bdschuym@xxxxxxxxxx wrote:
Patrick McHardy wrote:
I can reproduce this with forwarding between two bridges.
The reason is that skb->nf_bridge still contains the data
from the first bridge and so br_netfilter thinks this is
a bridged packet. I don't know how this is supposed to work,
but it seems to me that on packets going out a bridge device
this should be reset in case it originates from a different
bridge (actually I think it should be reset unconditionally
but that would probably break bridged DNAT).
Bart, what do you think about changing this:
(sorry for the webmail mess)
I think that would work. It shouldn't be reset unconditionally at that point since we allow IP dnating of bridged packets (bridged-and-DNAT'ed case).
Could you check the attached patch?
> Another solution I think is this:
in br_nf_post_routing():
change
if (!nf_bridge)
to
if (!nf_bridge || !(nf_bridge->mask & BRNF_BRIDGED_DNAT))
Wouldn't that break the regular case of packets forwarded
through a single bridge?
This regression was introduced when the ip_out sabotage stuff was removed. br_nf_post_routing should now only consider bridged IP packets.
Yes, though the underlying problem seems to be that skb->nf_bridge
has no clearly defined lifetime. We want to pass the bridge port
information up exactly one layer, and then it should disappear.
But that seems to require sprinkling nf_bridge_put in lots of places.
diff --git a/net/bridge/br_netfilter.c b/net/bridge/br_netfilter.c
index da22f90..b7cac8d 100644
--- a/net/bridge/br_netfilter.c
+++ b/net/bridge/br_netfilter.c
@@ -713,8 +713,11 @@ static unsigned int br_nf_local_out(unsigned int hook, struct sk_buff *skb,
return NF_ACCEPT;
nf_bridge = skb->nf_bridge;
- if (!(nf_bridge->mask & BRNF_BRIDGED_DNAT))
+ if (!(nf_bridge->mask & BRNF_BRIDGED_DNAT)) {
+ nf_bridge_put(skb->nf_bridge);
+ skb->nf_bridge = NULL;
return NF_ACCEPT;
+ }
/* Bridged, take PF_BRIDGE/FORWARD.
* (see big note in front of br_nf_pre_routing_finish) */