1 /* $NetBSD: ieee8023ad_lacp.c,v 1.3 2005/12/11 12:24:54 christos Exp $ */
4 * Copyright (c)2005 YAMAMOTO Takashi,
5 * Copyright (c)2008 Andrew Thompson <thompsa@FreeBSD.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
33 #include <sys/param.h>
34 #include <sys/callout.h>
35 #include <sys/eventhandler.h>
37 #include <sys/systm.h>
38 #include <sys/malloc.h>
39 #include <sys/kernel.h> /* hz */
40 #include <sys/socket.h> /* for net/if.h */
41 #include <sys/sockio.h>
42 #include <sys/sysctl.h>
43 #include <machine/stdarg.h>
45 #include <sys/rwlock.h>
46 #include <sys/taskqueue.h>
49 #include <net/if_var.h>
50 #include <net/if_dl.h>
51 #include <net/ethernet.h>
52 #include <net/if_media.h>
53 #include <net/if_types.h>
55 #include <net/if_lagg.h>
56 #include <net/ieee8023ad_lacp.h>
59 * actor system priority and port priority.
60 * XXX should be configurable.
63 #define LACP_SYSTEM_PRIO 0x8000
64 #define LACP_PORT_PRIO 0x8000
66 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
67 { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
69 static const struct tlv_template lacp_info_tlv_template[] = {
70 { LACP_TYPE_ACTORINFO,
71 sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
72 { LACP_TYPE_PARTNERINFO,
73 sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
74 { LACP_TYPE_COLLECTORINFO,
75 sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
79 static const struct tlv_template marker_info_tlv_template[] = {
81 sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
85 static const struct tlv_template marker_response_tlv_template[] = {
86 { MARKER_TYPE_RESPONSE,
87 sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
91 typedef void (*lacp_timer_func_t)(struct lacp_port *);
93 static void lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
94 static void lacp_fill_markerinfo(struct lacp_port *,
95 struct lacp_markerinfo *);
97 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
98 static void lacp_suppress_distributing(struct lacp_softc *,
99 struct lacp_aggregator *);
100 static void lacp_transit_expire(void *);
101 static void lacp_update_portmap(struct lacp_softc *);
102 static void lacp_select_active_aggregator(struct lacp_softc *);
103 static uint16_t lacp_compose_key(struct lacp_port *);
104 static int tlv_check(const void *, size_t, const struct tlvhdr *,
105 const struct tlv_template *, boolean_t);
106 static void lacp_tick(void *);
108 static void lacp_fill_aggregator_id(struct lacp_aggregator *,
109 const struct lacp_port *);
110 static void lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
111 const struct lacp_peerinfo *);
112 static int lacp_aggregator_is_compatible(const struct lacp_aggregator *,
113 const struct lacp_port *);
114 static int lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
115 const struct lacp_peerinfo *);
117 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
119 static void lacp_aggregator_addref(struct lacp_softc *,
120 struct lacp_aggregator *);
121 static void lacp_aggregator_delref(struct lacp_softc *,
122 struct lacp_aggregator *);
124 /* receive machine */
126 static int lacp_pdu_input(struct lacp_port *, struct mbuf *);
127 static int lacp_marker_input(struct lacp_port *, struct mbuf *);
128 static void lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
129 static void lacp_sm_rx_timer(struct lacp_port *);
130 static void lacp_sm_rx_set_expired(struct lacp_port *);
131 static void lacp_sm_rx_update_ntt(struct lacp_port *,
132 const struct lacpdu *);
133 static void lacp_sm_rx_record_pdu(struct lacp_port *,
134 const struct lacpdu *);
135 static void lacp_sm_rx_update_selected(struct lacp_port *,
136 const struct lacpdu *);
137 static void lacp_sm_rx_record_default(struct lacp_port *);
138 static void lacp_sm_rx_update_default_selected(struct lacp_port *);
139 static void lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
140 const struct lacp_peerinfo *);
144 static void lacp_sm_mux(struct lacp_port *);
145 static void lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
146 static void lacp_sm_mux_timer(struct lacp_port *);
148 /* periodic transmit machine */
150 static void lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
151 static void lacp_sm_ptx_tx_schedule(struct lacp_port *);
152 static void lacp_sm_ptx_timer(struct lacp_port *);
154 /* transmit machine */
156 static void lacp_sm_tx(struct lacp_port *);
157 static void lacp_sm_assert_ntt(struct lacp_port *);
159 static void lacp_run_timers(struct lacp_port *);
160 static int lacp_compare_peerinfo(const struct lacp_peerinfo *,
161 const struct lacp_peerinfo *);
162 static int lacp_compare_systemid(const struct lacp_systemid *,
163 const struct lacp_systemid *);
164 static void lacp_port_enable(struct lacp_port *);
165 static void lacp_port_disable(struct lacp_port *);
166 static void lacp_select(struct lacp_port *);
167 static void lacp_unselect(struct lacp_port *);
168 static void lacp_disable_collecting(struct lacp_port *);
169 static void lacp_enable_collecting(struct lacp_port *);
170 static void lacp_disable_distributing(struct lacp_port *);
171 static void lacp_enable_distributing(struct lacp_port *);
172 static int lacp_xmit_lacpdu(struct lacp_port *);
173 static int lacp_xmit_marker(struct lacp_port *);
177 static void lacp_dump_lacpdu(const struct lacpdu *);
178 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
180 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
181 const struct lacp_peerinfo *, char *, size_t);
182 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
184 static const char *lacp_format_state(uint8_t, char *, size_t);
185 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
186 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
188 static const char *lacp_format_portid(const struct lacp_portid *, char *,
190 static void lacp_dprintf(const struct lacp_port *, const char *, ...)
191 __attribute__((__format__(__printf__, 2, 3)));
193 static int lacp_debug = 0;
194 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
195 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_TUN,
196 &lacp_debug, 0, "Enable LACP debug logging (1=debug, 2=trace)");
197 TUNABLE_INT("net.link.lagg.lacp.debug", &lacp_debug);
199 #define LACP_DPRINTF(a) if (lacp_debug & 0x01) { lacp_dprintf a ; }
200 #define LACP_TRACE(a) if (lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
201 #define LACP_TPRINTF(a) if (lacp_debug & 0x04) { lacp_dprintf a ; }
204 * partner administration variables.
205 * XXX should be configurable.
208 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
209 .lip_systemid = { .lsi_prio = 0xffff },
210 .lip_portid = { .lpi_prio = 0xffff },
211 .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
212 LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
215 static const struct lacp_peerinfo lacp_partner_admin_strict = {
216 .lip_systemid = { .lsi_prio = 0xffff },
217 .lip_portid = { .lpi_prio = 0xffff },
221 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
222 [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
223 [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
224 [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
228 lacp_input(struct lagg_port *lgp, struct mbuf *m)
230 struct lacp_port *lp = LACP_PORT(lgp);
233 if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
238 m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
240 case SLOWPROTOCOLS_SUBTYPE_LACP:
241 lacp_pdu_input(lp, m);
244 case SLOWPROTOCOLS_SUBTYPE_MARKER:
245 lacp_marker_input(lp, m);
249 /* Not a subtype we are interested in */
254 * lacp_pdu_input: process lacpdu
257 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
259 struct lacp_softc *lsc = lp->lp_lsc;
263 if (m->m_pkthdr.len != sizeof(*du)) {
267 if ((m->m_flags & M_MCAST) == 0) {
271 if (m->m_len < sizeof(*du)) {
272 m = m_pullup(m, sizeof(*du));
278 du = mtod(m, struct lacpdu *);
280 if (memcmp(&du->ldu_eh.ether_dhost,
281 ðermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
286 * ignore the version for compatibility with
287 * the future protocol revisions.
290 if (du->ldu_sph.sph_version != 1) {
296 * ignore tlv types for compatibility with
297 * the future protocol revisions.
299 if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
300 lacp_info_tlv_template, FALSE)) {
304 if (lacp_debug > 0) {
305 lacp_dprintf(lp, "lacpdu receive\n");
306 lacp_dump_lacpdu(du);
309 if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
310 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
327 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
329 struct lagg_port *lgp = lp->lp_lagg;
330 struct lagg_softc *sc = lgp->lp_softc;
332 info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
333 memcpy(&info->lip_systemid.lsi_mac,
334 IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
335 info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
336 info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
337 info->lip_state = lp->lp_state;
341 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
343 struct ifnet *ifp = lp->lp_ifp;
345 /* Fill in the port index and system id (encoded as the MAC) */
346 info->mi_rq_port = htons(ifp->if_index);
347 memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
348 info->mi_rq_xid = htonl(0);
352 lacp_xmit_lacpdu(struct lacp_port *lp)
354 struct lagg_port *lgp = lp->lp_lagg;
359 LACP_LOCK_ASSERT(lp->lp_lsc);
361 m = m_gethdr(M_NOWAIT, MT_DATA);
365 m->m_len = m->m_pkthdr.len = sizeof(*du);
367 du = mtod(m, struct lacpdu *);
368 memset(du, 0, sizeof(*du));
370 memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
372 memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
373 du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
375 du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
376 du->ldu_sph.sph_version = 1;
378 TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
379 du->ldu_actor = lp->lp_actor;
381 TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
382 sizeof(du->ldu_partner));
383 du->ldu_partner = lp->lp_partner;
385 TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
386 sizeof(du->ldu_collector));
387 du->ldu_collector.lci_maxdelay = 0;
389 if (lacp_debug > 0) {
390 lacp_dprintf(lp, "lacpdu transmit\n");
391 lacp_dump_lacpdu(du);
394 m->m_flags |= M_MCAST;
397 * XXX should use higher priority queue.
398 * otherwise network congestion can break aggregation.
401 error = lagg_enqueue(lp->lp_ifp, m);
406 lacp_xmit_marker(struct lacp_port *lp)
408 struct lagg_port *lgp = lp->lp_lagg;
410 struct markerdu *mdu;
413 LACP_LOCK_ASSERT(lp->lp_lsc);
415 m = m_gethdr(M_NOWAIT, MT_DATA);
419 m->m_len = m->m_pkthdr.len = sizeof(*mdu);
421 mdu = mtod(m, struct markerdu *);
422 memset(mdu, 0, sizeof(*mdu));
424 memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
426 memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
427 mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
429 mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
430 mdu->mdu_sph.sph_version = 1;
432 /* Bump the transaction id and copy over the marker info */
433 lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
434 TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
435 mdu->mdu_info = lp->lp_marker;
437 LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
438 ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
439 ntohl(mdu->mdu_info.mi_rq_xid)));
441 m->m_flags |= M_MCAST;
442 error = lagg_enqueue(lp->lp_ifp, m);
447 lacp_linkstate(struct lagg_port *lgp)
449 struct lacp_port *lp = LACP_PORT(lgp);
450 struct lacp_softc *lsc = lp->lp_lsc;
451 struct ifnet *ifp = lgp->lp_ifp;
452 struct ifmediareq ifmr;
458 bzero((char *)&ifmr, sizeof(ifmr));
459 error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
464 media = ifmr.ifm_active;
465 LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
466 "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
467 (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
468 old_state = lp->lp_state;
469 old_key = lp->lp_key;
471 lp->lp_media = media;
473 * If the port is not an active full duplex Ethernet link then it can
476 if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
477 ifp->if_link_state != LINK_STATE_UP) {
478 lacp_port_disable(lp);
480 lacp_port_enable(lp);
482 lp->lp_key = lacp_compose_key(lp);
484 if (old_state != lp->lp_state || old_key != lp->lp_key) {
485 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
486 lp->lp_selected = LACP_UNSELECTED;
494 struct lacp_softc *lsc = arg;
495 struct lacp_port *lp;
497 LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
498 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
506 lacp_sm_ptx_tx_schedule(lp);
508 callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
512 lacp_port_create(struct lagg_port *lgp)
514 struct lagg_softc *sc = lgp->lp_softc;
515 struct lacp_softc *lsc = LACP_SOFTC(sc);
516 struct lacp_port *lp;
517 struct ifnet *ifp = lgp->lp_ifp;
518 struct sockaddr_dl sdl;
519 struct ifmultiaddr *rifma = NULL;
522 boolean_t active = TRUE; /* XXX should be configurable */
523 boolean_t fast = FALSE; /* XXX should be configurable */
525 link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
526 sdl.sdl_alen = ETHER_ADDR_LEN;
528 bcopy(ðermulticastaddr_slowprotocols,
529 LLADDR(&sdl), ETHER_ADDR_LEN);
530 error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
532 printf("%s: ADDMULTI failed on %s\n", __func__, lgp->lp_ifname);
536 lp = malloc(sizeof(struct lacp_port),
537 M_DEVBUF, M_NOWAIT|M_ZERO);
542 lgp->lp_psc = (caddr_t)lp;
548 LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
550 lacp_fill_actorinfo(lp, &lp->lp_actor);
551 lacp_fill_markerinfo(lp, &lp->lp_marker);
553 (active ? LACP_STATE_ACTIVITY : 0) |
554 (fast ? LACP_STATE_TIMEOUT : 0);
555 lp->lp_aggregator = NULL;
556 lacp_sm_rx_set_expired(lp);
564 lacp_port_destroy(struct lagg_port *lgp)
566 struct lacp_port *lp = LACP_PORT(lgp);
567 struct lacp_softc *lsc = lp->lp_lsc;
571 for (i = 0; i < LACP_NTIMER; i++) {
572 LACP_TIMER_DISARM(lp, i);
575 lacp_disable_collecting(lp);
576 lacp_disable_distributing(lp);
579 /* The address may have already been removed by if_purgemaddrs() */
580 if (!lgp->lp_detaching)
581 if_delmulti_ifma(lp->lp_ifma);
583 LIST_REMOVE(lp, lp_next);
589 lacp_req(struct lagg_softc *sc, caddr_t data)
591 struct lacp_opreq *req = (struct lacp_opreq *)data;
592 struct lacp_softc *lsc = LACP_SOFTC(sc);
593 struct lacp_aggregator *la = lsc->lsc_active_aggregator;
596 bzero(req, sizeof(struct lacp_opreq));
598 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
599 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
601 req->actor_key = ntohs(la->la_actor.lip_key);
602 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
603 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
604 req->actor_state = la->la_actor.lip_state;
606 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
607 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
609 req->partner_key = ntohs(la->la_partner.lip_key);
610 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
611 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
612 req->partner_state = la->la_partner.lip_state;
618 lacp_portreq(struct lagg_port *lgp, caddr_t data)
620 struct lacp_opreq *req = (struct lacp_opreq *)data;
621 struct lacp_port *lp = LACP_PORT(lgp);
622 struct lacp_softc *lsc = lp->lp_lsc;
625 req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
626 memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
628 req->actor_key = ntohs(lp->lp_actor.lip_key);
629 req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
630 req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
631 req->actor_state = lp->lp_actor.lip_state;
633 req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
634 memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
636 req->partner_key = ntohs(lp->lp_partner.lip_key);
637 req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
638 req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
639 req->partner_state = lp->lp_partner.lip_state;
644 lacp_disable_collecting(struct lacp_port *lp)
646 LACP_DPRINTF((lp, "collecting disabled\n"));
647 lp->lp_state &= ~LACP_STATE_COLLECTING;
651 lacp_enable_collecting(struct lacp_port *lp)
653 LACP_DPRINTF((lp, "collecting enabled\n"));
654 lp->lp_state |= LACP_STATE_COLLECTING;
658 lacp_disable_distributing(struct lacp_port *lp)
660 struct lacp_aggregator *la = lp->lp_aggregator;
661 struct lacp_softc *lsc = lp->lp_lsc;
662 struct lagg_softc *sc = lsc->lsc_softc;
663 char buf[LACP_LAGIDSTR_MAX+1];
665 LACP_LOCK_ASSERT(lsc);
667 if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
671 KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
672 KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
673 KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
675 LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
677 lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
678 la->la_nports, la->la_nports - 1));
680 TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
682 sc->sc_active = la->la_nports;
684 if (lsc->lsc_active_aggregator == la) {
685 lacp_suppress_distributing(lsc, la);
686 lacp_select_active_aggregator(lsc);
687 /* regenerate the port map, the active aggregator has changed */
688 lacp_update_portmap(lsc);
691 lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
695 lacp_enable_distributing(struct lacp_port *lp)
697 struct lacp_aggregator *la = lp->lp_aggregator;
698 struct lacp_softc *lsc = lp->lp_lsc;
699 struct lagg_softc *sc = lsc->lsc_softc;
700 char buf[LACP_LAGIDSTR_MAX+1];
702 LACP_LOCK_ASSERT(lsc);
704 if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
708 LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
710 lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
711 la->la_nports, la->la_nports + 1));
713 KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
714 TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
716 sc->sc_active = la->la_nports;
718 lp->lp_state |= LACP_STATE_DISTRIBUTING;
720 if (lsc->lsc_active_aggregator == la) {
721 lacp_suppress_distributing(lsc, la);
722 lacp_update_portmap(lsc);
724 /* try to become the active aggregator */
725 lacp_select_active_aggregator(lsc);
729 lacp_transit_expire(void *vp)
731 struct lacp_softc *lsc = vp;
733 LACP_LOCK_ASSERT(lsc);
737 lsc->lsc_suppress_distributing = FALSE;
741 lacp_attach_sysctl(struct lacp_softc *lsc, struct sysctl_oid *p_oid)
743 struct lagg_softc *sc = lsc->lsc_softc;
745 SYSCTL_ADD_UINT(&sc->ctx, SYSCTL_CHILDREN(p_oid), OID_AUTO,
748 &lsc->lsc_strict_mode,
749 lsc->lsc_strict_mode,
750 "Enable LACP strict mode");
754 lacp_attach_sysctl_debug(struct lacp_softc *lsc, struct sysctl_oid *p_oid)
756 struct lagg_softc *sc = lsc->lsc_softc;
757 struct sysctl_oid *oid;
759 /* Create a child of the parent lagg interface */
760 oid = SYSCTL_ADD_NODE(&sc->ctx, SYSCTL_CHILDREN(p_oid),
761 OID_AUTO, "debug", CTLFLAG_RD, NULL, "DEBUG");
763 SYSCTL_ADD_UINT(&sc->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
766 &lsc->lsc_debug.lsc_rx_test,
767 lsc->lsc_debug.lsc_rx_test,
768 "Bitmap of if_dunit entries to drop RX frames for");
769 SYSCTL_ADD_UINT(&sc->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
772 &lsc->lsc_debug.lsc_tx_test,
773 lsc->lsc_debug.lsc_tx_test,
774 "Bitmap of if_dunit entries to drop TX frames for");
778 lacp_attach(struct lagg_softc *sc)
780 struct lacp_softc *lsc;
781 struct sysctl_oid *oid;
783 lsc = malloc(sizeof(struct lacp_softc),
784 M_DEVBUF, M_NOWAIT|M_ZERO);
788 sc->sc_psc = (caddr_t)lsc;
791 lsc->lsc_hashkey = arc4random();
792 lsc->lsc_active_aggregator = NULL;
793 lsc->lsc_strict_mode = 1;
795 TAILQ_INIT(&lsc->lsc_aggregators);
796 LIST_INIT(&lsc->lsc_ports);
798 /* Create a child of the parent lagg interface */
799 oid = SYSCTL_ADD_NODE(&sc->ctx, SYSCTL_CHILDREN(sc->sc_oid),
800 OID_AUTO, "lacp", CTLFLAG_RD, NULL, "LACP");
802 /* Attach sysctl nodes */
803 lacp_attach_sysctl(lsc, oid);
804 lacp_attach_sysctl_debug(lsc, oid);
806 callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
807 callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
809 /* if the lagg is already up then do the same */
810 if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
817 lacp_detach(struct lagg_softc *sc)
819 struct lacp_softc *lsc = LACP_SOFTC(sc);
821 KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
822 ("aggregators still active"));
823 KASSERT(lsc->lsc_active_aggregator == NULL,
824 ("aggregator still attached"));
827 callout_drain(&lsc->lsc_transit_callout);
828 callout_drain(&lsc->lsc_callout);
830 LACP_LOCK_DESTROY(lsc);
836 lacp_init(struct lagg_softc *sc)
838 struct lacp_softc *lsc = LACP_SOFTC(sc);
841 callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
846 lacp_stop(struct lagg_softc *sc)
848 struct lacp_softc *lsc = LACP_SOFTC(sc);
851 callout_stop(&lsc->lsc_transit_callout);
852 callout_stop(&lsc->lsc_callout);
857 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
859 struct lacp_softc *lsc = LACP_SOFTC(sc);
860 struct lacp_portmap *pm;
861 struct lacp_port *lp;
864 if (__predict_false(lsc->lsc_suppress_distributing)) {
865 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
869 pm = &lsc->lsc_pmap[lsc->lsc_activemap];
870 if (pm->pm_count == 0) {
871 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
875 if (sc->use_flowid && (m->m_flags & M_FLOWID))
876 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
878 hash = lagg_hashmbuf(sc, m, lsc->lsc_hashkey);
879 hash %= pm->pm_count;
880 lp = pm->pm_map[hash];
882 KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
883 ("aggregated port is not distributing"));
885 return (lp->lp_lagg);
888 * lacp_suppress_distributing: drop transmit packets for a while
889 * to preserve packet ordering.
893 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
895 struct lacp_port *lp;
897 if (lsc->lsc_active_aggregator != la) {
903 lsc->lsc_suppress_distributing = TRUE;
905 /* send a marker frame down each port to verify the queues are empty */
906 LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
907 lp->lp_flags |= LACP_PORT_MARK;
908 lacp_xmit_marker(lp);
911 /* set a timeout for the marker frames */
912 callout_reset(&lsc->lsc_transit_callout,
913 LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
917 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
918 const struct lacp_peerinfo *b)
920 return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
924 lacp_compare_systemid(const struct lacp_systemid *a,
925 const struct lacp_systemid *b)
927 return (memcmp(a, b, sizeof(*a)));
932 lacp_compare_portid(const struct lacp_portid *a,
933 const struct lacp_portid *b)
935 return (memcmp(a, b, sizeof(*a)));
940 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
942 struct lacp_port *lp;
945 lp = TAILQ_FIRST(&la->la_ports);
950 speed = ifmedia_baudrate(lp->lp_media);
951 speed *= la->la_nports;
953 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
954 lp->lp_media, la->la_nports));
961 * lacp_select_active_aggregator: select an aggregator to be used to transmit
962 * packets from lagg(4) interface.
966 lacp_select_active_aggregator(struct lacp_softc *lsc)
968 struct lacp_aggregator *la;
969 struct lacp_aggregator *best_la = NULL;
970 uint64_t best_speed = 0;
971 char buf[LACP_LAGIDSTR_MAX+1];
975 TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
978 if (la->la_nports == 0) {
982 speed = lacp_aggregator_bandwidth(la);
983 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
984 lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
985 speed, la->la_nports));
987 /* This aggregator is chosen if
988 * the partner has a better system priority
989 * or, the total aggregated speed is higher
990 * or, it is already the chosen aggregator
992 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
993 LACP_SYS_PRI(best_la->la_partner)) ||
994 speed > best_speed ||
995 (speed == best_speed &&
996 la == lsc->lsc_active_aggregator)) {
1002 KASSERT(best_la == NULL || best_la->la_nports > 0,
1003 ("invalid aggregator refcnt"));
1004 KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1005 ("invalid aggregator list"));
1007 if (lsc->lsc_active_aggregator != best_la) {
1008 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1009 LACP_DPRINTF((NULL, "old %s\n",
1010 lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1011 buf, sizeof(buf))));
1013 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1015 LACP_DPRINTF((NULL, "new %s\n",
1016 lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1018 if (lsc->lsc_active_aggregator != best_la) {
1019 lsc->lsc_active_aggregator = best_la;
1020 lacp_update_portmap(lsc);
1022 lacp_suppress_distributing(lsc, best_la);
1028 * Updated the inactive portmap array with the new list of ports and
1032 lacp_update_portmap(struct lacp_softc *lsc)
1034 struct lagg_softc *sc = lsc->lsc_softc;
1035 struct lacp_aggregator *la;
1036 struct lacp_portmap *p;
1037 struct lacp_port *lp;
1042 newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1043 p = &lsc->lsc_pmap[newmap];
1044 la = lsc->lsc_active_aggregator;
1046 bzero(p, sizeof(struct lacp_portmap));
1048 if (la != NULL && la->la_nports > 0) {
1049 p->pm_count = la->la_nports;
1051 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1052 p->pm_map[i++] = lp;
1053 KASSERT(i == p->pm_count, ("Invalid port count"));
1054 speed = lacp_aggregator_bandwidth(la);
1056 sc->sc_ifp->if_baudrate = speed;
1058 /* switch the active portmap over */
1059 atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1060 LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1062 lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1066 lacp_compose_key(struct lacp_port *lp)
1068 struct lagg_port *lgp = lp->lp_lagg;
1069 struct lagg_softc *sc = lgp->lp_softc;
1070 u_int media = lp->lp_media;
1073 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1076 * non-aggregatable links should have unique keys.
1078 * XXX this isn't really unique as if_index is 16 bit.
1081 /* bit 0..14: (some bits of) if_index of this port */
1082 key = lp->lp_ifp->if_index;
1086 u_int subtype = IFM_SUBTYPE(media);
1088 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1089 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1091 /* bit 0..4: IFM_SUBTYPE modulo speed */
1116 case IFM_10G_TWINAX:
1117 case IFM_10G_TWINAX_LONG:
1130 /* bit 5..14: (some bits of) if_index of lagg device */
1131 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1134 return (htons(key));
1138 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1140 char buf[LACP_LAGIDSTR_MAX+1];
1142 LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1144 lacp_format_lagid(&la->la_actor, &la->la_partner,
1146 la->la_refcnt, la->la_refcnt + 1));
1148 KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1150 KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1154 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1156 char buf[LACP_LAGIDSTR_MAX+1];
1158 LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1160 lacp_format_lagid(&la->la_actor, &la->la_partner,
1162 la->la_refcnt, la->la_refcnt - 1));
1164 KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1166 if (la->la_refcnt > 0) {
1170 KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1171 KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1173 TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1179 * lacp_aggregator_get: allocate an aggregator.
1182 static struct lacp_aggregator *
1183 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1185 struct lacp_aggregator *la;
1187 la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1191 TAILQ_INIT(&la->la_ports);
1193 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1200 * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1204 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1206 lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1207 lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1209 la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1213 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1214 const struct lacp_peerinfo *lpi_port)
1216 memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1217 lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1218 lpi_aggr->lip_key = lpi_port->lip_key;
1222 * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1226 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1227 const struct lacp_port *lp)
1229 if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1230 !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1234 if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1238 if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1242 if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1250 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1251 const struct lacp_peerinfo *b)
1253 if (memcmp(&a->lip_systemid, &b->lip_systemid,
1254 sizeof(a->lip_systemid))) {
1258 if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1266 lacp_port_enable(struct lacp_port *lp)
1268 lp->lp_state |= LACP_STATE_AGGREGATION;
1272 lacp_port_disable(struct lacp_port *lp)
1274 lacp_set_mux(lp, LACP_MUX_DETACHED);
1276 lp->lp_state &= ~LACP_STATE_AGGREGATION;
1277 lp->lp_selected = LACP_UNSELECTED;
1278 lacp_sm_rx_record_default(lp);
1279 lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1280 lp->lp_state &= ~LACP_STATE_EXPIRED;
1284 * lacp_select: select an aggregator. create one if necessary.
1287 lacp_select(struct lacp_port *lp)
1289 struct lacp_softc *lsc = lp->lp_lsc;
1290 struct lacp_aggregator *la;
1291 char buf[LACP_LAGIDSTR_MAX+1];
1293 if (lp->lp_aggregator) {
1297 KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1298 ("timer_wait_while still active"));
1300 LACP_DPRINTF((lp, "port lagid=%s\n",
1301 lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1302 buf, sizeof(buf))));
1304 TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1305 if (lacp_aggregator_is_compatible(la, lp)) {
1311 la = lacp_aggregator_get(lsc, lp);
1313 LACP_DPRINTF((lp, "aggregator creation failed\n"));
1316 * will retry on the next tick.
1321 lacp_fill_aggregator_id(la, lp);
1322 LACP_DPRINTF((lp, "aggregator created\n"));
1324 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1325 if (la->la_refcnt == LACP_MAX_PORTS)
1327 lacp_aggregator_addref(lsc, la);
1330 LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1331 lacp_format_lagid(&la->la_actor, &la->la_partner,
1332 buf, sizeof(buf))));
1334 lp->lp_aggregator = la;
1335 lp->lp_selected = LACP_SELECTED;
1339 * lacp_unselect: finish unselect/detach process.
1343 lacp_unselect(struct lacp_port *lp)
1345 struct lacp_softc *lsc = lp->lp_lsc;
1346 struct lacp_aggregator *la = lp->lp_aggregator;
1348 KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1349 ("timer_wait_while still active"));
1355 lp->lp_aggregator = NULL;
1356 lacp_aggregator_delref(lsc, la);
1362 lacp_sm_mux(struct lacp_port *lp)
1364 struct lagg_port *lgp = lp->lp_lagg;
1365 struct lagg_softc *sc = lgp->lp_softc;
1366 enum lacp_mux_state new_state;
1368 (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1369 boolean_t p_collecting =
1370 (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1371 enum lacp_selected selected = lp->lp_selected;
1372 struct lacp_aggregator *la;
1375 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1376 "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1377 lp->lp_mux_state, selected, p_sync, p_collecting);
1380 la = lp->lp_aggregator;
1381 KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1382 ("MUX not detached"));
1383 new_state = lp->lp_mux_state;
1384 switch (lp->lp_mux_state) {
1385 case LACP_MUX_DETACHED:
1386 if (selected != LACP_UNSELECTED) {
1387 new_state = LACP_MUX_WAITING;
1390 case LACP_MUX_WAITING:
1391 KASSERT(la->la_pending > 0 ||
1392 !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1393 ("timer_wait_while still active"));
1394 if (selected == LACP_SELECTED && la->la_pending == 0) {
1395 new_state = LACP_MUX_ATTACHED;
1396 } else if (selected == LACP_UNSELECTED) {
1397 new_state = LACP_MUX_DETACHED;
1400 case LACP_MUX_ATTACHED:
1401 if (selected == LACP_SELECTED && p_sync) {
1402 new_state = LACP_MUX_COLLECTING;
1403 } else if (selected != LACP_SELECTED) {
1404 new_state = LACP_MUX_DETACHED;
1407 case LACP_MUX_COLLECTING:
1408 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1409 new_state = LACP_MUX_DISTRIBUTING;
1410 } else if (selected != LACP_SELECTED || !p_sync) {
1411 new_state = LACP_MUX_ATTACHED;
1414 case LACP_MUX_DISTRIBUTING:
1415 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1416 new_state = LACP_MUX_COLLECTING;
1417 lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1422 panic("%s: unknown state", __func__);
1425 if (lp->lp_mux_state == new_state) {
1429 lacp_set_mux(lp, new_state);
1434 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1436 struct lacp_aggregator *la = lp->lp_aggregator;
1438 if (lp->lp_mux_state == new_state) {
1442 switch (new_state) {
1443 case LACP_MUX_DETACHED:
1444 lp->lp_state &= ~LACP_STATE_SYNC;
1445 lacp_disable_distributing(lp);
1446 lacp_disable_collecting(lp);
1447 lacp_sm_assert_ntt(lp);
1449 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1450 KASSERT(la->la_pending > 0,
1451 ("timer_wait_while not active"));
1454 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1457 case LACP_MUX_WAITING:
1458 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1459 LACP_AGGREGATE_WAIT_TIME);
1462 case LACP_MUX_ATTACHED:
1463 lp->lp_state |= LACP_STATE_SYNC;
1464 lacp_disable_collecting(lp);
1465 lacp_sm_assert_ntt(lp);
1467 case LACP_MUX_COLLECTING:
1468 lacp_enable_collecting(lp);
1469 lacp_disable_distributing(lp);
1470 lacp_sm_assert_ntt(lp);
1472 case LACP_MUX_DISTRIBUTING:
1473 lacp_enable_distributing(lp);
1476 panic("%s: unknown state", __func__);
1479 LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1481 lp->lp_mux_state = new_state;
1485 lacp_sm_mux_timer(struct lacp_port *lp)
1487 struct lacp_aggregator *la = lp->lp_aggregator;
1488 char buf[LACP_LAGIDSTR_MAX+1];
1490 KASSERT(la->la_pending > 0, ("no pending event"));
1492 LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1493 lacp_format_lagid(&la->la_actor, &la->la_partner,
1495 la->la_pending, la->la_pending - 1));
1500 /* periodic transmit machine */
1503 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1505 if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1506 LACP_STATE_TIMEOUT)) {
1510 LACP_DPRINTF((lp, "partner timeout changed\n"));
1513 * FAST_PERIODIC -> SLOW_PERIODIC
1515 * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1517 * let lacp_sm_ptx_tx_schedule to update timeout.
1520 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1523 * if timeout has been shortened, assert NTT.
1526 if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1527 lacp_sm_assert_ntt(lp);
1532 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1536 if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1537 !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1543 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1547 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1551 timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1552 LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1554 LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1558 lacp_sm_ptx_timer(struct lacp_port *lp)
1560 lacp_sm_assert_ntt(lp);
1564 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1569 * check LACP_DISABLED first
1572 if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1577 * check loopback condition.
1580 if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1581 &lp->lp_actor.lip_systemid)) {
1586 * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1589 lacp_sm_rx_update_selected(lp, du);
1590 lacp_sm_rx_update_ntt(lp, du);
1591 lacp_sm_rx_record_pdu(lp, du);
1593 timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1594 LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1595 LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1597 lp->lp_state &= ~LACP_STATE_EXPIRED;
1600 * kick transmit machine without waiting the next tick.
1607 lacp_sm_rx_set_expired(struct lacp_port *lp)
1609 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1610 lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1611 LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1612 lp->lp_state |= LACP_STATE_EXPIRED;
1616 lacp_sm_rx_timer(struct lacp_port *lp)
1618 if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1619 /* CURRENT -> EXPIRED */
1620 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1621 lacp_sm_rx_set_expired(lp);
1623 /* EXPIRED -> DEFAULTED */
1624 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1625 lacp_sm_rx_update_default_selected(lp);
1626 lacp_sm_rx_record_default(lp);
1627 lp->lp_state &= ~LACP_STATE_EXPIRED;
1632 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1636 char buf[LACP_STATESTR_MAX+1];
1640 oldpstate = lp->lp_partner.lip_state;
1642 active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1643 || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1644 (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1646 lp->lp_partner = du->ldu_actor;
1648 ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1649 LACP_STATE_AGGREGATION) &&
1650 !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1651 || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1654 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1657 lp->lp_state &= ~LACP_STATE_DEFAULTED;
1659 if (oldpstate != lp->lp_partner.lip_state) {
1660 LACP_DPRINTF((lp, "old pstate %s\n",
1661 lacp_format_state(oldpstate, buf, sizeof(buf))));
1662 LACP_DPRINTF((lp, "new pstate %s\n",
1663 lacp_format_state(lp->lp_partner.lip_state, buf,
1667 /* XXX Hack, still need to implement 5.4.9 para 2,3,4 */
1668 if (lp->lp_lsc->lsc_strict_mode)
1669 lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1671 lacp_sm_ptx_update_timeout(lp, oldpstate);
1675 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1680 if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1681 !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1682 LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1683 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1684 lacp_sm_assert_ntt(lp);
1689 lacp_sm_rx_record_default(struct lacp_port *lp)
1695 oldpstate = lp->lp_partner.lip_state;
1696 if (lp->lp_lsc->lsc_strict_mode)
1697 lp->lp_partner = lacp_partner_admin_strict;
1699 lp->lp_partner = lacp_partner_admin_optimistic;;
1700 lp->lp_state |= LACP_STATE_DEFAULTED;
1701 lacp_sm_ptx_update_timeout(lp, oldpstate);
1705 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1706 const struct lacp_peerinfo *info)
1711 if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1712 !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1713 LACP_STATE_AGGREGATION)) {
1714 lp->lp_selected = LACP_UNSELECTED;
1715 /* mux machine will clean up lp->lp_aggregator */
1720 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1725 lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1729 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1734 if (lp->lp_lsc->lsc_strict_mode)
1735 lacp_sm_rx_update_selected_from_peerinfo(lp,
1736 &lacp_partner_admin_strict);
1738 lacp_sm_rx_update_selected_from_peerinfo(lp,
1739 &lacp_partner_admin_optimistic);
1742 /* transmit machine */
1745 lacp_sm_tx(struct lacp_port *lp)
1749 if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1751 || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1752 && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1755 lp->lp_flags &= ~LACP_PORT_NTT;
1758 if (!(lp->lp_flags & LACP_PORT_NTT)) {
1762 /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1763 if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1764 (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1765 LACP_DPRINTF((lp, "rate limited pdu\n"));
1769 if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1770 error = lacp_xmit_lacpdu(lp);
1772 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1776 lp->lp_flags &= ~LACP_PORT_NTT;
1778 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1784 lacp_sm_assert_ntt(struct lacp_port *lp)
1787 lp->lp_flags |= LACP_PORT_NTT;
1791 lacp_run_timers(struct lacp_port *lp)
1795 for (i = 0; i < LACP_NTIMER; i++) {
1796 KASSERT(lp->lp_timer[i] >= 0,
1797 ("invalid timer value %d", lp->lp_timer[i]));
1798 if (lp->lp_timer[i] == 0) {
1800 } else if (--lp->lp_timer[i] <= 0) {
1801 if (lacp_timer_funcs[i]) {
1802 (*lacp_timer_funcs[i])(lp);
1809 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1811 struct lacp_softc *lsc = lp->lp_lsc;
1812 struct lagg_port *lgp = lp->lp_lagg;
1813 struct lacp_port *lp2;
1814 struct markerdu *mdu;
1818 if (m->m_pkthdr.len != sizeof(*mdu)) {
1822 if ((m->m_flags & M_MCAST) == 0) {
1826 if (m->m_len < sizeof(*mdu)) {
1827 m = m_pullup(m, sizeof(*mdu));
1833 mdu = mtod(m, struct markerdu *);
1835 if (memcmp(&mdu->mdu_eh.ether_dhost,
1836 ðermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1840 if (mdu->mdu_sph.sph_version != 1) {
1844 switch (mdu->mdu_tlv.tlv_type) {
1845 case MARKER_TYPE_INFO:
1846 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1847 marker_info_tlv_template, TRUE)) {
1850 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1851 memcpy(&mdu->mdu_eh.ether_dhost,
1852 ðermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1853 memcpy(&mdu->mdu_eh.ether_shost,
1854 lgp->lp_lladdr, ETHER_ADDR_LEN);
1855 error = lagg_enqueue(lp->lp_ifp, m);
1858 case MARKER_TYPE_RESPONSE:
1859 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1860 marker_response_tlv_template, TRUE)) {
1863 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1864 ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1865 ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1867 /* Verify that it is the last marker we sent out */
1868 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1869 sizeof(struct lacp_markerinfo)))
1873 lp->lp_flags &= ~LACP_PORT_MARK;
1875 if (lsc->lsc_suppress_distributing) {
1876 /* Check if any ports are waiting for a response */
1877 LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1878 if (lp2->lp_flags & LACP_PORT_MARK) {
1885 /* All interface queues are clear */
1886 LACP_DPRINTF((NULL, "queue flush complete\n"));
1887 lsc->lsc_suppress_distributing = FALSE;
1901 LACP_DPRINTF((lp, "bad marker frame\n"));
1907 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1908 const struct tlv_template *tmpl, boolean_t check_type)
1910 while (/* CONSTCOND */ 1) {
1911 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1914 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1915 tlv->tlv_length != tmpl->tmpl_length) {
1918 if (tmpl->tmpl_type == 0) {
1921 tlv = (const struct tlvhdr *)
1922 ((const char *)tlv + tlv->tlv_length);
1931 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1933 snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1945 lacp_format_systemid(const struct lacp_systemid *sysid,
1946 char *buf, size_t buflen)
1948 char macbuf[LACP_MACSTR_MAX+1];
1950 snprintf(buf, buflen, "%04X,%s",
1951 ntohs(sysid->lsi_prio),
1952 lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1958 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
1960 snprintf(buf, buflen, "%04X,%04X",
1961 ntohs(portid->lpi_prio),
1962 ntohs(portid->lpi_portno));
1968 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
1970 char sysid[LACP_SYSTEMIDSTR_MAX+1];
1971 char portid[LACP_PORTIDSTR_MAX+1];
1973 snprintf(buf, buflen, "(%s,%04X,%s)",
1974 lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
1975 ntohs(peer->lip_key),
1976 lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
1982 lacp_format_lagid(const struct lacp_peerinfo *a,
1983 const struct lacp_peerinfo *b, char *buf, size_t buflen)
1985 char astr[LACP_PARTNERSTR_MAX+1];
1986 char bstr[LACP_PARTNERSTR_MAX+1];
1990 * there's a convention to display small numbered peer
1994 if (lacp_compare_peerinfo(a, b) > 0) {
1995 const struct lacp_peerinfo *t;
2003 snprintf(buf, buflen, "[%s,%s]",
2004 lacp_format_partner(a, astr, sizeof(astr)),
2005 lacp_format_partner(b, bstr, sizeof(bstr)));
2011 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2012 char *buf, size_t buflen)
2018 return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2022 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2024 snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2029 lacp_dump_lacpdu(const struct lacpdu *du)
2031 char buf[LACP_PARTNERSTR_MAX+1];
2032 char buf2[LACP_STATESTR_MAX+1];
2034 printf("actor=%s\n",
2035 lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2036 printf("actor.state=%s\n",
2037 lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2038 printf("partner=%s\n",
2039 lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2040 printf("partner.state=%s\n",
2041 lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2043 printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2047 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2052 printf("%s: ", lp->lp_ifp->if_xname);