We're experimenting with BGP-LU and frr in our lab. We have two host machines running frr. Each is connected to two TOR switches through their eth0 and eth1 links (four TORs total). Those are connected through another pair of switches. We have configured BGP-LU to distribute MPLS labels over ebpg throughout. Connectivity between loopback addresses on the two hosts works well at this point.

Then, I created some namespaces in the hosts and gave them addresses. We use BGP to announce those addresses as /32 routes with the loopback address as the next hop. The route between namespaces on the two machines is recursively resolved so that they push the same MPLS label as the path to the other loopback. This all works well to start off with and we did some iperf runs that showed pretty good results. ECMP was working because the bandwidth was higher than any single link.

After some link state changes, we seemed to lose the connection. However, pings between the loopback addresses still worked. After some time, we noticed that the MPLS labels in the routes to the namespace addresses (/32s) were different than the label in the route to the loopback. Since the former routes are resolved recursively using the latter, the labels should always be the same. Could this be a bug in FRR? Shouldn't the routes to the namespaces be invalidated or updated as soon as the route it was based on changed? The traffic between namespaces is getting dropped because the switch doesn't know about the label being pushed by the host.

Any insight would be very helpful.

Thanks!

Carl Baldwin

Here are /32 routes received. 10.112.128.1 is the loopback on the other host. The four routes are to four namespaces on the other host.

Below is the routing table as it looked when we lost connectivity. Notice that the mpls label for the loopback route is 306592 via eth1 but the label for the four namespace addresses is 306576.