Re: [PATCH net-next 1/3] udp_tunnel: allow to turn off path mtu discovery on encap sockets

[David Ahern <dsahern@gmail.com>]

On 7/19/20 3:49 PM, Stefano Brivio wrote:
>>
>> With this test case, the lookup fails:
>>
>> [  144.689378] vxlan: vxlan_xmit_one: dev vxlan_a 10.0.1.1/57864 ->
>> 10.0.0.0/4789 len 5010 gw 10.0.1.2
>> [  144.692755] vxlan: skb_tunnel_check_pmtu: dst dev br0 skb dev vxlan_a
>> skb len 5010 encap_mtu 4000 headroom 50
>> [  144.697682] vxlan: skb_dst_update_pmtu_no_confirm: calling
>> ip_rt_update_pmtu+0x0/0x160/ffffffff825ee850 for dev br0 mtu 3950
>> [  144.703601] IPv4: __ip_rt_update_pmtu: dev br0 mtu 3950 old_mtu 5000
>> 192.168.2.1 -> 192.168.2.2
>> [  144.708177] IPv4: __ip_rt_update_pmtu: fib_lookup failed for
>> 192.168.2.1 -> 192.168.2.2
>>
>> Because the lookup fails, __ip_rt_update_pmtu skips creating the exception.
>>
>> This hack gets the lookup to succeed:
>>
>> fl4->flowi4_oif = dst->dev->ifindex;
>> or
>> fl4->flowi4_oif = 0;
> 
> Oh, I didn't consider that... route. :) Here comes an added twist, which
> currently needs Florian's changes from:
> 	https://git.breakpoint.cc/cgit/fw/net-next.git/log/?h=udp_tun_pmtud_12
> 
> Test is as follows:
> 
> test_pmtu_ipv4_vxlan4_exception_bridge() {
> 	test_pmtu_ipvX_over_vxlanY_or_geneveY_exception vxlan  4 4
> 
> 	ip netns add ns-C
> 
> 	ip -n ns-C link add veth_c_a type veth peer name veth_a_c
> 	ip -n ns-C link set veth_a_c netns ns-A
> 
> 	ip -n ns-C addr add 192.168.2.100/24 dev veth_c
> 
> 	ip -n ns-C link set dev veth_c_a mtu 5000
> 	ip -n ns-C link set veth_c_a up
> 	ip -n ns-A link set dev veth_a_c mtu 5000
> 	ip -n ns-A link set veth_c_a up
> 
> 	ip -n ns-A link add br0 type bridge
> 	ip -n ns-A link set br0 up
> 	ip -n ns-A link set dev br0 mtu 5000
> 	ip -n ns-A link set veth_a_c master br0
> 	ip -n ns-A link set vxlan_a master br0
> 
> 	ip -n ns-A addr del 192.168.2.1/24 dev vxlan_a
> 	ip -n ns-A addr add 192.168.2.1/24 dev br0
> 
> 	ip -n ns-C exec ping -c 1 -w 2 -M want -s 5000 192.168.2.2
> }
> 
> I didn't check the test itself recently, I'm just copying from some
> local changes I was trying last week, some commands might be wrong.

I fixed the exec typo, but yes even with my flowi4_oif hack it fails.

> 
> The idea is: what if we now have another host (here, it's ns-C) sending
> traffic to that bridge? Then the exception on a local interface isn't
> enough, we actually need to send Fragmentation Needed back to where the
> packet came from, and the bridge won't do it for us (with routing, it
> already works).
> 
> I haven't tried your hack, but I guess it would have the same problem.
> 

What I saw in my tests and debug statements is that vxlan xmit does
compensate for the tunnel overhead (e.g., skb_tunnel_check_pmtu in
vxlan_xmit_one). It still feels like there are some minor details that
are wrong - like the fib_lookup failing when called from the
vxlan_xmit_one path. Does finding and fixing those make it work vs
adding another config item? I can send my debug diff if it helps.