udp: Fix udp_poll() and ioctl()
authorEric Dumazet <eric.dumazet@gmail.com>
Fri, 9 Oct 2009 04:43:40 +0000 (04:43 +0000)
committerDavid S. Miller <davem@davemloft.net>
Tue, 13 Oct 2009 10:16:54 +0000 (03:16 -0700)
udp_poll() can in some circumstances drop frames with incorrect checksums.

Problem is we now have to lock the socket while dropping frames, or risk
sk_forward corruption.

This bug is present since commit 95766fff6b9a78d1
([UDP]: Add memory accounting.)

While we are at it, we can correct ioctl(SIOCINQ) to also drop bad frames.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/udp.c

index 6ec6a8a..d0d436d 100644 (file)
@@ -841,6 +841,42 @@ out:
        return ret;
 }
 
+
+/**
+ *     first_packet_length     - return length of first packet in receive queue
+ *     @sk: socket
+ *
+ *     Drops all bad checksum frames, until a valid one is found.
+ *     Returns the length of found skb, or 0 if none is found.
+ */
+static unsigned int first_packet_length(struct sock *sk)
+{
+       struct sk_buff_head list_kill, *rcvq = &sk->sk_receive_queue;
+       struct sk_buff *skb;
+       unsigned int res;
+
+       __skb_queue_head_init(&list_kill);
+
+       spin_lock_bh(&rcvq->lock);
+       while ((skb = skb_peek(rcvq)) != NULL &&
+               udp_lib_checksum_complete(skb)) {
+               UDP_INC_STATS_BH(sock_net(sk), UDP_MIB_INERRORS,
+                                IS_UDPLITE(sk));
+               __skb_unlink(skb, rcvq);
+               __skb_queue_tail(&list_kill, skb);
+       }
+       res = skb ? skb->len : 0;
+       spin_unlock_bh(&rcvq->lock);
+
+       if (!skb_queue_empty(&list_kill)) {
+               lock_sock(sk);
+               __skb_queue_purge(&list_kill);
+               sk_mem_reclaim_partial(sk);
+               release_sock(sk);
+       }
+       return res;
+}
+
 /*
  *     IOCTL requests applicable to the UDP protocol
  */
@@ -857,21 +893,16 @@ int udp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 
        case SIOCINQ:
        {
-               struct sk_buff *skb;
-               unsigned long amount;
+               unsigned int amount = first_packet_length(sk);
 
-               amount = 0;
-               spin_lock_bh(&sk->sk_receive_queue.lock);
-               skb = skb_peek(&sk->sk_receive_queue);
-               if (skb != NULL) {
+               if (amount)
                        /*
                         * We will only return the amount
                         * of this packet since that is all
                         * that will be read.
                         */
-                       amount = skb->len - sizeof(struct udphdr);
-               }
-               spin_unlock_bh(&sk->sk_receive_queue.lock);
+                       amount -= sizeof(struct udphdr);
+
                return put_user(amount, (int __user *)arg);
        }
 
@@ -1540,29 +1571,11 @@ unsigned int udp_poll(struct file *file, struct socket *sock, poll_table *wait)
 {
        unsigned int mask = datagram_poll(file, sock, wait);
        struct sock *sk = sock->sk;
-       int     is_lite = IS_UDPLITE(sk);
 
        /* Check for false positives due to checksum errors */
-       if ((mask & POLLRDNORM) &&
-           !(file->f_flags & O_NONBLOCK) &&
-           !(sk->sk_shutdown & RCV_SHUTDOWN)) {
-               struct sk_buff_head *rcvq = &sk->sk_receive_queue;
-               struct sk_buff *skb;
-
-               spin_lock_bh(&rcvq->lock);
-               while ((skb = skb_peek(rcvq)) != NULL &&
-                      udp_lib_checksum_complete(skb)) {
-                       UDP_INC_STATS_BH(sock_net(sk),
-                                       UDP_MIB_INERRORS, is_lite);
-                       __skb_unlink(skb, rcvq);
-                       kfree_skb(skb);
-               }
-               spin_unlock_bh(&rcvq->lock);
-
-               /* nothing to see, move along */
-               if (skb == NULL)
-                       mask &= ~(POLLIN | POLLRDNORM);
-       }
+       if ((mask & POLLRDNORM) && !(file->f_flags & O_NONBLOCK) &&
+           !(sk->sk_shutdown & RCV_SHUTDOWN) && !first_packet_length(sk))
+               mask &= ~(POLLIN | POLLRDNORM);
 
        return mask;