net: bridge: fix per-port af_packet sockets
authorNikolay Aleksandrov <nikolay@cumulusnetworks.com>
Thu, 11 Apr 2019 10:56:39 +0000 (13:56 +0300)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 27 Apr 2019 07:35:33 +0000 (09:35 +0200)
[ Upstream commit 3b2e2904deb314cc77a2192f506f2fd44e3d10d0 ]

When the commit below was introduced it changed two visible things:
 - the skb was no longer passed through the protocol handlers with the
   original device
 - the skb was passed up the stack with skb->dev = bridge

The first change broke af_packet sockets on bridge ports. For example we
use them for hostapd which listens for ETH_P_PAE packets on the ports.
We discussed two possible fixes:
 - create a clone and pass it through NF_HOOK(), act on the original skb
   based on the result
 - somehow signal to the caller from the okfn() that it was called,
   meaning the skb is ok to be passed, which this patch is trying to
   implement via returning 1 from the bridge link-local okfn()

Note that we rely on the fact that NF_QUEUE/STOLEN would return 0 and
drop/error would return < 0 thus the okfn() is called only when the
return was 1, so we signal to the caller that it was called by preserving
the return value from nf_hook().

Fixes: 8626c56c8279 ("bridge: fix potential use-after-free when hook returns QUEUE or STOLEN verdict")
Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
net/bridge/br_input.c

index 7637f58c12263bae0bc907d01d0838c1fd68cf5c..10fa84056cb5286fff78256836e02a311ad55b13 100644 (file)
@@ -236,13 +236,10 @@ static void __br_handle_local_finish(struct sk_buff *skb)
 /* note: already called with rcu_read_lock */
 static int br_handle_local_finish(struct net *net, struct sock *sk, struct sk_buff *skb)
 {
-       struct net_bridge_port *p = br_port_get_rcu(skb->dev);
-
        __br_handle_local_finish(skb);
 
-       BR_INPUT_SKB_CB(skb)->brdev = p->br->dev;
-       br_pass_frame_up(skb);
-       return 0;
+       /* return 1 to signal the okfn() was called so it's ok to use the skb */
+       return 1;
 }
 
 /*
@@ -318,10 +315,18 @@ rx_handler_result_t br_handle_frame(struct sk_buff **pskb)
                                goto forward;
                }
 
-               /* Deliver packet to local host only */
-               NF_HOOK(NFPROTO_BRIDGE, NF_BR_LOCAL_IN, dev_net(skb->dev),
-                       NULL, skb, skb->dev, NULL, br_handle_local_finish);
-               return RX_HANDLER_CONSUMED;
+               /* The else clause should be hit when nf_hook():
+                *   - returns < 0 (drop/error)
+                *   - returns = 0 (stolen/nf_queue)
+                * Thus return 1 from the okfn() to signal the skb is ok to pass
+                */
+               if (NF_HOOK(NFPROTO_BRIDGE, NF_BR_LOCAL_IN,
+                           dev_net(skb->dev), NULL, skb, skb->dev, NULL,
+                           br_handle_local_finish) == 1) {
+                       return RX_HANDLER_PASS;
+               } else {
+                       return RX_HANDLER_CONSUMED;
+               }
        }
 
 forward: