bgpd: [EVPN GR] fix EVPN GR for non default vrf

Consider a case where default vrf is EVPN enabled and there are some
tenant VRFs with no peers which imports some routes from default VRF.

With the current code for EVPN GR, BGP declares GR completion
after the default VRF is complete and not wait for other VRFs to complete
 its deferral path selection.

Reason for this is because of the following sequence of events,
 a) Since there are no peers in tenant VRFs(non default VRFs), we
    prematurely send UPDATE_PENDING + UPDATE_COMPLETE in
    peer_unshut_after_cfg().
 b) When bgp_do_deferred_path_selection() is invoked for default VRF, it
    completes the deferred path calculation and calls
    bgp_evpn_handle_deferred_bestpath_for_vrfs() to enqueue the deferred
    path selection for non default VRFs.
 c) It then(for default VRF) sends UPDATE_COMPLETE followed by GR done
    for all instances from BGP pov because the deferral related variables
    are not set appropriately for non default VRFs (especially
    gr_route_sync_pending)

Fixing the above by setting deferral related variables for non default
VRFs appropriately and sending UPDATE_PENDING accordingly.

Signed-off-by: Rajasekar Raja <[email protected]>
Signed-off-by: Chirag Shah <[email protected]>
Signed-off-by: Vijayalaxmi Basavaraj <[email protected]>

Author: raja-rajasekar

bgpd/bgp_route.c

@@ -4281,6 +4281,33 @@ static inline void bgp_evpn_handle_deferred_bestpath_for_vrfs(void)
 					   bgp_vrf->name_pretty,
 					   get_afi_safi_str(tmp_afi, tmp_safi, false));
 
+			/*
+			 * Below piece of code is to handle non default EVPN VRF
+			 * instances where the variables are to be set
+			 * appropriately.
+			 *
+			 * NOTE:
+			 * - Prior to this change, peer_unshut_after_cfg() sends
+			 *   UPD_PENDING + UPD_COMPLETE for non default VRFs
+			 *   prematurely.
+			 * - So, when default vrf deferral calculation is
+			 *   complete, it invokes this function to queue the
+			 *   deferral for non default vrfs.
+			 * - However it then ends up sending the UPDATE_COMPLETE
+			 *   and marks GR done for all instances since the below
+			 *   variables (especially gr_route_sync_pending) are
+			 *   not set for non default vrfs.
+			 *
+			 * Sending UPDATE_PENDING here makes sense to tell zebra
+			 * that the non default VRF is a work in progress and is
+			 * yet to go through the deferred path selection.
+			 */
+			bgp_vrf->gr_route_sync_pending = true;
+			bgp_vrf->gr_info[tmp_afi][tmp_safi].af_enabled = true;
+			bgp_zebra_update(bgp_vrf, tmp_afi, tmp_safi,
+					 ZEBRA_CLIENT_ROUTE_UPDATE_PENDING);
+			bgp_vrf->gr_info[tmp_afi][tmp_safi].select_defer_over = true;
+
 			/*
 			 * The reason why we are starting the timer and
 			 * not doing deferred BP calculation in place is