[c-nsp] PIM convergence problem

[John Neiberger]

We ran into an interesting problem last night and I'm a little stumped. It
appears that PIM did not follow a unicast routing change after a BGP peer
was shutdown. Imagine this simple topology:

[A] ----- [B] ------ [C] -------  [D]
 |
 |
 |
 [D]

Router A is a CRS and is forwarding PIM joins toward Router D, which is
directly attached. We are not running an IGP here. There is only an eBGP
session between two ASes that we manage. We shutdown the BGP session
between A and D, which caused unicast traffic to switch to the path toward
Router B. However, it looks like Router A did not tear down the PIM joins
that are now no longer valid. It seems that it was still joining a lot of
traffic that it could no longer do anything with since it would now fail
RPF checks.

We didn't get snapshots during the event, so I can't prove that is what
happened, but it is the only thing we've found that makes any sense. We've
had quite a few engineers looking at it and we do have TAC on the case, but
I thought I'd check here, as well.

Have any of you seen a situation where PIM joins stay up even when they
shouldn't? Is there possibly an issue with an interaction between PIM and
BGP? I've never seen this sort of behavior before, so I'm not quite sure
what to think.

Thanks,
John