nf_conntrack: Use rcu_barrier()
authorJesper Dangaard Brouer <hawk@comx.dk>
Thu, 25 Jun 2009 14:32:52 +0000 (16:32 +0200)
committerPatrick McHardy <kaber@trash.net>
Thu, 25 Jun 2009 14:32:52 +0000 (16:32 +0200)
RCU barriers, rcu_barrier(), is inserted two places.

 In nf_conntrack_expect.c nf_conntrack_expect_fini() before the
 kmem_cache_destroy().  Firstly to make sure the callback to the
 nf_ct_expect_free_rcu() code is still around.  Secondly because I'm
 unsure about the consequence of having in flight
 nf_ct_expect_free_rcu/kmem_cache_free() calls while doing a
 kmem_cache_destroy() slab destroy.

 And in nf_conntrack_extend.c nf_ct_extend_unregister(), inorder to
 wait for completion of callbacks to __nf_ct_ext_free_rcu(), which is
 invoked by __nf_ct_ext_add().  It might be more efficient to call
 rcu_barrier() in nf_conntrack_core.c nf_conntrack_cleanup_net(), but
 thats make it more difficult to read the code (as the callback code
 in located in nf_conntrack_extend.c).

Signed-off-by: Jesper Dangaard Brouer <hawk@comx.dk>
Signed-off-by: Patrick McHardy <kaber@trash.net>
net/netfilter/nf_conntrack_expect.c
net/netfilter/nf_conntrack_extend.c

index afde8f9..2032dfe 100644 (file)
@@ -617,8 +617,10 @@ err1:
 void nf_conntrack_expect_fini(struct net *net)
 {
        exp_proc_remove(net);
-       if (net_eq(net, &init_net))
+       if (net_eq(net, &init_net)) {
+               rcu_barrier(); /* Wait for call_rcu() before destroy */
                kmem_cache_destroy(nf_ct_expect_cachep);
+       }
        nf_ct_free_hashtable(net->ct.expect_hash, net->ct.expect_vmalloc,
                             nf_ct_expect_hsize);
 }
index 4b2c769..fef95be 100644 (file)
@@ -186,6 +186,6 @@ void nf_ct_extend_unregister(struct nf_ct_ext_type *type)
        rcu_assign_pointer(nf_ct_ext_types[type->id], NULL);
        update_alloc_size(type);
        mutex_unlock(&nf_ct_ext_type_mutex);
-       synchronize_rcu();
+       rcu_barrier(); /* Wait for completion of call_rcu()'s */
 }
 EXPORT_SYMBOL_GPL(nf_ct_extend_unregister);