BIRDv2 OSPF: Stub for loopback potentially broken: Invalid Prefix in LSA

[Joakim Tjernlund]

On Tue, 2019-04-23 at 19:33 +0200, Ondrej Zajicek wrote:
> 
> On Tue, Apr 23, 2019 at 12:40:04PM +0000, Joakim Tjernlund wrote:
> > > I think the standard just does not consider the case of 'unnumbered'
> > > link with both local and remote address but without subnet and assumes
> > > 'real' unnumbered PtP link with no local IP address associated with
> > > the iface.
> > 
> > There is always some local address, in linux you have to assign it to the I/F,
> > but in others, like Cisco, you can assign an IP address to a dummy I/F and then
> > tell unnumbered I/Fs to use the dummy I/Fs' IP address.
> > There has to be some SRC IP address on pkgs sent by OSPF
> 
> Yes, but the way how the RFC is written it (IMHO) assumes the second case -
> unnumbered ifaces using SRC IP from some dummy iface. Therefore stub

yes, the second case, agreed.

> node is associated with the dummy iface and not with the unnumbered PtP iface.

No, this is a true unnumbered I/F(it has no local IP address) and will not generate
any stub route.

> 
> 
> > Since Linux always has an local IP adress it is not possible to deduce if the
> > user intended the link to be unnumbered or not, unless one wants to add explicit
> > config "unnumbered", it would be best to assume unnumbered I think.
> 
> I do not understand what you mean in this article. Linux could have active iface
> without any IP address and it can be used for PtP link (with SRC IP from some
> other iface), but it is not implemented in BIRD.

I expressed myself badly, I meant if one want to have control over what SRC ip is used,
you need to have it on the local I/F. Maybe there is a way to do that these days differently
but I don't think any PtoP impl. on Linux uses that. 

> 
> It is true that even a link with local /32 IP without specified remote peer could
> be used for OSPF PtP link. BIRD in this case assumes it is a stub /32 and i am not
> sure whether it could be configured as regular PtP. I did not realized before that
> this is a relevant case for PtP link.

This case is as close one gets to a true unnumbered I/F in Linux I think and pppd supports
such links. I think should treat /32 links as unnumbered even if there is an remote
IP address or make it configurable.

> 
> > Consider the use case with many /32 ptp links, all with the same local IP, there would
> > be a lot of redundant host routes in the Router LSA. Better to let the user
> > add an explicit stub network for all PtoP's if one needs it.
> 
> Well, that is solvable by not putting the same stub multiple times in the
> Router LSA.

Maybe but since it is the remote ip one puts here you will only save "space" when
you have multiple ptop links between the same two routers.

> 
> More debatable is a case like this:
> 
> eth0 10.0.0.1/24
> eth1 10.0.0.1/32 peer 10.0.1.1/32
> 
> 10.0.0.1 is already reachable by route for eth0 (either stub or network),
> so it is technically not necessary to have it in the Router LSA as
> a separate /32 stub. So it makes sense to me to have it configurable
> with three values: yes, no, if-not-covered.

Yes, this is what I was getting at too. Maybe have a conf option "unnumbered" to treat
eth1 as a "true" unnumbred I/F(no stub network).
Quietion is, what should be default? 
What is default in current 2.0 Bird?

 Jocke