On Jul 30, 2009, at 12:45 PM, der Mouse wrote:
In principle, it should not be necessary to have local and remote addresses on a point-to-point link at all.It's necessary if you expect to communicate between the nodes that share the link, since otherwise they have no way to name one another. For routing, addresses of some kind are necessary to tell the routing code where to route packets, though there's no reason in principle it couldn't be set up so as to make some kind of "the other end ofinterface $IFNAME" thing a valid address for routing purposes. For ourpurposes, use of (IP) addresses is compelled by the routing code's insistence on using IP addresses to name route destinations.
It's not conceptually necessary to name both ends; indeed, if I recall correctly Cisco routers are perfectly happy with unnumbered interfaces. What is necessary is is that each node have one (or more) IP addresses, and for the routing tables to point in the proper direction. If I want to talk to host B, all I need to know is "send it down this point-to-point link"; it's not really my concern whether or not it's the host that's directly connected to that end.
Way back when, there was no parameter to route(8) to route to an interface; you had to fake it by giving a 0 metric. The routing subsystem was also organized to route to a gateway address, and all interfaces had to have addresses. Given that, it was natural to require two addresses when ifconfiging, but as you say, in principle it's not necessary.
--Steve Bellovin, http://www.cs.columbia.edu/~smb