Netfilter changes PACKET_OTHERHOST to PACKET_HOST before invoking the
hooks as, while it's an expected value for a bridge, routing expects
PACKET_HOST. The change is undone later on after hook traversal. This
can be seen with pairs of functions updating skb>pkt_type and then
reverting it to its original value:

For hook NF_INET_PRE_ROUTING:
  setup_pre_routing / br_nf_pre_routing_finish

For hook NF_INET_FORWARD:
  br_nf_forward_ip / br_nf_forward_finish

But the third case where netfilter does this, for hook
NF_INET_POST_ROUTING, the packet type is changed in br_nf_post_routing
but never reverted. A comment says:

  /* We assume any code from br_dev_queue_push_xmit onwards doesn't care
   * about the value of skb->pkt_type. */

But when having a tunnel (say vxlan) attached to a bridge we have the
following call trace:

  br_nf_pre_routing
  br_nf_pre_routing_ipv6
     br_nf_pre_routing_finish
  br_nf_forward_ip
     br_nf_forward_finish
  br_nf_post_routing           <- pkt_type is updated to PACKET_HOST
     br_nf_dev_queue_xmit      <- but not reverted to its original value
  vxlan_xmit
     vxlan_xmit_one
        skb_tunnel_check_pmtu  <- a check on pkt_type is performed

In this specific case, this creates issues such as when an ICMPv6 PTB
should be sent back. When CONFIG_BRIDGE_NETFILTER is enabled, the PTB
isn't sent (as skb_tunnel_check_pmtu checks if pkt_type is PACKET_HOST
and returns early).

If the comment is right and no one cares about the value of
skb->pkt_type after br_dev_queue_push_xmit (which isn't true), resetting
it to its original value should be safe.

Fixes: 1da177e4 ("Linux-2.6.12-rc2")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Reviewed-by: Florian Westphal <fw@strlen.de>
Link: https://lore.kernel.org/r/20201123174902.622102-1-atenart@kernel.org


Signed-off-by: Jakub Kicinski <kuba@kernel.org>