ip6tnl: avoid touching dst refcount in ip6_tnl_xmit2()
authorEric Dumazet <eric.dumazet@gmail.com>
Thu, 28 Jul 2011 04:32:25 +0000 (04:32 +0000)
committerDavid S. Miller <davem@davemloft.net>
Mon, 1 Aug 2011 07:12:00 +0000 (00:12 -0700)
Even using percpu stats, we still hit tunnel dst_entry refcount in
ip6_tnl_xmit2()

Since we are in RCU locked section, we can use skb_dst_set_noref() and
avoid these atomic operations, leaving dst shared on cpus.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv6/ip6_tunnel.c

index 0bc98886c383c93c79e81a3cd7c4bfe220502b58..6fb1fb3624bf47f33e9960e403ceb8bd4a36368d 100644 (file)
@@ -889,7 +889,7 @@ static int ip6_tnl_xmit2(struct sk_buff *skb,
        struct net_device_stats *stats = &t->dev->stats;
        struct ipv6hdr *ipv6h = ipv6_hdr(skb);
        struct ipv6_tel_txoption opt;
-       struct dst_entry *dst;
+       struct dst_entry *dst, *ndst = NULL;
        struct net_device *tdev;
        int mtu;
        unsigned int max_headroom = sizeof(struct ipv6hdr);
@@ -897,19 +897,19 @@ static int ip6_tnl_xmit2(struct sk_buff *skb,
        int err = -1;
        int pkt_len;
 
-       if ((dst = ip6_tnl_dst_check(t)) != NULL)
-               dst_hold(dst);
-       else {
-               dst = ip6_route_output(net, NULL, fl6);
+       dst = ip6_tnl_dst_check(t);
+       if (!dst) {
+               ndst = ip6_route_output(net, NULL, fl6);
 
-               if (dst->error)
+               if (ndst->error)
                        goto tx_err_link_failure;
-               dst = xfrm_lookup(net, dst, flowi6_to_flowi(fl6), NULL, 0);
-               if (IS_ERR(dst)) {
-                       err = PTR_ERR(dst);
-                       dst = NULL;
+               ndst = xfrm_lookup(net, ndst, flowi6_to_flowi(fl6), NULL, 0);
+               if (IS_ERR(ndst)) {
+                       err = PTR_ERR(ndst);
+                       ndst = NULL;
                        goto tx_err_link_failure;
                }
+               dst = ndst;
        }
 
        tdev = dst->dev;
@@ -955,7 +955,7 @@ static int ip6_tnl_xmit2(struct sk_buff *skb,
                skb = new_skb;
        }
        skb_dst_drop(skb);
-       skb_dst_set(skb, dst_clone(dst));
+       skb_dst_set_noref(skb, dst);
 
        skb->transport_header = skb->network_header;
 
@@ -987,13 +987,14 @@ static int ip6_tnl_xmit2(struct sk_buff *skb,
                stats->tx_errors++;
                stats->tx_aborted_errors++;
        }
-       ip6_tnl_dst_store(t, dst);
+       if (ndst)
+               ip6_tnl_dst_store(t, ndst);
        return 0;
 tx_err_link_failure:
        stats->tx_carrier_errors++;
        dst_link_failure(skb);
 tx_err_dst_release:
-       dst_release(dst);
+       dst_release(ndst);
        return err;
 }