packet: fix use after free race in send path when dev is released
authorDaniel Borkmann <dborkman@redhat.com>
Thu, 21 Nov 2013 15:50:58 +0000 (16:50 +0100)
committerDavid S. Miller <davem@davemloft.net>
Thu, 21 Nov 2013 18:09:43 +0000 (13:09 -0500)
Salam reported a use after free bug in PF_PACKET that occurs when
we're sending out frames on a socket bound device and suddenly the
net device is being unregistered. It appears that commit 827d9780
introduced a possible race condition between {t,}packet_snd() and
packet_notifier(). In the case of a bound socket, packet_notifier()
can drop the last reference to the net_device and {t,}packet_snd()
might end up suddenly sending a packet over a freed net_device.

To avoid reverting 827d9780 and thus introducing a performance
regression compared to the current state of things, we decided to
hold a cached RCU protected pointer to the net device and maintain
it on write side via bind spin_lock protected register_prot_hook()
and __unregister_prot_hook() calls.

In {t,}packet_snd() path, we access this pointer under rcu_read_lock
through packet_cached_dev_get() that holds reference to the device
to prevent it from being freed through packet_notifier() while
we're in send path. This is okay to do as dev_put()/dev_hold() are
per-cpu counters, so this should not be a performance issue. Also,
the code simplifies a bit as we don't need need_rls_dev anymore.

Fixes: 827d978037d7 ("af-packet: Use existing netdev reference for bound sockets.")
Reported-by: Salam Noureddine <noureddine@aristanetworks.com>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Salam Noureddine <noureddine@aristanetworks.com>
Cc: Ben Greear <greearb@candelatech.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/packet/af_packet.c
net/packet/internal.h

index 61bd50adead1cc4211040cc21eae589953a27e1f..ac27c86ef6d11e00c2ecb1512b09bbda73c3eefc 100644 (file)
@@ -244,11 +244,15 @@ static void __fanout_link(struct sock *sk, struct packet_sock *po);
 static void register_prot_hook(struct sock *sk)
 {
        struct packet_sock *po = pkt_sk(sk);
+
        if (!po->running) {
-               if (po->fanout)
+               if (po->fanout) {
                        __fanout_link(sk, po);
-               else
+               } else {
                        dev_add_pack(&po->prot_hook);
+                       rcu_assign_pointer(po->cached_dev, po->prot_hook.dev);
+               }
+
                sock_hold(sk);
                po->running = 1;
        }
@@ -266,10 +270,13 @@ static void __unregister_prot_hook(struct sock *sk, bool sync)
        struct packet_sock *po = pkt_sk(sk);
 
        po->running = 0;
-       if (po->fanout)
+       if (po->fanout) {
                __fanout_unlink(sk, po);
-       else
+       } else {
                __dev_remove_pack(&po->prot_hook);
+               RCU_INIT_POINTER(po->cached_dev, NULL);
+       }
+
        __sock_put(sk);
 
        if (sync) {
@@ -2052,12 +2059,24 @@ static int tpacket_fill_skb(struct packet_sock *po, struct sk_buff *skb,
        return tp_len;
 }
 
+static struct net_device *packet_cached_dev_get(struct packet_sock *po)
+{
+       struct net_device *dev;
+
+       rcu_read_lock();
+       dev = rcu_dereference(po->cached_dev);
+       if (dev)
+               dev_hold(dev);
+       rcu_read_unlock();
+
+       return dev;
+}
+
 static int tpacket_snd(struct packet_sock *po, struct msghdr *msg)
 {
        struct sk_buff *skb;
        struct net_device *dev;
        __be16 proto;
-       bool need_rls_dev = false;
        int err, reserve = 0;
        void *ph;
        struct sockaddr_ll *saddr = (struct sockaddr_ll *)msg->msg_name;
@@ -2070,7 +2089,7 @@ static int tpacket_snd(struct packet_sock *po, struct msghdr *msg)
        mutex_lock(&po->pg_vec_lock);
 
        if (saddr == NULL) {
-               dev = po->prot_hook.dev;
+               dev     = packet_cached_dev_get(po);
                proto   = po->num;
                addr    = NULL;
        } else {
@@ -2084,19 +2103,17 @@ static int tpacket_snd(struct packet_sock *po, struct msghdr *msg)
                proto   = saddr->sll_protocol;
                addr    = saddr->sll_addr;
                dev = dev_get_by_index(sock_net(&po->sk), saddr->sll_ifindex);
-               need_rls_dev = true;
        }
 
        err = -ENXIO;
        if (unlikely(dev == NULL))
                goto out;
-
-       reserve = dev->hard_header_len;
-
        err = -ENETDOWN;
        if (unlikely(!(dev->flags & IFF_UP)))
                goto out_put;
 
+       reserve = dev->hard_header_len;
+
        size_max = po->tx_ring.frame_size
                - (po->tp_hdrlen - sizeof(struct sockaddr_ll));
 
@@ -2173,8 +2190,7 @@ out_status:
        __packet_set_status(po, ph, status);
        kfree_skb(skb);
 out_put:
-       if (need_rls_dev)
-               dev_put(dev);
+       dev_put(dev);
 out:
        mutex_unlock(&po->pg_vec_lock);
        return err;
@@ -2212,7 +2228,6 @@ static int packet_snd(struct socket *sock,
        struct sk_buff *skb;
        struct net_device *dev;
        __be16 proto;
-       bool need_rls_dev = false;
        unsigned char *addr;
        int err, reserve = 0;
        struct virtio_net_hdr vnet_hdr = { 0 };
@@ -2228,7 +2243,7 @@ static int packet_snd(struct socket *sock,
         */
 
        if (saddr == NULL) {
-               dev = po->prot_hook.dev;
+               dev     = packet_cached_dev_get(po);
                proto   = po->num;
                addr    = NULL;
        } else {
@@ -2240,19 +2255,17 @@ static int packet_snd(struct socket *sock,
                proto   = saddr->sll_protocol;
                addr    = saddr->sll_addr;
                dev = dev_get_by_index(sock_net(sk), saddr->sll_ifindex);
-               need_rls_dev = true;
        }
 
        err = -ENXIO;
-       if (dev == NULL)
+       if (unlikely(dev == NULL))
                goto out_unlock;
-       if (sock->type == SOCK_RAW)
-               reserve = dev->hard_header_len;
-
        err = -ENETDOWN;
-       if (!(dev->flags & IFF_UP))
+       if (unlikely(!(dev->flags & IFF_UP)))
                goto out_unlock;
 
+       if (sock->type == SOCK_RAW)
+               reserve = dev->hard_header_len;
        if (po->has_vnet_hdr) {
                vnet_hdr_len = sizeof(vnet_hdr);
 
@@ -2386,15 +2399,14 @@ static int packet_snd(struct socket *sock,
        if (err > 0 && (err = net_xmit_errno(err)) != 0)
                goto out_unlock;
 
-       if (need_rls_dev)
-               dev_put(dev);
+       dev_put(dev);
 
        return len;
 
 out_free:
        kfree_skb(skb);
 out_unlock:
-       if (dev && need_rls_dev)
+       if (dev)
                dev_put(dev);
 out:
        return err;
@@ -2614,6 +2626,7 @@ static int packet_create(struct net *net, struct socket *sock, int protocol,
        po = pkt_sk(sk);
        sk->sk_family = PF_PACKET;
        po->num = proto;
+       RCU_INIT_POINTER(po->cached_dev, NULL);
 
        sk->sk_destruct = packet_sock_destruct;
        sk_refcnt_debug_inc(sk);
index c4e4b4561207354c49c9797dd5ea405867de0bcd..1035fa2d909c7f18c100266c85ec07bce4a73a97 100644 (file)
@@ -113,6 +113,7 @@ struct packet_sock {
        unsigned int            tp_loss:1;
        unsigned int            tp_tx_has_off:1;
        unsigned int            tp_tstamp;
+       struct net_device __rcu *cached_dev;
        struct packet_type      prot_hook ____cacheline_aligned_in_smp;
 };