[PATCH net] bridge: Fix incorrect re-injection of STP packets

Commit 8626c56c8279 ("bridge: fix potential use-after-free when hook
returns QUEUE or STOLEN verdict") fixed incorrect usage of NF_HOOK's
return value by consuming packets in okfn via br_pass_frame_up().

However, this function re-injects packets to the Rx path with skb->dev
set to the bridge device, which breaks kernel's STP, as all STP packets
appear to originate from the bridge device itself.

Instead, if okfn was called for a packet, make bridge's rx_handler
re-inject it to the Rx path by returning RX_HANDLER_PASS. This is
consistent with previous behavior.

Cc: Florian Westphal <fw@strlen.de>
Fixes: 8626c56c8279 ("bridge: fix potential use-after-free when hook returns QUEUE or STOLEN verdict")
Reviewed-by: Jiri Pirko <redacted>
Signed-off-by: Ido Schimmel <redacted>
---
 net/bridge/br_input.c | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/net/bridge/br_input.c b/net/bridge/br_input.c
index 1607977..c73ed44 100644
--- a/net/bridge/br_input.c
+++ b/net/bridge/br_input.c

@@ -223,9 +223,7 @@ static int br_handle_local_finish(struct net *net, struct sock *sk, struct sk_bu
 	if (p->flags & BR_LEARNING && br_should_learn(p, skb, &vid))
 		br_fdb_update(p->br, p, eth_hdr(skb)->h_source, vid, false);
 
-	BR_INPUT_SKB_CB(skb)->brdev = p->br->dev;
-	br_pass_frame_up(skb);
-	return 0;
+	return RX_HANDLER_PASS;
 }
 
 /*

@@ -238,6 +236,7 @@ rx_handler_result_t br_handle_frame(struct sk_buff **pskb)
 	struct sk_buff *skb = *pskb;
 	const unsigned char *dest = eth_hdr(skb)->h_dest;
 	br_should_route_hook_t *rhook;
+	int err;
 
 	if (unlikely(skb->pkt_type == PACKET_LOOPBACK))
 		return RX_HANDLER_PASS;

@@ -287,8 +286,11 @@ rx_handler_result_t br_handle_frame(struct sk_buff **pskb)
 		}
 
 		/* 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);
+		err = NF_HOOK(NFPROTO_BRIDGE, NF_BR_LOCAL_IN, dev_net(skb->dev),
+			      NULL, skb, skb->dev, NULL,
+			      br_handle_local_finish);
+		if (err == RX_HANDLER_PASS)
+			return RX_HANDLER_PASS;
 		return RX_HANDLER_CONSUMED;
 	}
 

-- 
2.4.3

Hi,

On Fri, 3 Jun 2016 12:39:45 +0300 Ido Schimmel [off-list ref] wrote:

Seems '*pskb = skb' is needed prior returning RX_HANDLER_PASS - there's
a 'skb = skb_share_check()' at beginning of br_handle_frame.

(The *pskb = skb was present prior 8626c56c8279, but gone since there
was no longer an RX_HANDLER_PASS return).

One nit to consider:

The fix relies on the fact that RX_HANDLER_PASS != 0 (otherwise we end up
not knowing whether skb was STOLEN or br_handle_local_finish has
executed, which was the original problem 8626c56c8279 tried to address).

No reason to use the RX_HANDLER_xxx enumeration space as the ret code
of an 'okfn' (br_handle_local_finish in this case).
Some positive value locally defined in br_input.c (and documented in
br_handle_local_finish) would do.

Regards,
Shmulik

Sat, Jun 04, 2016 at 09:41:41AM IDT, shmulik.ladkani@gmail.com wrote:

Hi,

On Fri, 3 Jun 2016 12:39:45 +0300 Ido Schimmel [off-list ref] wrote:

quoted

diff --git a/net/bridge/br_input.c b/net/bridge/br_input.c
index 1607977..c73ed44 100644
--- a/net/bridge/br_input.c
+++ b/net/bridge/br_input.c

@@ -223,9 +223,7 @@ static int br_handle_local_finish(struct net *net, struct sock *sk, struct sk_bu
 	if (p->flags & BR_LEARNING && br_should_learn(p, skb, &vid))
 		br_fdb_update(p->br, p, eth_hdr(skb)->h_source, vid, false);
 
-	BR_INPUT_SKB_CB(skb)->brdev = p->br->dev;
-	br_pass_frame_up(skb);
-	return 0;
+	return RX_HANDLER_PASS;
 }
 
 /*

@@ -238,6 +236,7 @@ rx_handler_result_t br_handle_frame(struct sk_buff **pskb)
 	struct sk_buff *skb = *pskb;
 	const unsigned char *dest = eth_hdr(skb)->h_dest;
 	br_should_route_hook_t *rhook;
+	int err;
 
 	if (unlikely(skb->pkt_type == PACKET_LOOPBACK))
 		return RX_HANDLER_PASS;

@@ -287,8 +286,11 @@ rx_handler_result_t br_handle_frame(struct sk_buff **pskb)
 		}
 
 		/* 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);
+		err = NF_HOOK(NFPROTO_BRIDGE, NF_BR_LOCAL_IN, dev_net(skb->dev),
+			      NULL, skb, skb->dev, NULL,
+			      br_handle_local_finish);
+		if (err == RX_HANDLER_PASS)
+			return RX_HANDLER_PASS;
 		return RX_HANDLER_CONSUMED;

Seems '*pskb = skb' is needed prior returning RX_HANDLER_PASS - there's
a 'skb = skb_share_check()' at beginning of br_handle_frame.

(The *pskb = skb was present prior 8626c56c8279, but gone since there
was no longer an RX_HANDLER_PASS return).

Right! Will fix that in v2.

One nit to consider:

The fix relies on the fact that RX_HANDLER_PASS != 0 (otherwise we end up
not knowing whether skb was STOLEN or br_handle_local_finish has
executed, which was the original problem 8626c56c8279 tried to address).

No reason to use the RX_HANDLER_xxx enumeration space as the ret code
of an 'okfn' (br_handle_local_finish in this case).
Some positive value locally defined in br_input.c (and documented in
br_handle_local_finish) would do.

Yes, I agree it would be clearer to state that explicitly. Will add that
in v2.

Thanks for reviewing!

Hi Florian,

Fri, Jun 03, 2016 at 12:39:45PM IDT, idosch@mellanox.com wrote:

Commit 8626c56c8279 ("bridge: fix potential use-after-free when hook
returns QUEUE or STOLEN verdict") fixed incorrect usage of NF_HOOK's
return value by consuming packets in okfn via br_pass_frame_up().

However, this function re-injects packets to the Rx path with skb->dev
set to the bridge device, which breaks kernel's STP, as all STP packets
appear to originate from the bridge device itself.

Instead, if okfn was called for a packet, make bridge's rx_handler
re-inject it to the Rx path by returning RX_HANDLER_PASS. This is
consistent with previous behavior.

Cc: Florian Westphal <fw@strlen.de>
Fixes: 8626c56c8279 ("bridge: fix potential use-after-free when hook returns QUEUE or STOLEN verdict")
Reviewed-by: Jiri Pirko <redacted>
Signed-off-by: Ido Schimmel <redacted>

I read your commit more closely and also looked at nf_reinject() and I'm
not sure how the fix should be carried out.

After packet is processed by okfn it should return to the Rx path -
by making bridge's rx_handler return RX_HANDLER_PASS - so that it 
could be picked up by the packet handlers.

However, if verdict is NF_QUEUE and packet is later re-injected via
nf_reinject() then this can't happen, as rx_handler already returned
RX_HANDLER_CONSUMED for the packet.

So, my patch is wrong because it doesn't consume the packet in okfn, but
yours consumes it in a way which breaks current packet handlers.

Any suggestions regarding a fix? I have a feeling I'm missing something
obvious.

Thanks!