bgpd: speed up "no-hit" withdraws for routeservers
This accelerates handling of incoming Withdraw messages for routes that
don't exist in the table to begin with. Cisco IOS 12.4(24)T4 has a bug
in this regard - it sends withdraws instead of doing nothing for
prefixes that are filtered.
Pulling up the adj_in removal in Quagga should have no ill effect, but
we can avoid the costly iteration over all rsclients if there was no
adj_in entry.
Performance impact of this change on routeserver with 3 buggy peers,
startup/sync time:
before patch: 143.12 seconds (user cpu)
after patch: 7.01 seconds (user cpu)
Many thanks to Nick Hilliard & INEX for providing real-world test data!
Signed-off-by: David Lamparter <equinox@opensourcerouting.org>
Acked-by: Paul Jakma <paul@jakma.org>
diff --git a/bgpd/bgp_advertise.c b/bgpd/bgp_advertise.c
index e0fa58d..be9b480 100644
--- a/bgpd/bgp_advertise.c
+++ b/bgpd/bgp_advertise.c
@@ -361,7 +361,7 @@
XFREE (MTYPE_BGP_ADJ_IN, bai);
}
-void
+int
bgp_adj_in_unset (struct bgp_node *rn, struct peer *peer)
{
struct bgp_adj_in *adj;
@@ -371,10 +371,11 @@
break;
if (! adj)
- return;
+ return 0;
bgp_adj_in_remove (rn, adj);
bgp_unlock_node (rn);
+ return 1;
}
void