Okay, I'll try to hit all the questions.... :-) Take a look at temp.gif, attached. Now, if the A to B and A to C links are equal cost, no matter what their cost is:
-- B will mark A as the successor, and will not mark C as a feasible successor, because the RD == FD.
-- C will mark A as the successor, and will not mark B as an FS, because the RD == FD.
The problem is you can't prove, mathematically, that the path is loop free unless the RD < FD, really. Suppose B and C could point at each other as feasible successors. What happens if 10.1.1.0/24 itself fails? B will immediately flip over to using C as its path, and C will immediately flip over to using B as its path--a loop.
The loop would actually be permanent. B would send an update out towards A saying it has a new path through C, but not towards C, since C is its current path to 10.1.1.0/24 (split horizon rule, never send an update out the path you're using). C would do the same thing, and thus, the loop would sit there until the link between B and C fails because of some high volume traffic flow. :-)
There are a couple of ways around this. For instance, We could use poison reverse instead of split horizon, I think.... We could make it so when you see an RD == FD, you will send out a poison reverse when you flip to that route, so if it's a loop, it's killed pretty fast. But, you'll still have a transient loop. It might also be possible to do some sort of exploritory query in this case, but it's hard to figure out exactly how such a beast would work. So, we've explored some ideas in this space, but we don't have a lot of demand to fix this sort of thing.
What happens today is if the A->B links fails, for instance, B will go active, send a query to C, and C is going to answer with an ack that its path still good. You won't get subsecond convergence in this case (unless your routers are really fast, lightly loaded, and the link between then is really fast/lightly loaded as well), but you shouldn't ever encounter really long convergence, either, because the query range is one hop.
It 10.1.1.0/24 itself fails, B and C should query each other, and work it out in short order. The actual chain of events is going to depend on the timing of things, most likely one of the two is going to get the query and process it before the other (from A), since packets are serialized, etc....
For the second question--if the route isn't an FS, we won't ever install it in the routing table, no matter what the variance is. DUAL uses a "negative test," really. We can prove a route isn't a loop, but we can't prove it is a loop. So, we're conservative, and if we can't prove it isn't a loop, we assume it is a loop. Since we can't know any route with RD >= FD isn't a loop, we just won't use it....
:-)
Russ
