[c-nsp] bgp vpls rfc4761 - no failover at multihomed ce site

[Aaron]

After working for ~5 hours on the phone with Cisco TAC yesterday (a team of
engineers were working on this one, some XR engineers, some IOS engineers),
they told me yesterday that they are seeing in their notes and their
findings that RFC4761 Section 3.5 (Multi-homing) isn't supported on the
ME3600. (I'm running 15.4(2)S - which appears to be most recent release)
...we tested failover and it didn't work.

Keep in mind, my test architecture closely mimics my real world desired
deployment scenario, of which we tried during a maintenance window last week
and melted the network with bad l2 forwarding loops...(that was using
RFC4762 LDP Sig VPLS).  Which is, p net has 9k and (2) me's in a triangle,
off the 9k I have a pc, off the (2) me's I have (2) occam (nka, calix b
series) dslams.  The calix dslams do not provice l2 loop prevention on their
tagged gige uplinks, hence my desires to stop loops in my sp vpls
architecture with 4761 section 3.5.  Off the occams I have (2) other pc's.

While working with the TAC we saw that multi-homing appears to work and does
stop loops that 4762 won't*, but, if/when one of the pe's on the multihomed
site has a pe-ce handoff phy int/svi int/vfi context go down, there is no
bgp withdraw or mac withdraw (whatever is supposed to happen), sent from
that pe to the others to signal that this side of the multihomed site is
down and start using the other pe.  It wasn't until I completely removed the
vfi context from the configuration that we see a bgp withdraw message go out
and tell other pe's that this NLRI is no longer reachable via this pe....
then we see connectivity restored.  So it seems that the problem is that the
AC going down isn't communicated to the bgp process in order to signal a
withdraw of that l2vpn vpls nlri.

Please share if you know anything more about this.

Aaron

* as I understand it, the way that 4761 stops loops is by simply using a
common ve-id across pe's that share a common L2 customer site.