I noticed that the throttling for peers which was added some time ago is
incomplete. The following diff solved these issues.
In rde_update_queue_runner() only process peers which are currently not
throttled. Additionally only run the runners if not too many imsg are
pending in the RDE. Both these changes should help imporve responsiveness.
In the SE only set the throttled flag if the imsg sending was successful.
Does work fine on my systems with little or no effect on convergance time.
Please test on your systems and look for preformance changes.
--
:wq Claudio
Index: rde.c
===================================================================
RCS file: /cvs/src/usr.sbin/bgpd/rde.c,v
retrieving revision 1.435
diff -u -p -r1.435 rde.c
--- rde.c 15 Oct 2018 10:44:47 -0000 1.435
+++ rde.c 17 Oct 2018 12:18:51 -0000
@@ -334,15 +334,16 @@ rde_main(int debug, int verbose)
mctx = LIST_NEXT(mctx, entry);
}
- rde_update_queue_runner();
- for (aid = AID_INET6; aid < AID_MAX; aid++)
- rde_update6_queue_runner(aid);
+ if (ibuf_se && ibuf_se->w.queued < SESS_MSG_HIGH_MARK) {
+ rde_update_queue_runner();
+ for (aid = AID_INET6; aid < AID_MAX; aid++)
+ rde_update6_queue_runner(aid);
+ }
if (rde_dump_pending() &&
ibuf_se_ctl && ibuf_se_ctl->w.queued <= 10)
rde_dump_runner();
- if (softreconfig) {
+ if (softreconfig)
rde_reload_runner();
- }
}
/* do not clean up on shutdown on production, it takes ages. */
@@ -2665,6 +2666,8 @@ rde_update_queue_runner(void)
continue;
if (peer->state != PEER_UP)
continue;
+ if (peer->throttled)
+ continue;
/* first withdraws */
wpos = 2; /* reserve space for the length field */
r = up_dump_prefix(queue_buf + wpos, len - wpos - 2,
@@ -2732,6 +2735,8 @@ rde_update6_queue_runner(u_int8_t aid)
continue;
if (peer->state != PEER_UP)
continue;
+ if (peer->throttled)
+ continue;
len = sizeof(queue_buf) - MSGSIZE_HEADER;
b = up_dump_mp_unreach(queue_buf, &len, peer, aid);
@@ -2755,6 +2760,8 @@ rde_update6_queue_runner(u_int8_t aid)
if (peer->conf.id == 0)
continue;
if (peer->state != PEER_UP)
+ continue;
+ if (peer->throttled)
continue;
len = sizeof(queue_buf) - MSGSIZE_HEADER;
r = up_dump_mp_reach(queue_buf, &len, peer, aid);
Index: session.c
===================================================================
RCS file: /cvs/src/usr.sbin/bgpd/session.c,v
retrieving revision 1.369
diff -u -p -r1.369 session.c
--- session.c 29 Sep 2018 07:58:06 -0000 1.369
+++ session.c 17 Oct 2018 12:18:51 -0000
@@ -1382,7 +1382,8 @@ session_sendmsg(struct bgp_msg *msg, str
if (!p->throttled && p->wbuf.queued > SESS_MSG_HIGH_MARK) {
if (imsg_rde(IMSG_XOFF, p->conf.id, NULL, 0) == -1)
log_peer_warn(&p->conf, "imsg_compose XOFF");
- p->throttled = 1;
+ else
+ p->throttled = 1;
}
free(msg);
@@ -1773,7 +1774,8 @@ session_dispatch_msg(struct pollfd *pfd,
if (p->throttled && p->wbuf.queued < SESS_MSG_LOW_MARK) {
if (imsg_rde(IMSG_XON, p->conf.id, NULL, 0) == -1)
log_peer_warn(&p->conf, "imsg_compose XON");
- p->throttled = 0;
+ else
+ p->throttled = 0;
}
if (!(pfd->revents & POLLIN))
return (1);
