From: Theo Buehler <tb@theobuehler.org>
Subject: Re: bgpd: adjust peer_dump to properly blast updates out
To: tech@openbsd.org
Date: Wed, 20 May 2026 20:06:24 +0200

On Wed, May 20, 2026 at 05:57:58PM +0200, Claudio Jeker wrote:
> I replaced the peer_blast() call in peer_dump_done() with calling
> peer_blast_done() instead. The idea was that peer_dump() already does all
> the pend_prefix_add() calls and so peer_blast() just redoes all the work.
> 
> This is true for most cases but there is the case where peer_up() calls
> peer_dump() with a pre-populated (cached) Adj-RIB-Out. In that case it is
> possible that peer_dump() via peer_generate_update and up_generate_updates
> does not really alter the Adj-RIB-Out entry and in that case
> adjout_prefix_update() will simply skip sending an update out.
> 
> I don't want to walk the RIB twice for peer_up() calls so instead add a
> force_update flag to adjout_prefix_update() and all callers upwards.
> This way peer_dump_upcall() can call peer_generate_update() with
> force_update set to true which ensures that all prefixes call
> pend_prefix_add().
> 
> I switched up_generate_default() to always send one update (since it is
> just one) so this code can now also use peer_blast_done() directly.
> 
> The good thing is the explanation for this is a lot more complicated than
> the diff itself.

ok