udp: Fix rcv socket locking
authorHerbert Xu <herbert@gondor.apana.org.au>
Mon, 15 Sep 2008 18:48:46 +0000 (11:48 -0700)
committerDavid S. Miller <davem@davemloft.net>
Mon, 15 Sep 2008 18:48:46 +0000 (11:48 -0700)
The previous patch in response to the recursive locking on IPsec
reception is broken as it tries to drop the BH socket lock while in
user context.

This patch fixes it by shrinking the section protected by the
socket lock to sock_queue_rcv_skb only.  The only reason we added
the lock is for the accounting which happens in that function.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/udp.c

index 8e42fbbd57616316df2692a39379398f9486ac22..57e26fa66185affd152d8613800ee5a9e6230991 100644 (file)
@@ -951,6 +951,27 @@ int udp_disconnect(struct sock *sk, int flags)
        return 0;
 }
 
+static int __udp_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
+{
+       int is_udplite = IS_UDPLITE(sk);
+       int rc;
+
+       if ((rc = sock_queue_rcv_skb(sk, skb)) < 0) {
+               /* Note that an ENOMEM error is charged twice */
+               if (rc == -ENOMEM)
+                       UDP_INC_STATS_BH(sock_net(sk), UDP_MIB_RCVBUFERRORS,
+                                        is_udplite);
+               goto drop;
+       }
+
+       return 0;
+
+drop:
+       UDP_INC_STATS_BH(sock_net(sk), UDP_MIB_INERRORS, is_udplite);
+       kfree_skb(skb);
+       return -1;
+}
+
 /* returns:
  *  -1: error
  *   0: success
@@ -989,9 +1010,7 @@ int udp_queue_rcv_skb(struct sock * sk, struct sk_buff *skb)
                    up->encap_rcv != NULL) {
                        int ret;
 
-                       bh_unlock_sock(sk);
                        ret = (*up->encap_rcv)(sk, skb);
-                       bh_lock_sock(sk);
                        if (ret <= 0) {
                                UDP_INC_STATS_BH(sock_net(sk),
                                                 UDP_MIB_INDATAGRAMS,
@@ -1044,17 +1063,16 @@ int udp_queue_rcv_skb(struct sock * sk, struct sk_buff *skb)
                        goto drop;
        }
 
-       if ((rc = sock_queue_rcv_skb(sk,skb)) < 0) {
-               /* Note that an ENOMEM error is charged twice */
-               if (rc == -ENOMEM) {
-                       UDP_INC_STATS_BH(sock_net(sk),
-                                       UDP_MIB_RCVBUFERRORS, is_udplite);
-                       atomic_inc(&sk->sk_drops);
-               }
-               goto drop;
-       }
+       rc = 0;
 
-       return 0;
+       bh_lock_sock(sk);
+       if (!sock_owned_by_user(sk))
+               rc = __udp_queue_rcv_skb(sk, skb);
+       else
+               sk_add_backlog(sk, skb);
+       bh_unlock_sock(sk);
+
+       return rc;
 
 drop:
        UDP_INC_STATS_BH(sock_net(sk), UDP_MIB_INERRORS, is_udplite);
@@ -1092,15 +1110,7 @@ static int __udp4_lib_mcast_deliver(struct net *net, struct sk_buff *skb,
                                skb1 = skb_clone(skb, GFP_ATOMIC);
 
                        if (skb1) {
-                               int ret = 0;
-
-                               bh_lock_sock(sk);
-                               if (!sock_owned_by_user(sk))
-                                       ret = udp_queue_rcv_skb(sk, skb1);
-                               else
-                                       sk_add_backlog(sk, skb1);
-                               bh_unlock_sock(sk);
-
+                               int ret = udp_queue_rcv_skb(sk, skb1);
                                if (ret > 0)
                                        /* we should probably re-process instead
                                         * of dropping packets here. */
@@ -1195,13 +1205,7 @@ int __udp4_lib_rcv(struct sk_buff *skb, struct hlist_head udptable[],
                        uh->dest, inet_iif(skb), udptable);
 
        if (sk != NULL) {
-               int ret = 0;
-               bh_lock_sock(sk);
-               if (!sock_owned_by_user(sk))
-                       ret = udp_queue_rcv_skb(sk, skb);
-               else
-                       sk_add_backlog(sk, skb);
-               bh_unlock_sock(sk);
+               int ret = udp_queue_rcv_skb(sk, skb);
                sock_put(sk);
 
                /* a return value > 0 means to resubmit the input, but
@@ -1494,7 +1498,7 @@ struct proto udp_prot = {
        .sendmsg           = udp_sendmsg,
        .recvmsg           = udp_recvmsg,
        .sendpage          = udp_sendpage,
-       .backlog_rcv       = udp_queue_rcv_skb,
+       .backlog_rcv       = __udp_queue_rcv_skb,
        .hash              = udp_lib_hash,
        .unhash            = udp_lib_unhash,
        .get_port          = udp_v4_get_port,