From: Claudio Jeker <cjeker@diehard.n-r-g.com>
Subject: bgpd: fix adj-rib-out cache issue with add-path
To: tech@openbsd.org
Date: Mon, 20 Jul 2026 16:31:30 +0200

The adj-rib-out remains cached when a peer goes down.
When the session is reestablished the code checks if the sync is still
valid or if a full peer_dump is needed. This depends also on the
capabilities of the session.

If a session drops from add-path send X to no add-path send then the
adj-rib-out is a bit desynced and up_generate_updates() does not all the
required cleanup.

In the EVAL_SYNC case up_generate_updates() should check if the
adjout_prefix (p) returned was for an add-path path. This is the case if
path_id_tx is not 0. In that case withdraw all paths before calling
up_process_prefix().

The add-path side already does this by default.
Most other cases where the adj-rib-out gets desynced is during config
reloads (e.g. switching export modes). There the code already does a
RIB flush and the resync.

-- 
:wq Claudio

Index: rde_update.c
===================================================================
RCS file: /cvs/src/usr.sbin/bgpd/rde_update.c,v
diff -u -p -r1.202 rde_update.c
--- rde_update.c	20 Jul 2026 13:25:49 -0000	1.202
+++ rde_update.c	20 Jul 2026 14:22:36 -0000
@@ -228,6 +228,18 @@ up_generate_updates(struct rde_peer *pee
 
 	p = adjout_prefix_first(re->prefix, peer->adjout_bid);
 
+	/*
+	 * In case the peer switched from add-path send to no add-path,
+	 * flush all the cached prefixes during initial sync.
+	 */
+	if (mode == EVAL_SYNC && p != NULL && p->path_id_tx != 0) {
+		do {
+			adjout_prefix_withdraw(peer, re->prefix, p, 1);
+			p = adjout_prefix_next(re->prefix,
+			    peer->adjout_bid, p);
+		} while(p != NULL);
+	}
+
 	new = prefix_best(re);
 	while (new != NULL) {
 		switch (up_process_prefix(peer, new, p, mode)) {