[c-nsp] MPLS VPN over mGRE

[John Neiberger]

I was reading through the configuration guide for MPLS VPN over mGRE to try
to reverse engineer a configuration we have at work. This kind of hurts my
head, but I think I've almost got it. The method we use is basically the
same as this:

http://www.cisco.com/en/US/docs/ios/interface/configuration/guide/ir_mplsvpnomgre.html

The config basically consists of:

* VRF with RD and import/export RTs
* l3vpn encapsulation method using GRE as the protocol
* VPNv4 peer relationships between all endpoints needing access to this VPN
* An ingress route policy on the VPNv4 peers that set the ip next-hop
encapsulation to the l3vpn encapsulation method configured earlier

That seems to be about it. The thing I don't yet understand is what starts
the endpoint discovery process. I read somewhere that VPNv4 prefixes
related to tunnels use a SAFI of 64, so when a peer receives those prefixes
it would know that the sender wants to participate in the multipoint GRE
VPN. If that's the case, what is it about the configuration that would
cause the router to use that SAFI in the VPNv4 updates?

I see nothing in the VRF or BGP config that specifically would cause it to
behave any different. Why wouldn't it send regular VPNv4 routes? Other than
the ingress route policy that associates incoming routes to the mGRE, I
don't see what would cause the router to set the SAFI to 64 on advertised
routes.

Is that even how this works, or is there some other mechanism that triggers
tunnel endpoint discovery?

Thanks!
John