again!

Fix this by making the total_scan calculations independent of
shrinker->nr, and making the shrinker->nr updates atomic w.r.t. to
other updates via cmpxchg loops.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
mm/vmscan.c

index 255226554a3d15fc050fd8a58e1b43fed5538ace..2f7c6ae8fae07c23c0e539ef97f756448943920e 100644 (file)
@@ -251,17 +251,29 @@ unsigned long shrink_slab(struct shrink_control *shrink,
                unsigned long total_scan;
                unsigned long max_pass;
                int shrink_ret = 0;
+               long nr;
+               long new_nr;
 
+               /*
+                * copy the current shrinker scan count into a local variable
+                * and zero it so that other concurrent shrinker invocations
+                * don't also do this scanning work.
+                */
+               do {
+                       nr = shrinker->nr;
+               } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
+
+               total_scan = nr;
                max_pass = do_shrinker_shrink(shrinker, shrink, 0);
                delta = (4 * nr_pages_scanned) / shrinker->seeks;
                delta *= max_pass;
                do_div(delta, lru_pages + 1);
-               shrinker->nr += delta;
-               if (shrinker->nr < 0) {
+               total_scan += delta;
+               if (total_scan < 0) {
                        printk(KERN_ERR "shrink_slab: %pF negative objects to "
                               "delete nr=%ld\n",
-                              shrinker->shrink, shrinker->nr);
-                       shrinker->nr = max_pass;
+                              shrinker->shrink, total_scan);
+                       total_scan = max_pass;
                }
 
                /*
@@ -269,13 +281,10 @@ unsigned long shrink_slab(struct shrink_control *shrink,
                 * never try to free more than twice the estimate number of
                 * freeable entries.
                 */
-               if (shrinker->nr > max_pass * 2)
-                       shrinker->nr = max_pass * 2;
-
-               total_scan = shrinker->nr;
-               shrinker->nr = 0;
+               if (total_scan > max_pass * 2)
+                       total_scan = max_pass * 2;
 
-               trace_mm_shrink_slab_start(shrinker, shrink, total_scan,
+               trace_mm_shrink_slab_start(shrinker, shrink, nr,
                                        nr_pages_scanned, lru_pages,
                                        max_pass, delta, total_scan);
 
@@ -296,9 +305,19 @@ unsigned long shrink_slab(struct shrink_control *shrink,
                        cond_resched();
                }
 
-               shrinker->nr += total_scan;
-               trace_mm_shrink_slab_end(shrinker, shrink_ret, total_scan,
-                                        shrinker->nr);
+               /*
+                * move the unused scan count back into the shrinker in a
+                * manner that handles concurrent updates. If we exhausted the
+                * scan, there is no need to do an update.
+                */
+               do {
+                       nr = shrinker->nr;
+                       new_nr = total_scan + nr;
+                       if (total_scan <= 0)
+                               break;
+               } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
+
+               trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
        }
        up_read(&shrinker_rwsem);
 out: