On Mon, 7 Sept 2026 at 15:05, Tore Anderson via NANOG <nanog@lists.nanog.org> wrote:
Any ICMPv4/v6 errors can be sourced from the loopback interface, as you say.
We loose interface information in traceroutes if use loopbacks for ICMP errors. I'd immagine people have different opinions about whether this is a problem or not. Personally, since I'm often in the position to troubleshoot problems in other networks, I would rather have this information at hand. With a loopback return IP I cannot even see different ECMP paths if the nodes are the same. I understand that IPv4 has a cost. But using a loopbacks for ICMP errors arguably also does; it's just someone else that is picking up the tab, when troubleshooting issues in the network that is saving a buck. Networks have NOT redistributed public point-to-point prefixes in their IGP for decades to avoid attacks against infrastructure IPs. That doesn't break traceroute. Using a loopback on the other hand takes away a lot of useful information from traceroutes. For this reason I would suggest - for IPv6 at least - to use global unicast adressing on point-to-point links, as oppossed to non routable prefixes. Again that doesn't mean the /127 needs to be actually routed/redistributed. For IPv4 money will drive that decision anyway, although if I'm thinking about cross connect fees in carrier hotels I have a hard time seeing how a public /31 would be a showstopper. lukas