1 /* $NetBSD: ieee8023ad_lacp.c,v 1.3 2005/12/11 12:24:54 christos Exp $ */
4 * SPDX-License-Identifier: BSD-2-Clause-NetBSD
6 * Copyright (c)2005 YAMAMOTO Takashi,
7 * Copyright (c)2008 Andrew Thompson <thompsa@FreeBSD.org>
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
35 #include "opt_kern_tls.h"
36 #include "opt_ratelimit.h"
38 #include <sys/param.h>
39 #include <sys/callout.h>
40 #include <sys/eventhandler.h>
42 #include <sys/systm.h>
43 #include <sys/malloc.h>
44 #include <sys/kernel.h> /* hz */
45 #include <sys/socket.h> /* for net/if.h */
46 #include <sys/sockio.h>
47 #include <sys/sysctl.h>
48 #include <machine/stdarg.h>
50 #include <sys/rwlock.h>
51 #include <sys/taskqueue.h>
54 #include <net/if_var.h>
55 #include <net/if_dl.h>
56 #include <net/ethernet.h>
57 #include <net/if_media.h>
58 #include <net/if_types.h>
60 #include <net/if_lagg.h>
61 #include <net/ieee8023ad_lacp.h>
64 * actor system priority and port priority.
65 * XXX should be configurable.
68 #define LACP_SYSTEM_PRIO 0x8000
69 #define LACP_PORT_PRIO 0x8000
71 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
72 { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
74 static const struct tlv_template lacp_info_tlv_template[] = {
75 { LACP_TYPE_ACTORINFO,
76 sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
77 { LACP_TYPE_PARTNERINFO,
78 sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
79 { LACP_TYPE_COLLECTORINFO,
80 sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
84 static const struct tlv_template marker_info_tlv_template[] = {
86 sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
90 static const struct tlv_template marker_response_tlv_template[] = {
91 { MARKER_TYPE_RESPONSE,
92 sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
96 typedef void (*lacp_timer_func_t)(struct lacp_port *);
98 static void lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
99 static void lacp_fill_markerinfo(struct lacp_port *,
100 struct lacp_markerinfo *);
102 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
103 static void lacp_suppress_distributing(struct lacp_softc *,
104 struct lacp_aggregator *);
105 static void lacp_transit_expire(void *);
106 static void lacp_update_portmap(struct lacp_softc *);
107 static void lacp_select_active_aggregator(struct lacp_softc *);
108 static uint16_t lacp_compose_key(struct lacp_port *);
109 static int tlv_check(const void *, size_t, const struct tlvhdr *,
110 const struct tlv_template *, boolean_t);
111 static void lacp_tick(void *);
113 static void lacp_fill_aggregator_id(struct lacp_aggregator *,
114 const struct lacp_port *);
115 static void lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
116 const struct lacp_peerinfo *);
117 static int lacp_aggregator_is_compatible(const struct lacp_aggregator *,
118 const struct lacp_port *);
119 static int lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
120 const struct lacp_peerinfo *);
122 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
124 static void lacp_aggregator_addref(struct lacp_softc *,
125 struct lacp_aggregator *);
126 static void lacp_aggregator_delref(struct lacp_softc *,
127 struct lacp_aggregator *);
129 /* receive machine */
131 static int lacp_pdu_input(struct lacp_port *, struct mbuf *);
132 static int lacp_marker_input(struct lacp_port *, struct mbuf *);
133 static void lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
134 static void lacp_sm_rx_timer(struct lacp_port *);
135 static void lacp_sm_rx_set_expired(struct lacp_port *);
136 static void lacp_sm_rx_update_ntt(struct lacp_port *,
137 const struct lacpdu *);
138 static void lacp_sm_rx_record_pdu(struct lacp_port *,
139 const struct lacpdu *);
140 static void lacp_sm_rx_update_selected(struct lacp_port *,
141 const struct lacpdu *);
142 static void lacp_sm_rx_record_default(struct lacp_port *);
143 static void lacp_sm_rx_update_default_selected(struct lacp_port *);
144 static void lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
145 const struct lacp_peerinfo *);
149 static void lacp_sm_mux(struct lacp_port *);
150 static void lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
151 static void lacp_sm_mux_timer(struct lacp_port *);
153 /* periodic transmit machine */
155 static void lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
156 static void lacp_sm_ptx_tx_schedule(struct lacp_port *);
157 static void lacp_sm_ptx_timer(struct lacp_port *);
159 /* transmit machine */
161 static void lacp_sm_tx(struct lacp_port *);
162 static void lacp_sm_assert_ntt(struct lacp_port *);
164 static void lacp_run_timers(struct lacp_port *);
165 static int lacp_compare_peerinfo(const struct lacp_peerinfo *,
166 const struct lacp_peerinfo *);
167 static int lacp_compare_systemid(const struct lacp_systemid *,
168 const struct lacp_systemid *);
169 static void lacp_port_enable(struct lacp_port *);
170 static void lacp_port_disable(struct lacp_port *);
171 static void lacp_select(struct lacp_port *);
172 static void lacp_unselect(struct lacp_port *);
173 static void lacp_disable_collecting(struct lacp_port *);
174 static void lacp_enable_collecting(struct lacp_port *);
175 static void lacp_disable_distributing(struct lacp_port *);
176 static void lacp_enable_distributing(struct lacp_port *);
177 static int lacp_xmit_lacpdu(struct lacp_port *);
178 static int lacp_xmit_marker(struct lacp_port *);
182 static void lacp_dump_lacpdu(const struct lacpdu *);
183 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
185 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
186 const struct lacp_peerinfo *, char *, size_t);
187 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
189 static const char *lacp_format_state(uint8_t, char *, size_t);
190 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
191 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
193 static const char *lacp_format_portid(const struct lacp_portid *, char *,
195 static void lacp_dprintf(const struct lacp_port *, const char *, ...)
196 __attribute__((__format__(__printf__, 2, 3)));
198 VNET_DEFINE_STATIC(int, lacp_debug);
199 #define V_lacp_debug VNET(lacp_debug)
200 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
201 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
202 &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
204 VNET_DEFINE_STATIC(int, lacp_default_strict_mode) = 1;
205 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode,
206 CTLFLAG_RWTUN | CTLFLAG_VNET, &VNET_NAME(lacp_default_strict_mode), 0,
207 "LACP strict protocol compliance default");
209 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
210 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
211 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
214 * partner administration variables.
215 * XXX should be configurable.
218 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
219 .lip_systemid = { .lsi_prio = 0xffff },
220 .lip_portid = { .lpi_prio = 0xffff },
221 .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
222 LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
225 static const struct lacp_peerinfo lacp_partner_admin_strict = {
226 .lip_systemid = { .lsi_prio = 0xffff },
227 .lip_portid = { .lpi_prio = 0xffff },
231 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
232 [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
233 [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
234 [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
238 lacp_input(struct lagg_port *lgp, struct mbuf *m)
240 struct lacp_port *lp = LACP_PORT(lgp);
243 if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
248 m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
250 case SLOWPROTOCOLS_SUBTYPE_LACP:
251 lacp_pdu_input(lp, m);
254 case SLOWPROTOCOLS_SUBTYPE_MARKER:
255 lacp_marker_input(lp, m);
259 /* Not a subtype we are interested in */
264 * lacp_pdu_input: process lacpdu
267 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
269 struct lacp_softc *lsc = lp->lp_lsc;
273 if (m->m_pkthdr.len != sizeof(*du)) {
277 if ((m->m_flags & M_MCAST) == 0) {
281 if (m->m_len < sizeof(*du)) {
282 m = m_pullup(m, sizeof(*du));
288 du = mtod(m, struct lacpdu *);
290 if (memcmp(&du->ldu_eh.ether_dhost,
291 ðermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
296 * ignore the version for compatibility with
297 * the future protocol revisions.
300 if (du->ldu_sph.sph_version != 1) {
306 * ignore tlv types for compatibility with
307 * the future protocol revisions.
309 if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
310 lacp_info_tlv_template, FALSE)) {
314 if (V_lacp_debug > 0) {
315 lacp_dprintf(lp, "lacpdu receive\n");
316 lacp_dump_lacpdu(du);
319 if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
320 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
337 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
339 struct lagg_port *lgp = lp->lp_lagg;
340 struct lagg_softc *sc = lgp->lp_softc;
342 info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
343 memcpy(&info->lip_systemid.lsi_mac,
344 IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
345 info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
346 info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
347 info->lip_state = lp->lp_state;
351 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
353 struct ifnet *ifp = lp->lp_ifp;
355 /* Fill in the port index and system id (encoded as the MAC) */
356 info->mi_rq_port = htons(ifp->if_index);
357 memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
358 info->mi_rq_xid = htonl(0);
362 lacp_xmit_lacpdu(struct lacp_port *lp)
364 struct lagg_port *lgp = lp->lp_lagg;
369 LACP_LOCK_ASSERT(lp->lp_lsc);
371 m = m_gethdr(M_NOWAIT, MT_DATA);
375 m->m_len = m->m_pkthdr.len = sizeof(*du);
377 du = mtod(m, struct lacpdu *);
378 memset(du, 0, sizeof(*du));
380 memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
382 memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
383 du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
385 du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
386 du->ldu_sph.sph_version = 1;
388 TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
389 du->ldu_actor = lp->lp_actor;
391 TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
392 sizeof(du->ldu_partner));
393 du->ldu_partner = lp->lp_partner;
395 TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
396 sizeof(du->ldu_collector));
397 du->ldu_collector.lci_maxdelay = 0;
399 if (V_lacp_debug > 0) {
400 lacp_dprintf(lp, "lacpdu transmit\n");
401 lacp_dump_lacpdu(du);
404 m->m_flags |= M_MCAST;
407 * XXX should use higher priority queue.
408 * otherwise network congestion can break aggregation.
411 error = lagg_enqueue(lp->lp_ifp, m);
416 lacp_xmit_marker(struct lacp_port *lp)
418 struct lagg_port *lgp = lp->lp_lagg;
420 struct markerdu *mdu;
423 LACP_LOCK_ASSERT(lp->lp_lsc);
425 m = m_gethdr(M_NOWAIT, MT_DATA);
429 m->m_len = m->m_pkthdr.len = sizeof(*mdu);
431 mdu = mtod(m, struct markerdu *);
432 memset(mdu, 0, sizeof(*mdu));
434 memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
436 memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
437 mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
439 mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
440 mdu->mdu_sph.sph_version = 1;
442 /* Bump the transaction id and copy over the marker info */
443 lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
444 TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
445 mdu->mdu_info = lp->lp_marker;
447 LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
448 ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
449 ntohl(mdu->mdu_info.mi_rq_xid)));
451 m->m_flags |= M_MCAST;
452 error = lagg_enqueue(lp->lp_ifp, m);
457 lacp_linkstate(struct lagg_port *lgp)
459 struct lacp_port *lp = LACP_PORT(lgp);
460 struct lacp_softc *lsc = lp->lp_lsc;
461 struct ifnet *ifp = lgp->lp_ifp;
462 struct ifmediareq ifmr;
468 bzero((char *)&ifmr, sizeof(ifmr));
469 error = (*ifp->if_ioctl)(ifp, SIOCGIFXMEDIA, (caddr_t)&ifmr);
471 bzero((char *)&ifmr, sizeof(ifmr));
472 error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
478 media = ifmr.ifm_active;
479 LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
480 "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
481 (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
482 old_state = lp->lp_state;
483 old_key = lp->lp_key;
485 lp->lp_media = media;
487 * If the port is not an active full duplex Ethernet link then it can
490 if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
491 ifp->if_link_state != LINK_STATE_UP) {
492 lacp_port_disable(lp);
494 lacp_port_enable(lp);
496 lp->lp_key = lacp_compose_key(lp);
498 if (old_state != lp->lp_state || old_key != lp->lp_key) {
499 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
500 lp->lp_selected = LACP_UNSELECTED;
508 struct lacp_softc *lsc = arg;
509 struct lacp_port *lp;
511 LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
512 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
515 CURVNET_SET(lp->lp_ifp->if_vnet);
521 lacp_sm_ptx_tx_schedule(lp);
524 callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
528 lacp_port_create(struct lagg_port *lgp)
530 struct lagg_softc *sc = lgp->lp_softc;
531 struct lacp_softc *lsc = LACP_SOFTC(sc);
532 struct lacp_port *lp;
533 struct ifnet *ifp = lgp->lp_ifp;
534 struct sockaddr_dl sdl;
535 struct ifmultiaddr *rifma = NULL;
538 link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
539 sdl.sdl_alen = ETHER_ADDR_LEN;
541 bcopy(ðermulticastaddr_slowprotocols,
542 LLADDR(&sdl), ETHER_ADDR_LEN);
543 error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
545 printf("%s: ADDMULTI failed on %s\n", __func__,
546 lgp->lp_ifp->if_xname);
550 lp = malloc(sizeof(struct lacp_port),
551 M_DEVBUF, M_NOWAIT|M_ZERO);
562 LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
564 lacp_fill_actorinfo(lp, &lp->lp_actor);
565 lacp_fill_markerinfo(lp, &lp->lp_marker);
566 lp->lp_state = LACP_STATE_ACTIVITY;
567 lp->lp_aggregator = NULL;
568 lacp_sm_rx_set_expired(lp);
576 lacp_port_destroy(struct lagg_port *lgp)
578 struct lacp_port *lp = LACP_PORT(lgp);
579 struct lacp_softc *lsc = lp->lp_lsc;
583 for (i = 0; i < LACP_NTIMER; i++) {
584 LACP_TIMER_DISARM(lp, i);
587 lacp_disable_collecting(lp);
588 lacp_disable_distributing(lp);
591 LIST_REMOVE(lp, lp_next);
594 /* The address may have already been removed by if_purgemaddrs() */
595 if (!lgp->lp_detaching)
596 if_delmulti_ifma(lp->lp_ifma);
602 lacp_req(struct lagg_softc *sc, void *data)
604 struct lacp_opreq *req = (struct lacp_opreq *)data;
605 struct lacp_softc *lsc = LACP_SOFTC(sc);
606 struct lacp_aggregator *la;
608 bzero(req, sizeof(struct lacp_opreq));
611 * If the LACP softc is NULL, return with the opreq structure full of
612 * zeros. It is normal for the softc to be NULL while the lagg is
618 la = lsc->lsc_active_aggregator;
621 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
622 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
624 req->actor_key = ntohs(la->la_actor.lip_key);
625 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
626 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
627 req->actor_state = la->la_actor.lip_state;
629 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
630 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
632 req->partner_key = ntohs(la->la_partner.lip_key);
633 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
634 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
635 req->partner_state = la->la_partner.lip_state;
641 lacp_portreq(struct lagg_port *lgp, void *data)
643 struct lacp_opreq *req = (struct lacp_opreq *)data;
644 struct lacp_port *lp = LACP_PORT(lgp);
645 struct lacp_softc *lsc = lp->lp_lsc;
648 req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
649 memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
651 req->actor_key = ntohs(lp->lp_actor.lip_key);
652 req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
653 req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
654 req->actor_state = lp->lp_actor.lip_state;
656 req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
657 memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
659 req->partner_key = ntohs(lp->lp_partner.lip_key);
660 req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
661 req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
662 req->partner_state = lp->lp_partner.lip_state;
667 lacp_disable_collecting(struct lacp_port *lp)
669 LACP_DPRINTF((lp, "collecting disabled\n"));
670 lp->lp_state &= ~LACP_STATE_COLLECTING;
674 lacp_enable_collecting(struct lacp_port *lp)
676 LACP_DPRINTF((lp, "collecting enabled\n"));
677 lp->lp_state |= LACP_STATE_COLLECTING;
681 lacp_disable_distributing(struct lacp_port *lp)
683 struct lacp_aggregator *la = lp->lp_aggregator;
684 struct lacp_softc *lsc = lp->lp_lsc;
685 struct lagg_softc *sc = lsc->lsc_softc;
686 char buf[LACP_LAGIDSTR_MAX+1];
688 LACP_LOCK_ASSERT(lsc);
690 if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
694 KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
695 KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
696 KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
698 LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
700 lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
701 la->la_nports, la->la_nports - 1));
703 TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
705 sc->sc_active = la->la_nports;
707 if (lsc->lsc_active_aggregator == la) {
708 lacp_suppress_distributing(lsc, la);
709 lacp_select_active_aggregator(lsc);
710 /* regenerate the port map, the active aggregator has changed */
711 lacp_update_portmap(lsc);
714 lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
715 if_link_state_change(sc->sc_ifp,
716 sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
720 lacp_enable_distributing(struct lacp_port *lp)
722 struct lacp_aggregator *la = lp->lp_aggregator;
723 struct lacp_softc *lsc = lp->lp_lsc;
724 struct lagg_softc *sc = lsc->lsc_softc;
725 char buf[LACP_LAGIDSTR_MAX+1];
727 LACP_LOCK_ASSERT(lsc);
729 if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
733 LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
735 lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
736 la->la_nports, la->la_nports + 1));
738 KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
739 TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
741 sc->sc_active = la->la_nports;
743 lp->lp_state |= LACP_STATE_DISTRIBUTING;
745 if (lsc->lsc_active_aggregator == la) {
746 lacp_suppress_distributing(lsc, la);
747 lacp_update_portmap(lsc);
749 /* try to become the active aggregator */
750 lacp_select_active_aggregator(lsc);
752 if_link_state_change(sc->sc_ifp,
753 sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
757 lacp_transit_expire(void *vp)
759 struct lacp_softc *lsc = vp;
761 LACP_LOCK_ASSERT(lsc);
763 CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
767 lsc->lsc_suppress_distributing = FALSE;
771 lacp_attach(struct lagg_softc *sc)
773 struct lacp_softc *lsc;
775 lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
780 lsc->lsc_hashkey = m_ether_tcpip_hash_init();
781 lsc->lsc_active_aggregator = NULL;
782 lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
784 TAILQ_INIT(&lsc->lsc_aggregators);
785 LIST_INIT(&lsc->lsc_ports);
787 callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
788 callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
790 /* if the lagg is already up then do the same */
791 if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
796 lacp_detach(void *psc)
798 struct lacp_softc *lsc = (struct lacp_softc *)psc;
800 KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
801 ("aggregators still active"));
802 KASSERT(lsc->lsc_active_aggregator == NULL,
803 ("aggregator still attached"));
805 callout_drain(&lsc->lsc_transit_callout);
806 callout_drain(&lsc->lsc_callout);
808 LACP_LOCK_DESTROY(lsc);
813 lacp_init(struct lagg_softc *sc)
815 struct lacp_softc *lsc = LACP_SOFTC(sc);
818 callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
823 lacp_stop(struct lagg_softc *sc)
825 struct lacp_softc *lsc = LACP_SOFTC(sc);
828 callout_stop(&lsc->lsc_transit_callout);
829 callout_stop(&lsc->lsc_callout);
834 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
836 struct lacp_softc *lsc = LACP_SOFTC(sc);
837 struct lacp_portmap *pm;
838 struct lacp_port *lp;
839 struct lacp_port **map;
843 if (__predict_false(lsc->lsc_suppress_distributing)) {
844 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
848 pm = &lsc->lsc_pmap[lsc->lsc_activemap];
849 if (pm->pm_count == 0) {
850 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
855 if ((sc->sc_opts & LAGG_OPT_USE_NUMA) &&
856 pm->pm_num_dom > 1 && m->m_pkthdr.numa_domain < MAXMEMDOM) {
857 count = pm->pm_numa[m->m_pkthdr.numa_domain].count;
859 map = pm->pm_numa[m->m_pkthdr.numa_domain].map;
861 /* No ports on this domain; use global hash. */
863 count = pm->pm_count;
869 count = pm->pm_count;
871 if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
872 M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
873 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
875 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
880 KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
881 ("aggregated port is not distributing"));
883 return (lp->lp_lagg);
886 #if defined(RATELIMIT) || defined(KERN_TLS)
888 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t flowid)
890 struct lacp_softc *lsc = LACP_SOFTC(sc);
891 struct lacp_portmap *pm;
892 struct lacp_port *lp;
895 if (__predict_false(lsc->lsc_suppress_distributing)) {
896 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
900 pm = &lsc->lsc_pmap[lsc->lsc_activemap];
901 if (pm->pm_count == 0) {
902 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
906 hash = flowid >> sc->flowid_shift;
907 hash %= pm->pm_count;
908 lp = pm->pm_map[hash];
910 return (lp->lp_lagg);
915 * lacp_suppress_distributing: drop transmit packets for a while
916 * to preserve packet ordering.
920 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
922 struct lacp_port *lp;
924 if (lsc->lsc_active_aggregator != la) {
930 lsc->lsc_suppress_distributing = TRUE;
932 /* send a marker frame down each port to verify the queues are empty */
933 LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
934 lp->lp_flags |= LACP_PORT_MARK;
935 lacp_xmit_marker(lp);
938 /* set a timeout for the marker frames */
939 callout_reset(&lsc->lsc_transit_callout,
940 LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
944 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
945 const struct lacp_peerinfo *b)
947 return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
951 lacp_compare_systemid(const struct lacp_systemid *a,
952 const struct lacp_systemid *b)
954 return (memcmp(a, b, sizeof(*a)));
959 lacp_compare_portid(const struct lacp_portid *a,
960 const struct lacp_portid *b)
962 return (memcmp(a, b, sizeof(*a)));
967 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
969 struct lacp_port *lp;
972 lp = TAILQ_FIRST(&la->la_ports);
977 speed = ifmedia_baudrate(lp->lp_media);
978 speed *= la->la_nports;
980 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
981 lp->lp_media, la->la_nports));
988 * lacp_select_active_aggregator: select an aggregator to be used to transmit
989 * packets from lagg(4) interface.
993 lacp_select_active_aggregator(struct lacp_softc *lsc)
995 struct lacp_aggregator *la;
996 struct lacp_aggregator *best_la = NULL;
997 uint64_t best_speed = 0;
998 char buf[LACP_LAGIDSTR_MAX+1];
1002 TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1005 if (la->la_nports == 0) {
1009 speed = lacp_aggregator_bandwidth(la);
1010 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
1011 lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
1012 speed, la->la_nports));
1015 * This aggregator is chosen if the partner has a better
1016 * system priority or, the total aggregated speed is higher
1017 * or, it is already the chosen aggregator
1019 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
1020 LACP_SYS_PRI(best_la->la_partner)) ||
1021 speed > best_speed ||
1022 (speed == best_speed &&
1023 la == lsc->lsc_active_aggregator)) {
1029 KASSERT(best_la == NULL || best_la->la_nports > 0,
1030 ("invalid aggregator refcnt"));
1031 KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1032 ("invalid aggregator list"));
1034 if (lsc->lsc_active_aggregator != best_la) {
1035 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1036 LACP_DPRINTF((NULL, "old %s\n",
1037 lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1038 buf, sizeof(buf))));
1040 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1042 LACP_DPRINTF((NULL, "new %s\n",
1043 lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1045 if (lsc->lsc_active_aggregator != best_la) {
1046 lsc->lsc_active_aggregator = best_la;
1047 lacp_update_portmap(lsc);
1049 lacp_suppress_distributing(lsc, best_la);
1055 * Updated the inactive portmap array with the new list of ports and
1059 lacp_update_portmap(struct lacp_softc *lsc)
1061 struct lagg_softc *sc = lsc->lsc_softc;
1062 struct lacp_aggregator *la;
1063 struct lacp_portmap *p;
1064 struct lacp_port *lp;
1073 newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1074 p = &lsc->lsc_pmap[newmap];
1075 la = lsc->lsc_active_aggregator;
1077 bzero(p, sizeof(struct lacp_portmap));
1079 if (la != NULL && la->la_nports > 0) {
1080 p->pm_count = la->la_nports;
1082 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q) {
1083 p->pm_map[i++] = lp;
1085 domain = lp->lp_ifp->if_numa_domain;
1086 if (domain >= MAXMEMDOM)
1088 count = p->pm_numa[domain].count;
1089 p->pm_numa[domain].map[count] = lp;
1090 p->pm_numa[domain].count++;
1093 KASSERT(i == p->pm_count, ("Invalid port count"));
1096 for (i = 0; i < MAXMEMDOM; i++) {
1097 if (p->pm_numa[i].count != 0)
1101 speed = lacp_aggregator_bandwidth(la);
1103 sc->sc_ifp->if_baudrate = speed;
1105 /* switch the active portmap over */
1106 atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1107 LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1109 lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1113 lacp_compose_key(struct lacp_port *lp)
1115 struct lagg_port *lgp = lp->lp_lagg;
1116 struct lagg_softc *sc = lgp->lp_softc;
1117 u_int media = lp->lp_media;
1120 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1123 * non-aggregatable links should have unique keys.
1125 * XXX this isn't really unique as if_index is 16 bit.
1128 /* bit 0..14: (some bits of) if_index of this port */
1129 key = lp->lp_ifp->if_index;
1133 u_int subtype = IFM_SUBTYPE(media);
1135 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1136 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1138 /* bit 0..4: IFM_SUBTYPE modulo speed */
1161 case IFM_1000_SGMII:
1162 case IFM_1000_CX_SGMII:
1168 case IFM_10G_TWINAX:
1169 case IFM_10G_TWINAX_LONG:
1199 case IFM_50G_LAUI2_AC:
1201 case IFM_50G_AUI2_AC:
1207 case IFM_50G_KR_PAM4:
1208 case IFM_50G_AUI1_AC:
1236 case IFM_40G_XLAUI_AC:
1244 case IFM_100G_CAUI4_AC:
1245 case IFM_100G_CAUI4:
1246 case IFM_100G_AUI4_AC:
1248 case IFM_100G_CR_PAM4:
1249 case IFM_100G_KR_PAM4:
1253 case IFM_100G_KR2_PAM4:
1254 case IFM_100G_CAUI2_AC:
1255 case IFM_100G_CAUI2:
1256 case IFM_100G_AUI2_AC:
1260 case IFM_200G_CR4_PAM4:
1265 case IFM_200G_KR4_PAM4:
1266 case IFM_200G_AUI4_AC:
1268 case IFM_200G_AUI8_AC:
1270 key = IFM_200G_CR4_PAM4;
1275 case IFM_400G_AUI8_AC:
1283 /* bit 5..14: (some bits of) if_index of lagg device */
1284 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1287 return (htons(key));
1291 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1293 char buf[LACP_LAGIDSTR_MAX+1];
1295 LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1297 lacp_format_lagid(&la->la_actor, &la->la_partner,
1299 la->la_refcnt, la->la_refcnt + 1));
1301 KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1303 KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1307 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1309 char buf[LACP_LAGIDSTR_MAX+1];
1311 LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1313 lacp_format_lagid(&la->la_actor, &la->la_partner,
1315 la->la_refcnt, la->la_refcnt - 1));
1317 KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1319 if (la->la_refcnt > 0) {
1323 KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1324 KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1326 TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1332 * lacp_aggregator_get: allocate an aggregator.
1335 static struct lacp_aggregator *
1336 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1338 struct lacp_aggregator *la;
1340 la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1344 TAILQ_INIT(&la->la_ports);
1346 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1353 * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1357 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1359 lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1360 lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1362 la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1366 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1367 const struct lacp_peerinfo *lpi_port)
1369 memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1370 lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1371 lpi_aggr->lip_key = lpi_port->lip_key;
1375 * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1379 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1380 const struct lacp_port *lp)
1382 if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1383 !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1387 if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1391 if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1395 if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1403 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1404 const struct lacp_peerinfo *b)
1406 if (memcmp(&a->lip_systemid, &b->lip_systemid,
1407 sizeof(a->lip_systemid))) {
1411 if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1419 lacp_port_enable(struct lacp_port *lp)
1421 lp->lp_state |= LACP_STATE_AGGREGATION;
1425 lacp_port_disable(struct lacp_port *lp)
1427 lacp_set_mux(lp, LACP_MUX_DETACHED);
1429 lp->lp_state &= ~LACP_STATE_AGGREGATION;
1430 lp->lp_selected = LACP_UNSELECTED;
1431 lacp_sm_rx_record_default(lp);
1432 lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1433 lp->lp_state &= ~LACP_STATE_EXPIRED;
1437 * lacp_select: select an aggregator. create one if necessary.
1440 lacp_select(struct lacp_port *lp)
1442 struct lacp_softc *lsc = lp->lp_lsc;
1443 struct lacp_aggregator *la;
1444 char buf[LACP_LAGIDSTR_MAX+1];
1446 if (lp->lp_aggregator) {
1450 /* If we haven't heard from our peer, skip this step. */
1451 if (lp->lp_state & LACP_STATE_DEFAULTED)
1454 KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1455 ("timer_wait_while still active"));
1457 LACP_DPRINTF((lp, "port lagid=%s\n",
1458 lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1459 buf, sizeof(buf))));
1461 TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1462 if (lacp_aggregator_is_compatible(la, lp)) {
1468 la = lacp_aggregator_get(lsc, lp);
1470 LACP_DPRINTF((lp, "aggregator creation failed\n"));
1473 * will retry on the next tick.
1478 lacp_fill_aggregator_id(la, lp);
1479 LACP_DPRINTF((lp, "aggregator created\n"));
1481 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1482 if (la->la_refcnt == LACP_MAX_PORTS)
1484 lacp_aggregator_addref(lsc, la);
1487 LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1488 lacp_format_lagid(&la->la_actor, &la->la_partner,
1489 buf, sizeof(buf))));
1491 lp->lp_aggregator = la;
1492 lp->lp_selected = LACP_SELECTED;
1496 * lacp_unselect: finish unselect/detach process.
1500 lacp_unselect(struct lacp_port *lp)
1502 struct lacp_softc *lsc = lp->lp_lsc;
1503 struct lacp_aggregator *la = lp->lp_aggregator;
1505 KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1506 ("timer_wait_while still active"));
1512 lp->lp_aggregator = NULL;
1513 lacp_aggregator_delref(lsc, la);
1519 lacp_sm_mux(struct lacp_port *lp)
1521 struct lagg_port *lgp = lp->lp_lagg;
1522 struct lagg_softc *sc = lgp->lp_softc;
1523 enum lacp_mux_state new_state;
1525 (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1526 boolean_t p_collecting =
1527 (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1528 enum lacp_selected selected = lp->lp_selected;
1529 struct lacp_aggregator *la;
1531 if (V_lacp_debug > 1)
1532 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1533 "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1534 lp->lp_mux_state, selected, p_sync, p_collecting);
1537 la = lp->lp_aggregator;
1538 KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1539 ("MUX not detached"));
1540 new_state = lp->lp_mux_state;
1541 switch (lp->lp_mux_state) {
1542 case LACP_MUX_DETACHED:
1543 if (selected != LACP_UNSELECTED) {
1544 new_state = LACP_MUX_WAITING;
1547 case LACP_MUX_WAITING:
1548 KASSERT(la->la_pending > 0 ||
1549 !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1550 ("timer_wait_while still active"));
1551 if (selected == LACP_SELECTED && la->la_pending == 0) {
1552 new_state = LACP_MUX_ATTACHED;
1553 } else if (selected == LACP_UNSELECTED) {
1554 new_state = LACP_MUX_DETACHED;
1557 case LACP_MUX_ATTACHED:
1558 if (selected == LACP_SELECTED && p_sync) {
1559 new_state = LACP_MUX_COLLECTING;
1560 } else if (selected != LACP_SELECTED) {
1561 new_state = LACP_MUX_DETACHED;
1564 case LACP_MUX_COLLECTING:
1565 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1566 new_state = LACP_MUX_DISTRIBUTING;
1567 } else if (selected != LACP_SELECTED || !p_sync) {
1568 new_state = LACP_MUX_ATTACHED;
1571 case LACP_MUX_DISTRIBUTING:
1572 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1573 new_state = LACP_MUX_COLLECTING;
1574 lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1579 panic("%s: unknown state", __func__);
1582 if (lp->lp_mux_state == new_state) {
1586 lacp_set_mux(lp, new_state);
1591 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1593 struct lacp_aggregator *la = lp->lp_aggregator;
1595 if (lp->lp_mux_state == new_state) {
1599 switch (new_state) {
1600 case LACP_MUX_DETACHED:
1601 lp->lp_state &= ~LACP_STATE_SYNC;
1602 lacp_disable_distributing(lp);
1603 lacp_disable_collecting(lp);
1604 lacp_sm_assert_ntt(lp);
1606 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1607 KASSERT(la->la_pending > 0,
1608 ("timer_wait_while not active"));
1611 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1614 case LACP_MUX_WAITING:
1615 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1616 LACP_AGGREGATE_WAIT_TIME);
1619 case LACP_MUX_ATTACHED:
1620 lp->lp_state |= LACP_STATE_SYNC;
1621 lacp_disable_collecting(lp);
1622 lacp_sm_assert_ntt(lp);
1624 case LACP_MUX_COLLECTING:
1625 lacp_enable_collecting(lp);
1626 lacp_disable_distributing(lp);
1627 lacp_sm_assert_ntt(lp);
1629 case LACP_MUX_DISTRIBUTING:
1630 lacp_enable_distributing(lp);
1633 panic("%s: unknown state", __func__);
1636 LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1638 lp->lp_mux_state = new_state;
1642 lacp_sm_mux_timer(struct lacp_port *lp)
1644 struct lacp_aggregator *la = lp->lp_aggregator;
1645 char buf[LACP_LAGIDSTR_MAX+1];
1647 KASSERT(la->la_pending > 0, ("no pending event"));
1649 LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1650 lacp_format_lagid(&la->la_actor, &la->la_partner,
1652 la->la_pending, la->la_pending - 1));
1657 /* periodic transmit machine */
1660 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1662 if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1663 LACP_STATE_TIMEOUT)) {
1667 LACP_DPRINTF((lp, "partner timeout changed\n"));
1670 * FAST_PERIODIC -> SLOW_PERIODIC
1672 * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1674 * let lacp_sm_ptx_tx_schedule to update timeout.
1677 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1680 * if timeout has been shortened, assert NTT.
1683 if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1684 lacp_sm_assert_ntt(lp);
1689 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1693 if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1694 !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1700 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1704 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1708 timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1709 LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1711 LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1715 lacp_sm_ptx_timer(struct lacp_port *lp)
1717 lacp_sm_assert_ntt(lp);
1721 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1726 * check LACP_DISABLED first
1729 if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1734 * check loopback condition.
1737 if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1738 &lp->lp_actor.lip_systemid)) {
1743 * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1746 lacp_sm_rx_update_selected(lp, du);
1747 lacp_sm_rx_update_ntt(lp, du);
1748 lacp_sm_rx_record_pdu(lp, du);
1750 timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1751 LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1752 LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1754 lp->lp_state &= ~LACP_STATE_EXPIRED;
1757 * kick transmit machine without waiting the next tick.
1764 lacp_sm_rx_set_expired(struct lacp_port *lp)
1766 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1767 lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1768 LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1769 lp->lp_state |= LACP_STATE_EXPIRED;
1773 lacp_sm_rx_timer(struct lacp_port *lp)
1775 if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1776 /* CURRENT -> EXPIRED */
1777 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1778 lacp_sm_rx_set_expired(lp);
1780 /* EXPIRED -> DEFAULTED */
1781 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1782 lacp_sm_rx_update_default_selected(lp);
1783 lacp_sm_rx_record_default(lp);
1784 lp->lp_state &= ~LACP_STATE_EXPIRED;
1789 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1793 char buf[LACP_STATESTR_MAX+1];
1797 oldpstate = lp->lp_partner.lip_state;
1799 active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1800 || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1801 (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1803 lp->lp_partner = du->ldu_actor;
1805 ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1806 LACP_STATE_AGGREGATION) &&
1807 !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1808 || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1810 * XXX Maintain legacy behavior of leaving the
1811 * LACP_STATE_SYNC bit unchanged from the partner's
1812 * advertisement if lsc_strict_mode is false.
1813 * TODO: We should re-examine the concept of the "strict mode"
1814 * to ensure it makes sense to maintain a non-strict mode.
1816 if (lp->lp_lsc->lsc_strict_mode)
1817 lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1819 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1822 lp->lp_state &= ~LACP_STATE_DEFAULTED;
1824 if (oldpstate != lp->lp_partner.lip_state) {
1825 LACP_DPRINTF((lp, "old pstate %s\n",
1826 lacp_format_state(oldpstate, buf, sizeof(buf))));
1827 LACP_DPRINTF((lp, "new pstate %s\n",
1828 lacp_format_state(lp->lp_partner.lip_state, buf,
1832 lacp_sm_ptx_update_timeout(lp, oldpstate);
1836 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1841 if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1842 !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1843 LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1844 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1845 lacp_sm_assert_ntt(lp);
1850 lacp_sm_rx_record_default(struct lacp_port *lp)
1856 oldpstate = lp->lp_partner.lip_state;
1857 if (lp->lp_lsc->lsc_strict_mode)
1858 lp->lp_partner = lacp_partner_admin_strict;
1860 lp->lp_partner = lacp_partner_admin_optimistic;
1861 lp->lp_state |= LACP_STATE_DEFAULTED;
1862 lacp_sm_ptx_update_timeout(lp, oldpstate);
1866 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1867 const struct lacp_peerinfo *info)
1872 if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1873 !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1874 LACP_STATE_AGGREGATION)) {
1875 lp->lp_selected = LACP_UNSELECTED;
1876 /* mux machine will clean up lp->lp_aggregator */
1881 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1886 lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1890 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1895 if (lp->lp_lsc->lsc_strict_mode)
1896 lacp_sm_rx_update_selected_from_peerinfo(lp,
1897 &lacp_partner_admin_strict);
1899 lacp_sm_rx_update_selected_from_peerinfo(lp,
1900 &lacp_partner_admin_optimistic);
1903 /* transmit machine */
1906 lacp_sm_tx(struct lacp_port *lp)
1910 if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1912 || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1913 && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1916 lp->lp_flags &= ~LACP_PORT_NTT;
1919 if (!(lp->lp_flags & LACP_PORT_NTT)) {
1923 /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1924 if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1925 (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1926 LACP_DPRINTF((lp, "rate limited pdu\n"));
1930 if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1931 error = lacp_xmit_lacpdu(lp);
1933 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1937 lp->lp_flags &= ~LACP_PORT_NTT;
1939 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1945 lacp_sm_assert_ntt(struct lacp_port *lp)
1948 lp->lp_flags |= LACP_PORT_NTT;
1952 lacp_run_timers(struct lacp_port *lp)
1956 for (i = 0; i < LACP_NTIMER; i++) {
1957 KASSERT(lp->lp_timer[i] >= 0,
1958 ("invalid timer value %d", lp->lp_timer[i]));
1959 if (lp->lp_timer[i] == 0) {
1961 } else if (--lp->lp_timer[i] <= 0) {
1962 if (lacp_timer_funcs[i]) {
1963 (*lacp_timer_funcs[i])(lp);
1970 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1972 struct lacp_softc *lsc = lp->lp_lsc;
1973 struct lagg_port *lgp = lp->lp_lagg;
1974 struct lacp_port *lp2;
1975 struct markerdu *mdu;
1979 if (m->m_pkthdr.len != sizeof(*mdu)) {
1983 if ((m->m_flags & M_MCAST) == 0) {
1987 if (m->m_len < sizeof(*mdu)) {
1988 m = m_pullup(m, sizeof(*mdu));
1994 mdu = mtod(m, struct markerdu *);
1996 if (memcmp(&mdu->mdu_eh.ether_dhost,
1997 ðermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
2001 if (mdu->mdu_sph.sph_version != 1) {
2005 switch (mdu->mdu_tlv.tlv_type) {
2006 case MARKER_TYPE_INFO:
2007 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2008 marker_info_tlv_template, TRUE)) {
2011 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
2012 memcpy(&mdu->mdu_eh.ether_dhost,
2013 ðermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
2014 memcpy(&mdu->mdu_eh.ether_shost,
2015 lgp->lp_lladdr, ETHER_ADDR_LEN);
2016 error = lagg_enqueue(lp->lp_ifp, m);
2019 case MARKER_TYPE_RESPONSE:
2020 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2021 marker_response_tlv_template, TRUE)) {
2024 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
2025 ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
2026 ":", ntohl(mdu->mdu_info.mi_rq_xid)));
2028 /* Verify that it is the last marker we sent out */
2029 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
2030 sizeof(struct lacp_markerinfo)))
2034 lp->lp_flags &= ~LACP_PORT_MARK;
2036 if (lsc->lsc_suppress_distributing) {
2037 /* Check if any ports are waiting for a response */
2038 LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
2039 if (lp2->lp_flags & LACP_PORT_MARK) {
2046 /* All interface queues are clear */
2047 LACP_DPRINTF((NULL, "queue flush complete\n"));
2048 lsc->lsc_suppress_distributing = FALSE;
2062 LACP_DPRINTF((lp, "bad marker frame\n"));
2068 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
2069 const struct tlv_template *tmpl, boolean_t check_type)
2071 while (/* CONSTCOND */ 1) {
2072 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
2075 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
2076 tlv->tlv_length != tmpl->tmpl_length) {
2079 if (tmpl->tmpl_type == 0) {
2082 tlv = (const struct tlvhdr *)
2083 ((const char *)tlv + tlv->tlv_length);
2092 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
2094 snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
2106 lacp_format_systemid(const struct lacp_systemid *sysid,
2107 char *buf, size_t buflen)
2109 char macbuf[LACP_MACSTR_MAX+1];
2111 snprintf(buf, buflen, "%04X,%s",
2112 ntohs(sysid->lsi_prio),
2113 lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2119 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2121 snprintf(buf, buflen, "%04X,%04X",
2122 ntohs(portid->lpi_prio),
2123 ntohs(portid->lpi_portno));
2129 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2131 char sysid[LACP_SYSTEMIDSTR_MAX+1];
2132 char portid[LACP_PORTIDSTR_MAX+1];
2134 snprintf(buf, buflen, "(%s,%04X,%s)",
2135 lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2136 ntohs(peer->lip_key),
2137 lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2143 lacp_format_lagid(const struct lacp_peerinfo *a,
2144 const struct lacp_peerinfo *b, char *buf, size_t buflen)
2146 char astr[LACP_PARTNERSTR_MAX+1];
2147 char bstr[LACP_PARTNERSTR_MAX+1];
2151 * there's a convention to display small numbered peer
2155 if (lacp_compare_peerinfo(a, b) > 0) {
2156 const struct lacp_peerinfo *t;
2164 snprintf(buf, buflen, "[%s,%s]",
2165 lacp_format_partner(a, astr, sizeof(astr)),
2166 lacp_format_partner(b, bstr, sizeof(bstr)));
2172 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2173 char *buf, size_t buflen)
2179 return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2183 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2185 snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2190 lacp_dump_lacpdu(const struct lacpdu *du)
2192 char buf[LACP_PARTNERSTR_MAX+1];
2193 char buf2[LACP_STATESTR_MAX+1];
2195 printf("actor=%s\n",
2196 lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2197 printf("actor.state=%s\n",
2198 lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2199 printf("partner=%s\n",
2200 lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2201 printf("partner.state=%s\n",
2202 lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2204 printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2208 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2213 printf("%s: ", lp->lp_ifp->if_xname);