Re: [PATCH v2 10/10] nf_conntrack: Use rcu_barrier() and fix kmem_cache_create flags
From: Patrick McHardy <hidden>
Date: 2009-06-24 13:58:46
Also in:
linux-ext4, linux-nfs, linux-wireless, lkml, netfilter-devel
Jesper Dangaard Brouer wrote:
Adjusting SLAB_DESTROY_BY_RCU flags.
kmem_cache_create("nf_conntrack", ...) does not need the
SLAB_DESTROY_BY_RCU flag.It does need it. We're using it instead of call_rcu() for conntracks.
But the
kmem_cache_create("nf_conntrack_expect", ...) should use the
SLAB_DESTROY_BY_RCU flag, because it uses a call_rcu() callback to
invoke kmem_cache_free().No, using call_rcu() means we don't need SLAB_DESTROY_BY_RCU. Please see the note in include/linux/slab.h.
RCU barriers, rcu_barrier(), is inserted two places. In nf_conntrack_expect.c nf_conntrack_expect_fini() before the kmem_cache_destroy(), even though the use of the SLAB_DESTROY_BY_RCU flag, because slub does not (currently) handle rcu sync correctly.
I think that should be fixed in slub then.
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).
This one looks fine.
quoted hunk ↗ jump to hunk
diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c index 5f72b94..438ce84 100644 --- a/net/netfilter/nf_conntrack_core.c +++ b/net/netfilter/nf_conntrack_core.c@@ -1242,7 +1242,7 @@ static int nf_conntrack_init_init_net(void) nf_conntrack_cachep = kmem_cache_create("nf_conntrack", sizeof(struct nf_conn), - 0, SLAB_DESTROY_BY_RCU, NULL); + 0, 0, NULL); if (!nf_conntrack_cachep) { printk(KERN_ERR "Unable to create nf_conn slab cache\n"); ret = -ENOMEM;diff --git a/net/netfilter/nf_conntrack_expect.c b/net/netfilter/nf_conntrack_expect.c index afde8f9..56227c2 100644 --- a/net/netfilter/nf_conntrack_expect.c +++ b/net/netfilter/nf_conntrack_expect.c@@ -593,7 +593,7 @@ int nf_conntrack_expect_init(struct net *net) if (net_eq(net, &init_net)) { nf_ct_expect_cachep = kmem_cache_create("nf_conntrack_expect", sizeof(struct nf_conntrack_expect), - 0, 0, NULL); + 0, SLAB_DESTROY_BY_RCU, NULL); if (!nf_ct_expect_cachep) goto err2; }@@ -617,8 +617,15 @@ 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)) { + /* hawk@comx.dk 2009-06-24: The rcu_barrier() can be + * removed once the sl*b allocators has been fixed + * regarding handling the SLAB_DESTROY_BY_RCU flag + * correctly. + */ + 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); }diff --git a/net/netfilter/nf_conntrack_extend.c b/net/netfilter/nf_conntrack_extend.c index 4b2c769..fef95be 100644 --- a/net/netfilter/nf_conntrack_extend.c +++ b/net/netfilter/nf_conntrack_extend.c@@ -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);