Generally all OSs will if sending from a local process will use the
address of the outgoing interface for the packet.
If the packet is forwarded and no NAT is used the address will be routed
via the interface suggested by the routing table.
So local routing can be a real pain, policy based routing is an option.
The other option could be to setup an 'output' NAT to an address which
is multi-homed.
I have a system running which is multi-homed with out issue other than
the actual routing machine. This machine is BGP connected to three
locations.
There is no NAT setup and because I also add the wireguard link
addresses to the BGP sessions.
Cheers
On 19/2/2023 6:44 am, Nico Schottelius wrote:
Dear group,
I was wondering how wireguard [Linux kernel] or wireguard-go [FreeBSD]
are supposed to decide which IP address to use for replying?
I have seen both on FreeBSD and Linux that wireguard seems to use the IP
address of the outgoing interface, i.e. the one with the route returning
to the sender. However in multi homed situations, this can be wrong,
let's take this example:
19:57:24.607526 net1 In IP 194.5.220.43.60770 > 147.78.195.254.51820:
UDP, length 148
19:57:24.608358 net2 Out IP 195.141.200.73.51820 > 194.5.220.43.60770:
UDP, length 92
The initiator sends from 194.5.220.43 to the receiver 147.78.195.254.
Wireguard then replies with the source IP of 195.141.200.73 instead of
147.78.195.254.
As the node is multi homed, the packet might leave through any of its
uplinks and thus return with a random (unexpected) IP address and will
not pass NAT rules on firewalls and finally be dropped. F.i. in above
example the firewall drops the packet from 195.141.200.73, because there
is no session entry for that.
I have observed this behaviour both on Linux 6.1.11 as well as
wireguard-go 0.0.20220316_8,1 on FreeBSD and in both cases the
connection will break depending on which active interface is taken as
exit.
I would argue that wireguard should by default invert the IP
addresses, i.e. switch dst=src, src=dst and then reply with that,
instead of adapting an interface specific address, or is there a good
reason for the current behaviour?
Best regards,
Nico
--
Sustainable and modern Infrastructures by ungleich.ch
