netfilter: nfnetlink_queue: provide rcu enabled callbacks
authorEric Dumazet <eric.dumazet@gmail.com>
Mon, 18 Jul 2011 14:08:27 +0000 (16:08 +0200)
committerPatrick McHardy <kaber@trash.net>
Mon, 18 Jul 2011 14:08:27 +0000 (16:08 +0200)
nenetlink_queue operations on SMP are not efficent if several queues are
used, because of nfnl_mutex contention when applications give packet
verdict.

Use new call_rcu field in struct nfnl_callback to advertize a callback
that is called under rcu_read_lock instead of nfnl_mutex.

On my 2x4x2 machine, I was able to reach 2.000.000 pps going through
user land returning NF_ACCEPT verdicts without losses, instead of less
than 500.000 pps before patch.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Florian Westphal <fw@strlen.de>
CC: Eric Leblond <eric@regit.org>
Signed-off-by: Patrick McHardy <kaber@trash.net>
net/netfilter/nfnetlink_queue.c

index b83123f12b42e30612481796bc6c9a4d3bacab60..c645b87915b845aae8918e4b25f9c302266b2c10 100644 (file)
@@ -619,39 +619,26 @@ nfqnl_recv_verdict(struct sock *ctnl, struct sk_buff *skb,
        struct nfqnl_instance *queue;
        unsigned int verdict;
        struct nf_queue_entry *entry;
-       int err;
 
-       rcu_read_lock();
        queue = instance_lookup(queue_num);
-       if (!queue) {
-               err = -ENODEV;
-               goto err_out_unlock;
-       }
+       if (!queue)
+               return -ENODEV;
 
-       if (queue->peer_pid != NETLINK_CB(skb).pid) {
-               err = -EPERM;
-               goto err_out_unlock;
-       }
+       if (queue->peer_pid != NETLINK_CB(skb).pid)
+               return -EPERM;
 
-       if (!nfqa[NFQA_VERDICT_HDR]) {
-               err = -EINVAL;
-               goto err_out_unlock;
-       }
+       if (!nfqa[NFQA_VERDICT_HDR])
+               return -EINVAL;
 
        vhdr = nla_data(nfqa[NFQA_VERDICT_HDR]);
        verdict = ntohl(vhdr->verdict);
 
-       if ((verdict & NF_VERDICT_MASK) > NF_MAX_VERDICT) {
-               err = -EINVAL;
-               goto err_out_unlock;
-       }
+       if ((verdict & NF_VERDICT_MASK) > NF_MAX_VERDICT)
+               return -EINVAL;
 
        entry = find_dequeue_entry(queue, ntohl(vhdr->id));
-       if (entry == NULL) {
-               err = -ENOENT;
-               goto err_out_unlock;
-       }
-       rcu_read_unlock();
+       if (entry == NULL)
+               return -ENOENT;
 
        if (nfqa[NFQA_PAYLOAD]) {
                if (nfqnl_mangle(nla_data(nfqa[NFQA_PAYLOAD]),
@@ -664,10 +651,6 @@ nfqnl_recv_verdict(struct sock *ctnl, struct sk_buff *skb,
 
        nf_reinject(entry, verdict);
        return 0;
-
-err_out_unlock:
-       rcu_read_unlock();
-       return err;
 }
 
 static int
@@ -780,9 +763,9 @@ err_out_unlock:
 }
 
 static const struct nfnl_callback nfqnl_cb[NFQNL_MSG_MAX] = {
-       [NFQNL_MSG_PACKET]      = { .call = nfqnl_recv_unsupp,
+       [NFQNL_MSG_PACKET]      = { .call_rcu = nfqnl_recv_unsupp,
                                    .attr_count = NFQA_MAX, },
-       [NFQNL_MSG_VERDICT]     = { .call = nfqnl_recv_verdict,
+       [NFQNL_MSG_VERDICT]     = { .call_rcu = nfqnl_recv_verdict,
                                    .attr_count = NFQA_MAX,
                                    .policy = nfqa_verdict_policy },
        [NFQNL_MSG_CONFIG]      = { .call = nfqnl_recv_config,