[PATCH] slab: Fix kmem_cache_destroy() on NUMA

[Roland Dreier <rdreier@cisco.com>]

With CONFIG_NUMA set, kmem_cache_destroy() may fail and say "Can't
free all objects."  The problem is caused by sequences such as the
following (suppose we are on a NUMA machine with two nodes, 0 and 1):

 * Allocate an object from cache on node 0.
 * Free the object on node 1.  The object is put into node 1's alien
   array_cache for node 0.
 * Call kmem_cache_destroy(), which ultimately ends up in __cache_shrink().
 * __cache_shrink() does drain_cpu_caches(), which loops through all nodes.
   For each node it drains the shared array_cache and then handles the
   alien array_cache for the other node.

However this means that node 0's shared array_cache will be drained,
and then node 1 will move the contents of its alien[0] array_cache
into that same shared array_cache.  node 0's shared array_cache is
never looked at again, so the objects left there will appear to be in
use when __cache_shrink() calls __node_shrink() for node 0.  So
__node_shrink() will return 1 and kmem_cache_destroy() will fail.

This patch fixes this by having drain_cpu_caches() do
drain_alien_cache() on every node before it does drain_array() on the
nodes' shared array_caches.

The problem was originally reported by Or Gerlitz <ogerlitz@voltaire.com>.

Cc: Christoph Lameter <clameter@sgi.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>

Signed-off-by: Roland Dreier <rolandd@cisco.com>

---

I get a nervous feeling about touching NUMA slab code, because just
the topic alone makes it sound hairy.  But I think my diagnosis and
fix are pretty clear, and this definitely fixes crashes seen when
unloading IB modules.  It's a regression from 2.6.16, and x86_64
machines with > 1 NUMA node are quite common, so this probably should
go into 2.6.17.

diff --git a/mm/slab.c b/mm/slab.c
index c32af7e..cb747be 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -2192,11 +2192,14 @@ static void drain_cpu_caches(struct kmem
 	check_irq_on();
 	for_each_online_node(node) {
 		l3 = cachep->nodelists[node];
-		if (l3) {
+		if (l3 && l3->alien)
+			drain_alien_cache(cachep, l3->alien);
+	}
+
+	for_each_online_node(node) {
+		l3 = cachep->nodelists[node];
+		if (l3)
 			drain_array(cachep, l3, l3->shared, 1, node);
-			if (l3->alien)
-				drain_alien_cache(cachep, l3->alien);
-		}
 	}
 }