PMTUD broken inside network namespace with multipath routing

Hi,

I have observed that PMTUD (Path MTU discovery) is broken using
multipath routing inside a network namespace. This breaks TCP, because
it keeps trying to send oversized packets.
Observed on kernel 5.4.44, other kernels weren't tested. However i
went through net/ipv4/route.c and haven't spotted changes in this
area, so i believe this bug is still there.

Host test with multipath routing:
---------------------------------
root@host1:~# ip route add 192.168.247.100/32 dev vmbr2 nexthop via
192.168.252.250 dev vmbr2 nexthop via 192.168.252.252 dev vmbr2
root@host1:~# ip route | grep -A2 192.168.247.100
192.168.247.100
 nexthop via 192.168.252.250 dev vmbr2 weight 1
 nexthop via 192.168.252.252 dev vmbr2 weight 1
root@host1:~# ping -M do -s 1380 192.168.247.100
PING 192.168.247.100 (192.168.247.100) 1380(1408) bytes of data.
From 192.168.252.252 icmp_seq=1 Frag needed and DF set (mtu = 1406)
ping: local error: Message too long, mtu=1406
ping: local error: Message too long, mtu=1406
ping: local error: Message too long, mtu=1406
ping: local error: Message too long, mtu=1406
^C

--- 192.168.247.100 ping statistics ---

5 packets transmitted, 0 received, +5 errors, 100% packet loss, time 80ms
root@host1:~# ip route get 192.168.247.100
192.168.247.100 via 192.168.252.250 dev vmbr2 src 192.168.252.15 uid 0
    cache expires 583sec mtu 1406

LXC container inside that host with multipath routing:
------------------------------------------------------
[root@lxctest ~]# ip route add 192.168.247.100/32 dev eth0 nexthop via
192.168.252.250 dev eth0 nexthop via 192.168.252.252 dev eth0
[root@lxctest ~]# ip route
default via 192.168.252.100 dev eth0 proto static metric 100
192.168.247.100
 nexthop via 192.168.252.250 dev eth0 weight 1
 nexthop via 192.168.252.252 dev eth0 weight 1
192.168.252.0/24 dev eth0 proto kernel scope link src 192.168.252.207 metric 100
[root@lxctest ~]# ping -M do -s 1380 192.168.247.100
PING 192.168.247.100 (192.168.247.100) 1380(1408) bytes of data.
From 192.168.252.252 icmp_seq=1 Frag needed and DF set (mtu = 1406)
From 192.168.252.252 icmp_seq=2 Frag needed and DF set (mtu = 1406)
From 192.168.252.252 icmp_seq=3 Frag needed and DF set (mtu = 1406)
From 192.168.252.252 icmp_seq=4 Frag needed and DF set (mtu = 1406)
[root@lxctest ~]# ip route get 192.168.247.100
192.168.247.100 via 192.168.252.252 dev eth0 src 192.168.252.207 uid 0
    cache

LXC container inside that host with regular routing:
----------------------------------------------------
[root@lxctest ~]# ip route add 192.168.247.100/32 via 192.168.252.252 dev eth0
[root@lxctest ~]# ip route
default via 192.168.252.100 dev eth0 proto static metric 100
192.168.247.100 via 192.168.252.252 dev eth0
192.168.252.0/24 dev eth0 proto kernel scope link src 192.168.252.207 metric 100
[root@lxctest ~]# ping -M do -s 1380 192.168.247.100
PING 192.168.247.100 (192.168.247.100) 1380(1408) bytes of data.
From 192.168.252.252 icmp_seq=1 Frag needed and DF set (mtu = 1406)
ping: local error: Message too long, mtu=1406
ping: local error: Message too long, mtu=1406
ping: local error: Message too long, mtu=1406
^C

--- 192.168.247.100 ping statistics ---

4 packets transmitted, 0 received, +4 errors, 100% packet loss, time 82ms
[root@lxctest ~]# ip route get 192.168.247.100
192.168.247.100 via 192.168.252.252 dev eth0 src 192.168.252.207 uid 0
    cache expires 591sec mtu 1406


What seems to be happening, is that when multipath routing is used
inside LXC (or any network namespace), the kernel doesn't generate a
routing exception to force the lower MTU.
I believe this is a bug inside the kernel.


Kfir Itzhak