This looks like another network-namespace regression.
icmp_send() does:

        net = rt->u.dst.dev->nd_net;

The bridge netfilter code attaches a fake dst_entry to the
skb which has dev == NULL when passing it to IPv4 netfilter.

Pavel, do you have a better ideas for fixing this than
instantiating a dst_entry in br_netfilter.c for every
device (or at least for every namespace)?



--

Hm... Why not make this dst entry point to looback device? This would
allow us to make the dst entry per-namespace and instantiate it with
the ns's lo.

Thanks,
Pavel
--

I thought of that myself, but that will result in MTU problems.
The current way is not ideal either (hardcoded MTU of 1500 for
the fake net_device), but that at least works in normal bridge
setups.
--

[CC list trimmed slightly]


The description of the problem was not entirely correct, the
bridge-netfilter dst_entry does have a proper device pointer,
it points to a dummy net_device structure however that doesn't
have a namespace associated with it. This blows up in
__ip_route_output_key.

--

For now I suggest this fix. Andy, could you please test whether it
fixes the problem?

I'm guessing it will.  I've been testing my older kernels today and
somehow I'm unable to reproduce the original problem, so I need to
figure out what is different about my config now and before.

I'll definitely post back when I test this, but I wanted to make sure
you guys knew I was at least trying it. :)

--

I just got done testing 159d83363b629c91d020734207c1bc788b96af5a and I
can confirm that is resolves my issue.
--

Thanks Andy.


--

From: Patrick McHardy <kaber@trash.net>

Even though we don't have a test report back yet, I've applied this,
thanks Patrick.
--