Some BGP routers do not have redundant control plane processor modules or do not support BGP HA with the same quality or coverage as 7450 ESS, 7750 SR, or 7950 XRS routes. When dealing with such routers or specific error conditions, BGP graceful restart (GR) is a good option for minimizing the network disruption caused by a control plane reset.
BGP GR assumes that the router restarting its BGP sessions has the ability and architecture to continue packet forwarding throughout the control plane reset. If this is the case, then the peers of the restarting router act as helpers and ‟hide” the control plane reset from the rest of the network so that forwarding can continue uninterrupted. Forwarding based on stale routes and hiding the ‟staleness” from other routers is considered acceptable because the duration of the control plane outage is expected to be relatively short (a few minutes). For BGP GR to be used on a session, both routers must advertise the BGP GR capability during the OPEN message exchange; see the BGP advertisement section for more details.
BGP GR is enabled on one or more BGP sessions by configuring the graceful-restart command in the global, group, or neighbor context. The command causes GR mode to be supported for the following active families:
IPv4 unicast
IPv6 unicast
VPN-IPv4
VPN-IPv6
label-IPv4
label-IPv6
L2-VPN
route-target (RTC)
flow-IPv4 (IPv4 FlowSpec)
flow-IPv6 (IPv6 FlowSpec)
Helper mode is activated when one of the following events affects an 'Established' session:
TCP socket error
new inbound TCP connection from the peer
hold timer expiry
peer unreachable
BFD down
sent NOTIFICATION message (only if enable-notification is configured under graceful-restart and the peer set the ‟N” bit in its GR capability, and the NOTIFICATION is not a 'Cease' with subcode 'Hard Reset')
received NOTIFICATION message (only if enable-notification is configured under graceful-restart and the peer set the ‟N” bit in its GR capability, and the NOTIFICATION is not a 'Cease' with subcode 'Hard Reset')
As soon as the failure is detected, the helping 7450 ESS, 7750 SR, or 7950 XRS router marks all the routes received from the peer as stale and starts a restart timer. The stale state is not factored into the BGP decision process, and it is not made visible to other routers in the network. The restart timer derives its initial value from the Restart Time carried in the last GR capability of the peer. The default advertised Restart Time is 300 seconds, but it can be changed using the restart-time command.
When the restart timer expires, helping stops if the session is not yet re-established. If the session is re-established before the restart timer expires and the new GR capability from the restarting router indicates that the forwarding state has been preserved, then helping continues and the peers exchange routes per the normal procedure.
When each router has advertised all its routes for a specific address family, it sends an End-of-RIB marker (EOR) for the address family. The EOR is a minimal UPDATE message with no reachable or unreachable NLRI for the AFI or SAFI. When the helping router receives an EOR, it deletes all remaining stale routes of the AFI or SAFI that were not refreshed in the most recent set of UPDATE messages. The maximum amount of time that routes can remain stale (before being deleted if they are not refreshed) is configurable using the stale-routes-time.