On Sun, Dec 28, 2025 at 05:34:23PM +0100, Claudio Jeker wrote:
> All done.
>  
> New diff below.

Thanks, I like this better. I think there's one logic error introduced
with the conversion, other than that this is ok.

> @@ -290,8 +302,66 @@ rde_generate_updates(struct rib_entry *r
>  {
>  	struct rde_peer	*peer;
>  
> -	RB_FOREACH(peer, peer_tree, &peertable)
> -		peer_generate_update(peer, re, newpath, old_pathid_tx, mode);
> +	switch (mode) {
> +	case EVAL_RECONF:
> +		/* skip peers which don't need to reconfigure */
> +		RB_FOREACH(peer, peer_tree, &peertable) {
> +			if (peer->reconf_out == 0)
> +				continue;
> +			peer_generate_update(peer, re, NULL, 0, EVAL_RECONF);
> +		}
> +		return;
> +	case EVAL_DEFAULT:
> +		break;
> +	case EVAL_ALL:
> +		/*
> +		 * EVAL_DEFAULT is triggered when a new best path is selected.
> +		 * EVAL_ALL is sent for any other update (needed for peers with
> +		 * addpath or evaluate all set).
> +		 * There can be only one EVAL_DEFAULT queued, it replaces the
> +		 * previous one. A flag is enough.
> +		 * A path can only exist once in the queue (old or new).
> +		 */
> +		if (re->pq_mode == EVAL_DEFAULT)
> +			/* already a best path update pending, nothing to do */
> +			return;
> +
> +		break;
> +	case EVAL_NONE:
> +		fatalx("bad eval mode in %s", __func__);
> +	}
> +
> +	if (re->pq_mode == EVAL_NONE) {

Should this not be != EVAL_NONE? This would match the condition you had
before. My understanding is that entries with EVAL_NONE should never be
present in the queue.

> +		peer = peer_get(re->pq_peer_id);
> +		TAILQ_REMOVE(&peer->rib_pq_head, re, rib_queue);
> +	}
> +	if (newpath != NULL)
> +		peer = prefix_peer(newpath);
> +	else
> +		peer = peerself;
> +	re->pq_mode = mode;
> +	re->pq_peer_id = peer->conf.id;
> +	TAILQ_INSERT_TAIL(&peer->rib_pq_head, re, rib_queue);
> +}