On 2020/10/19 11:47, David Gwynne wrote: > On Sun, Oct 18, 2020 at 08:57:34PM +0100, Stuart Henderson wrote: > > On 2020/10/18 14:04, David Gwynne wrote: > > > the problem i'm hitting is that i have a multihomed box where the > > > service it provides listens on an IP address that's assigned to lo1. > > > it's a host running a service, it's not a router, so the > > > net.inet.ip.forwarding sysctl is not set to 1. > > > > I ran into this, I just turned on the forwarding sysctl to avoid the > > problem. > > > > > i came up with this diff, which adds even more special casing for > > > loopback interfaces. it says addreesses on loopbacks are globally > > > reachable, even if ip forwarding is disabled. > > > > I don't see why loopbacks should be special. Another place this > > might show up is services running on carp addresses (I haven't updated > > those machines yet but there's a fair chance they'll be affected too). > > I would prefer an explicit sysctl to disable "strong host model". > > loopback is already special. if a packet comes from an loopback > interface, we allow it to talk to any IP on the local machine. i think > this is mostly to cope with the semantic we've had where local traffic > get's tied to a loopback interface instead of going anywhere near the > physical ones. > > carp is also special. > > let me paste the ip_laddr function instead of the diff to it, it's a bit > more obvious what's going on:
Thanks, that will already work for the machines I was thinking of then. > back to loopback and receiving packets. loopback is special because it > is not connected to the outside world. it is impossible to send a packet > via a loopback interface from another host, so configuring a globally > (externally) routable IP on it is currently pointless unless you enable > forwarding. i think making loopback more special and allowing it > to be globally reachable makes sense. i can't think of any downsides to > this at the moment, except that the behaviour would be subtle/not > obvious ok, so it makes sense for this to be independent of any possible separate lever. > is there a need to configure a globally reachable IP on a non-loopback > interface on a host (not router)? if so, then i'd be more convinced that > we need a separate lever to pull. I'm not using it this way, but here's a scenario. Say there are a couple of webservers with addresses from a carp on ethernet/vlan, with a link to their upstream router on some separate interface. They announce the carp prefix into ospf. They aren't routing themselves so the only reason to have forwarding=1 is to have them use "weak host model". With forwarding=0 I think they'll have to use "stub router no" otherwise everything will be announced high metric (rather than being dependent on carp state), but ospfd explicitly handles this; it's marked in parse.y with "/* allow to force non stub mode */".
