The other issue that comes up when this sort of thing is discussed is what
source address do you put on the packets? BSD has traditionally put the
source address associated with the interface, but if it's unnumbered...

Cisco solves this problem by requiring the declaration of another numbered
interface as being "associated" with the unnumbered interface (typically, a
local Ethernet).

The right thing here is a half-router - two routers that act as if the link
between them simply doesn't exist and they're just one router. This is
tricky, though, and has implications for things like traceroute and
pathchar, and the two devices have to agree on how to exchange these
packets. It's also messy for the routing daemons to deal with - they have
to know, and I'm not sure that routed or gated can deal with this sort of
thing; it almost requires the two routers to share state above and beyond
what the protocols allow.

The other place to look into this sort of stuff is the KA9Q TCP/IP code.
Phil Karn tried to make it possible to have exactly one IP address per
node, and had a full "half-router" concept working.

If we're going to go down this road, it should be with eyes wide open, and
with the goal of a completely clean solution that is easy for an admin to
configure and get right. There are lots of things that can go wrong, alas;
I tried to make this kind of thing work with a pair of cisco routers that I
controlled at Apple, and while I made it work, it required a whole lot more
config lashup than I really cared for.

	Erik <fair@clock.org>