Re: 2.6.17-rc4-mm1

[Eric Dumazet]

Andrew Morton a écrit :
> Eric Dumazet <dada1@xxxxxxxxxxxxx> wrote:
>   
> > Hi Andrew
> >
> > It seems latest kernels have a problem in kmem_cache_destroy()
> >     
>
> Mainline, or just -mm?
>
>   
Mainline and mm it seems, only if NUMA is ON, on 2.6.17 only (2.6.16.XX 
is OK)

I added some printks and it seems slab_partials is not empty in 
__node_shrink() for node 0

I am not a mm/slab.c expert but the following patch cures the problem 
for me :

[PATCH] slab : NUMA may need __cache_shrink() doing two loops.

Signed-off-by: Eric Dumazet <dada1@xxxxxxxxxxxxx>




--- linux-2.6.17-rc4/mm/slab.c	2006-05-15 14:58:09.000000000 +0200
+++ linux-2.6.17-rc4-ed/mm/slab.c	2006-05-15 15:23:07.000000000 +0200
@@ -2225,25 +2225,32 @@
 		spin_lock_irq(&l3->list_lock);
 	}
 	ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
+	if (ret)
+		printk(KERN_ERR "__node_shrink(name=%s,node=%d) ret=%d\n",
+			cachep->name, node, ret);
 	return ret;
 }
 
 static int __cache_shrink(struct kmem_cache *cachep)
 {
-	int ret = 0, i = 0;
+	int ret, i;
 	struct kmem_list3 *l3;
+	int loopcount = 0;
 
-	drain_cpu_caches(cachep);
+	do {
+		ret = 0;
+		drain_cpu_caches(cachep);
 
-	check_irq_on();
-	for_each_online_node(i) {
-		l3 = cachep->nodelists[i];
-		if (l3) {
-			spin_lock_irq(&l3->list_lock);
-			ret += __node_shrink(cachep, i);
-			spin_unlock_irq(&l3->list_lock);
+		check_irq_on();
+		for_each_online_node(i) {
+			l3 = cachep->nodelists[i];
+			if (l3) {
+				spin_lock_irq(&l3->list_lock);
+				ret += __node_shrink(cachep, i);
+				spin_unlock_irq(&l3->list_lock);
+			}
 		}
-	}
+	} while (ret && ++loopcount < 2);
 	return (ret ? 1 : 0);
 }