xref: /freebsd/sys/net/ieee8023ad_lacp.c (revision 3e0efd2ec4fcb4cd68fb8ccf8aea6fc6151c454b)
1 /*	$NetBSD: ieee8023ad_lacp.c,v 1.3 2005/12/11 12:24:54 christos Exp $	*/
2 
3 /*-
4  * Copyright (c)2005 YAMAMOTO Takashi,
5  * Copyright (c)2008 Andrew Thompson <thompsa@FreeBSD.org>
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
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.
16  *
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
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include <sys/param.h>
34 #include <sys/callout.h>
35 #include <sys/mbuf.h>
36 #include <sys/systm.h>
37 #include <sys/malloc.h>
38 #include <sys/kernel.h> /* hz */
39 #include <sys/socket.h> /* for net/if.h */
40 #include <sys/sockio.h>
41 #include <sys/sysctl.h>
42 #include <machine/stdarg.h>
43 #include <sys/lock.h>
44 #include <sys/rwlock.h>
45 
46 #include <net/if.h>
47 #include <net/if_dl.h>
48 #include <net/ethernet.h>
49 #include <net/if_media.h>
50 #include <net/if_types.h>
51 
52 #include <net/if_lagg.h>
53 #include <net/ieee8023ad_lacp.h>
54 
55 /*
56  * actor system priority and port priority.
57  * XXX should be configurable.
58  */
59 
60 #define	LACP_SYSTEM_PRIO	0x8000
61 #define	LACP_PORT_PRIO		0x8000
62 
63 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
64     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
65 
66 static const struct tlv_template lacp_info_tlv_template[] = {
67 	{ LACP_TYPE_ACTORINFO,
68 	    sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
69 	{ LACP_TYPE_PARTNERINFO,
70 	    sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
71 	{ LACP_TYPE_COLLECTORINFO,
72 	    sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
73 	{ 0, 0 },
74 };
75 
76 static const struct tlv_template marker_info_tlv_template[] = {
77 	{ MARKER_TYPE_INFO,
78 	    sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
79 	{ 0, 0 },
80 };
81 
82 static const struct tlv_template marker_response_tlv_template[] = {
83 	{ MARKER_TYPE_RESPONSE,
84 	    sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
85 	{ 0, 0 },
86 };
87 
88 typedef void (*lacp_timer_func_t)(struct lacp_port *);
89 
90 static void	lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
91 static void	lacp_fill_markerinfo(struct lacp_port *,
92 		    struct lacp_markerinfo *);
93 
94 static uint64_t	lacp_aggregator_bandwidth(struct lacp_aggregator *);
95 static void	lacp_suppress_distributing(struct lacp_softc *,
96 		    struct lacp_aggregator *);
97 static void	lacp_transit_expire(void *);
98 static void	lacp_update_portmap(struct lacp_softc *);
99 static void	lacp_select_active_aggregator(struct lacp_softc *);
100 static uint16_t	lacp_compose_key(struct lacp_port *);
101 static int	tlv_check(const void *, size_t, const struct tlvhdr *,
102 		    const struct tlv_template *, boolean_t);
103 static void	lacp_tick(void *);
104 
105 static void	lacp_fill_aggregator_id(struct lacp_aggregator *,
106 		    const struct lacp_port *);
107 static void	lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
108 		    const struct lacp_peerinfo *);
109 static int	lacp_aggregator_is_compatible(const struct lacp_aggregator *,
110 		    const struct lacp_port *);
111 static int	lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
112 		    const struct lacp_peerinfo *);
113 
114 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
115 		    struct lacp_port *);
116 static void	lacp_aggregator_addref(struct lacp_softc *,
117 		    struct lacp_aggregator *);
118 static void	lacp_aggregator_delref(struct lacp_softc *,
119 		    struct lacp_aggregator *);
120 
121 /* receive machine */
122 
123 static int	lacp_pdu_input(struct lacp_port *, struct mbuf *);
124 static int	lacp_marker_input(struct lacp_port *, struct mbuf *);
125 static void	lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
126 static void	lacp_sm_rx_timer(struct lacp_port *);
127 static void	lacp_sm_rx_set_expired(struct lacp_port *);
128 static void	lacp_sm_rx_update_ntt(struct lacp_port *,
129 		    const struct lacpdu *);
130 static void	lacp_sm_rx_record_pdu(struct lacp_port *,
131 		    const struct lacpdu *);
132 static void	lacp_sm_rx_update_selected(struct lacp_port *,
133 		    const struct lacpdu *);
134 static void	lacp_sm_rx_record_default(struct lacp_port *);
135 static void	lacp_sm_rx_update_default_selected(struct lacp_port *);
136 static void	lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
137 		    const struct lacp_peerinfo *);
138 
139 /* mux machine */
140 
141 static void	lacp_sm_mux(struct lacp_port *);
142 static void	lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
143 static void	lacp_sm_mux_timer(struct lacp_port *);
144 
145 /* periodic transmit machine */
146 
147 static void	lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
148 static void	lacp_sm_ptx_tx_schedule(struct lacp_port *);
149 static void	lacp_sm_ptx_timer(struct lacp_port *);
150 
151 /* transmit machine */
152 
153 static void	lacp_sm_tx(struct lacp_port *);
154 static void	lacp_sm_assert_ntt(struct lacp_port *);
155 
156 static void	lacp_run_timers(struct lacp_port *);
157 static int	lacp_compare_peerinfo(const struct lacp_peerinfo *,
158 		    const struct lacp_peerinfo *);
159 static int	lacp_compare_systemid(const struct lacp_systemid *,
160 		    const struct lacp_systemid *);
161 static void	lacp_port_enable(struct lacp_port *);
162 static void	lacp_port_disable(struct lacp_port *);
163 static void	lacp_select(struct lacp_port *);
164 static void	lacp_unselect(struct lacp_port *);
165 static void	lacp_disable_collecting(struct lacp_port *);
166 static void	lacp_enable_collecting(struct lacp_port *);
167 static void	lacp_disable_distributing(struct lacp_port *);
168 static void	lacp_enable_distributing(struct lacp_port *);
169 static int	lacp_xmit_lacpdu(struct lacp_port *);
170 static int	lacp_xmit_marker(struct lacp_port *);
171 
172 /* Debugging */
173 
174 static void	lacp_dump_lacpdu(const struct lacpdu *);
175 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
176 		    size_t);
177 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
178 		    const struct lacp_peerinfo *, char *, size_t);
179 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
180 		    char *, size_t);
181 static const char *lacp_format_state(uint8_t, char *, size_t);
182 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
183 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
184 		    size_t);
185 static const char *lacp_format_portid(const struct lacp_portid *, char *,
186 		    size_t);
187 static void	lacp_dprintf(const struct lacp_port *, const char *, ...)
188 		    __attribute__((__format__(__printf__, 2, 3)));
189 
190 static int lacp_debug = 0;
191 SYSCTL_INT(_net, OID_AUTO, lacp_debug, CTLFLAG_RW | CTLFLAG_TUN,
192     &lacp_debug, 0, "Enable LACP debug logging (1=debug, 2=trace)");
193 TUNABLE_INT("net.lacp_debug", &lacp_debug);
194 
195 #define LACP_DPRINTF(a) if (lacp_debug > 0) { lacp_dprintf a ; }
196 #define LACP_TRACE(a) if (lacp_debug > 1) { lacp_dprintf(a,"%s\n",__func__); }
197 
198 /*
199  * partner administration variables.
200  * XXX should be configurable.
201  */
202 
203 static const struct lacp_peerinfo lacp_partner_admin = {
204 	.lip_systemid = { .lsi_prio = 0xffff },
205 	.lip_portid = { .lpi_prio = 0xffff },
206 #if 1
207 	/* optimistic */
208 	.lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
209 	    LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
210 #else
211 	/* pessimistic */
212 	.lip_state = 0,
213 #endif
214 };
215 
216 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
217 	[LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
218 	[LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
219 	[LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
220 };
221 
222 struct mbuf *
223 lacp_input(struct lagg_port *lgp, struct mbuf *m)
224 {
225 	struct lacp_port *lp = LACP_PORT(lgp);
226 	uint8_t subtype;
227 
228 	if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
229 		m_freem(m);
230 		return (NULL);
231 	}
232 
233 	m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
234 	switch (subtype) {
235 		case SLOWPROTOCOLS_SUBTYPE_LACP:
236 			lacp_pdu_input(lp, m);
237 			return (NULL);
238 
239 		case SLOWPROTOCOLS_SUBTYPE_MARKER:
240 			lacp_marker_input(lp, m);
241 			return (NULL);
242 	}
243 
244 	/* Not a subtype we are interested in */
245 	return (m);
246 }
247 
248 /*
249  * lacp_pdu_input: process lacpdu
250  */
251 static int
252 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
253 {
254 	struct lacp_softc *lsc = lp->lp_lsc;
255 	struct lacpdu *du;
256 	int error = 0;
257 
258 	if (m->m_pkthdr.len != sizeof(*du)) {
259 		goto bad;
260 	}
261 
262 	if ((m->m_flags & M_MCAST) == 0) {
263 		goto bad;
264 	}
265 
266 	if (m->m_len < sizeof(*du)) {
267 		m = m_pullup(m, sizeof(*du));
268 		if (m == NULL) {
269 			return (ENOMEM);
270 		}
271 	}
272 
273 	du = mtod(m, struct lacpdu *);
274 
275 	if (memcmp(&du->ldu_eh.ether_dhost,
276 	    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
277 		goto bad;
278 	}
279 
280 	/*
281 	 * ignore the version for compatibility with
282 	 * the future protocol revisions.
283 	 */
284 #if 0
285 	if (du->ldu_sph.sph_version != 1) {
286 		goto bad;
287 	}
288 #endif
289 
290 	/*
291 	 * ignore tlv types for compatibility with
292 	 * the future protocol revisions.
293 	 */
294 	if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
295 	    lacp_info_tlv_template, FALSE)) {
296 		goto bad;
297 	}
298 
299         if (lacp_debug > 0) {
300 		lacp_dprintf(lp, "lacpdu receive\n");
301 		lacp_dump_lacpdu(du);
302 	}
303 
304 	LACP_LOCK(lsc);
305 	lacp_sm_rx(lp, du);
306 	LACP_UNLOCK(lsc);
307 
308 	m_freem(m);
309 	return (error);
310 
311 bad:
312 	m_freem(m);
313 	return (EINVAL);
314 }
315 
316 static void
317 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
318 {
319 	struct lagg_port *lgp = lp->lp_lagg;
320 	struct lagg_softc *sc = lgp->lp_softc;
321 
322 	info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
323 	memcpy(&info->lip_systemid.lsi_mac,
324 	    IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
325 	info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
326 	info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
327 	info->lip_state = lp->lp_state;
328 }
329 
330 static void
331 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
332 {
333 	struct ifnet *ifp = lp->lp_ifp;
334 
335 	/* Fill in the port index and system id (encoded as the MAC) */
336 	info->mi_rq_port = htons(ifp->if_index);
337 	memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
338 	info->mi_rq_xid = htonl(0);
339 }
340 
341 static int
342 lacp_xmit_lacpdu(struct lacp_port *lp)
343 {
344 	struct lagg_port *lgp = lp->lp_lagg;
345 	struct mbuf *m;
346 	struct lacpdu *du;
347 	int error;
348 
349 	LACP_LOCK_ASSERT(lp->lp_lsc);
350 
351 	m = m_gethdr(M_DONTWAIT, MT_DATA);
352 	if (m == NULL) {
353 		return (ENOMEM);
354 	}
355 	m->m_len = m->m_pkthdr.len = sizeof(*du);
356 
357 	du = mtod(m, struct lacpdu *);
358 	memset(du, 0, sizeof(*du));
359 
360 	memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
361 	    ETHER_ADDR_LEN);
362 	memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
363 	du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
364 
365 	du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
366 	du->ldu_sph.sph_version = 1;
367 
368 	TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
369 	du->ldu_actor = lp->lp_actor;
370 
371 	TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
372 	    sizeof(du->ldu_partner));
373 	du->ldu_partner = lp->lp_partner;
374 
375 	TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
376 	    sizeof(du->ldu_collector));
377 	du->ldu_collector.lci_maxdelay = 0;
378 
379 	if (lacp_debug > 0) {
380 		lacp_dprintf(lp, "lacpdu transmit\n");
381 		lacp_dump_lacpdu(du);
382 	}
383 
384 	m->m_flags |= M_MCAST;
385 
386 	/*
387 	 * XXX should use higher priority queue.
388 	 * otherwise network congestion can break aggregation.
389 	 */
390 
391 	error = lagg_enqueue(lp->lp_ifp, m);
392 	return (error);
393 }
394 
395 static int
396 lacp_xmit_marker(struct lacp_port *lp)
397 {
398 	struct lagg_port *lgp = lp->lp_lagg;
399 	struct mbuf *m;
400 	struct markerdu *mdu;
401 	int error;
402 
403 	LACP_LOCK_ASSERT(lp->lp_lsc);
404 
405 	m = m_gethdr(M_DONTWAIT, MT_DATA);
406 	if (m == NULL) {
407 		return (ENOMEM);
408 	}
409 	m->m_len = m->m_pkthdr.len = sizeof(*mdu);
410 
411 	mdu = mtod(m, struct markerdu *);
412 	memset(mdu, 0, sizeof(*mdu));
413 
414 	memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
415 	    ETHER_ADDR_LEN);
416 	memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
417 	mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
418 
419 	mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
420 	mdu->mdu_sph.sph_version = 1;
421 
422 	/* Bump the transaction id and copy over the marker info */
423 	lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
424 	TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
425 	mdu->mdu_info = lp->lp_marker;
426 
427 	LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
428 	    ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
429 	    ntohl(mdu->mdu_info.mi_rq_xid)));
430 
431 	m->m_flags |= M_MCAST;
432 	error = lagg_enqueue(lp->lp_ifp, m);
433 	return (error);
434 }
435 
436 void
437 lacp_linkstate(struct lagg_port *lgp)
438 {
439 	struct lacp_port *lp = LACP_PORT(lgp);
440 	struct lacp_softc *lsc = lp->lp_lsc;
441 	struct ifnet *ifp = lgp->lp_ifp;
442 	struct ifmediareq ifmr;
443 	int error = 0;
444 	u_int media;
445 	uint8_t old_state;
446 	uint16_t old_key;
447 
448 	bzero((char *)&ifmr, sizeof(ifmr));
449 	error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
450 	if (error != 0)
451 		return;
452 
453 	LACP_LOCK(lsc);
454 	media = ifmr.ifm_active;
455 	LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
456 	    "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
457 	    (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
458 	old_state = lp->lp_state;
459 	old_key = lp->lp_key;
460 
461 	lp->lp_media = media;
462 	/*
463 	 * If the port is not an active full duplex Ethernet link then it can
464 	 * not be aggregated.
465 	 */
466 	if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
467 	    ifp->if_link_state != LINK_STATE_UP) {
468 		lacp_port_disable(lp);
469 	} else {
470 		lacp_port_enable(lp);
471 	}
472 	lp->lp_key = lacp_compose_key(lp);
473 
474 	if (old_state != lp->lp_state || old_key != lp->lp_key) {
475 		LACP_DPRINTF((lp, "-> UNSELECTED\n"));
476 		lp->lp_selected = LACP_UNSELECTED;
477 	}
478 	LACP_UNLOCK(lsc);
479 }
480 
481 static void
482 lacp_tick(void *arg)
483 {
484 	struct lacp_softc *lsc = arg;
485 	struct lacp_port *lp;
486 
487 	LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
488 		if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
489 			continue;
490 
491 		lacp_run_timers(lp);
492 
493 		lacp_select(lp);
494 		lacp_sm_mux(lp);
495 		lacp_sm_tx(lp);
496 		lacp_sm_ptx_tx_schedule(lp);
497 	}
498 	callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
499 }
500 
501 int
502 lacp_port_create(struct lagg_port *lgp)
503 {
504 	struct lagg_softc *sc = lgp->lp_softc;
505 	struct lacp_softc *lsc = LACP_SOFTC(sc);
506 	struct lacp_port *lp;
507 	struct ifnet *ifp = lgp->lp_ifp;
508 	struct sockaddr_dl sdl;
509 	struct ifmultiaddr *rifma = NULL;
510 	int error;
511 
512 	boolean_t active = TRUE; /* XXX should be configurable */
513 	boolean_t fast = FALSE; /* XXX should be configurable */
514 
515 	bzero((char *)&sdl, sizeof(sdl));
516 	sdl.sdl_len = sizeof(sdl);
517 	sdl.sdl_family = AF_LINK;
518 	sdl.sdl_index = ifp->if_index;
519 	sdl.sdl_type = IFT_ETHER;
520 	sdl.sdl_alen = ETHER_ADDR_LEN;
521 
522 	bcopy(&ethermulticastaddr_slowprotocols,
523 	    LLADDR(&sdl), ETHER_ADDR_LEN);
524 	error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
525 	if (error) {
526 		printf("%s: ADDMULTI failed on %s\n", __func__, lgp->lp_ifname);
527 		return (error);
528 	}
529 
530 	lp = malloc(sizeof(struct lacp_port),
531 	    M_DEVBUF, M_NOWAIT|M_ZERO);
532 	if (lp == NULL)
533 		return (ENOMEM);
534 
535 	LACP_LOCK(lsc);
536 	lgp->lp_psc = (caddr_t)lp;
537 	lp->lp_ifp = ifp;
538 	lp->lp_lagg = lgp;
539 	lp->lp_lsc = lsc;
540 	lp->lp_ifma = rifma;
541 
542 	LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
543 
544 	lacp_fill_actorinfo(lp, &lp->lp_actor);
545 	lacp_fill_markerinfo(lp, &lp->lp_marker);
546 	lp->lp_state =
547 	    (active ? LACP_STATE_ACTIVITY : 0) |
548 	    (fast ? LACP_STATE_TIMEOUT : 0);
549 	lp->lp_aggregator = NULL;
550 	lacp_sm_rx_set_expired(lp);
551 	LACP_UNLOCK(lsc);
552 	lacp_linkstate(lgp);
553 
554 	return (0);
555 }
556 
557 void
558 lacp_port_destroy(struct lagg_port *lgp)
559 {
560 	struct lacp_port *lp = LACP_PORT(lgp);
561 	struct lacp_softc *lsc = lp->lp_lsc;
562 	int i;
563 
564 	LACP_LOCK(lsc);
565 	for (i = 0; i < LACP_NTIMER; i++) {
566 		LACP_TIMER_DISARM(lp, i);
567 	}
568 
569 	lacp_disable_collecting(lp);
570 	lacp_disable_distributing(lp);
571 	lacp_unselect(lp);
572 
573 	/* The address may have already been removed by if_purgemaddrs() */
574 	if (!lgp->lp_detaching)
575 		if_delmulti_ifma(lp->lp_ifma);
576 
577 	LIST_REMOVE(lp, lp_next);
578 	LACP_UNLOCK(lsc);
579 	free(lp, M_DEVBUF);
580 }
581 
582 void
583 lacp_req(struct lagg_softc *sc, caddr_t data)
584 {
585 	struct lacp_opreq *req = (struct lacp_opreq *)data;
586 	struct lacp_softc *lsc = LACP_SOFTC(sc);
587 	struct lacp_aggregator *la = lsc->lsc_active_aggregator;
588 
589 	LACP_LOCK(lsc);
590 	bzero(req, sizeof(struct lacp_opreq));
591 	if (la != NULL) {
592 		req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
593 		memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
594 		    ETHER_ADDR_LEN);
595 		req->actor_key = ntohs(la->la_actor.lip_key);
596 		req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
597 		req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
598 		req->actor_state = la->la_actor.lip_state;
599 
600 		req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
601 		memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
602 		    ETHER_ADDR_LEN);
603 		req->partner_key = ntohs(la->la_partner.lip_key);
604 		req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
605 		req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
606 		req->partner_state = la->la_partner.lip_state;
607 	}
608 	LACP_UNLOCK(lsc);
609 }
610 
611 void
612 lacp_portreq(struct lagg_port *lgp, caddr_t data)
613 {
614 	struct lacp_opreq *req = (struct lacp_opreq *)data;
615 	struct lacp_port *lp = LACP_PORT(lgp);
616 	struct lacp_softc *lsc = lp->lp_lsc;
617 
618 	LACP_LOCK(lsc);
619 	req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
620 	memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
621 	    ETHER_ADDR_LEN);
622 	req->actor_key = ntohs(lp->lp_actor.lip_key);
623 	req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
624 	req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
625 	req->actor_state = lp->lp_actor.lip_state;
626 
627 	req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
628 	memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
629 	    ETHER_ADDR_LEN);
630 	req->partner_key = ntohs(lp->lp_partner.lip_key);
631 	req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
632 	req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
633 	req->partner_state = lp->lp_partner.lip_state;
634 	LACP_UNLOCK(lsc);
635 }
636 
637 static void
638 lacp_disable_collecting(struct lacp_port *lp)
639 {
640 	LACP_DPRINTF((lp, "collecting disabled\n"));
641 	lp->lp_state &= ~LACP_STATE_COLLECTING;
642 }
643 
644 static void
645 lacp_enable_collecting(struct lacp_port *lp)
646 {
647 	LACP_DPRINTF((lp, "collecting enabled\n"));
648 	lp->lp_state |= LACP_STATE_COLLECTING;
649 }
650 
651 static void
652 lacp_disable_distributing(struct lacp_port *lp)
653 {
654 	struct lacp_aggregator *la = lp->lp_aggregator;
655 	struct lacp_softc *lsc = lp->lp_lsc;
656 	char buf[LACP_LAGIDSTR_MAX+1];
657 
658 	LACP_LOCK_ASSERT(lsc);
659 
660 	if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
661 		return;
662 	}
663 
664 	KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
665 	KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
666 	KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
667 
668 	LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
669 	    "nports %d -> %d\n",
670 	    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
671 	    la->la_nports, la->la_nports - 1));
672 
673 	TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
674 	la->la_nports--;
675 
676 	if (lsc->lsc_active_aggregator == la) {
677 		lacp_suppress_distributing(lsc, la);
678 		lacp_select_active_aggregator(lsc);
679 		/* regenerate the port map, the active aggregator has changed */
680 		lacp_update_portmap(lsc);
681 	}
682 
683 	lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
684 }
685 
686 static void
687 lacp_enable_distributing(struct lacp_port *lp)
688 {
689 	struct lacp_aggregator *la = lp->lp_aggregator;
690 	struct lacp_softc *lsc = lp->lp_lsc;
691 	char buf[LACP_LAGIDSTR_MAX+1];
692 
693 	LACP_LOCK_ASSERT(lsc);
694 
695 	if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
696 		return;
697 	}
698 
699 	LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
700 	    "nports %d -> %d\n",
701 	    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
702 	    la->la_nports, la->la_nports + 1));
703 
704 	KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
705 	TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
706 	la->la_nports++;
707 
708 	lp->lp_state |= LACP_STATE_DISTRIBUTING;
709 
710 	if (lsc->lsc_active_aggregator == la) {
711 		lacp_suppress_distributing(lsc, la);
712 		lacp_update_portmap(lsc);
713 	} else
714 		/* try to become the active aggregator */
715 		lacp_select_active_aggregator(lsc);
716 }
717 
718 static void
719 lacp_transit_expire(void *vp)
720 {
721 	struct lacp_softc *lsc = vp;
722 
723 	LACP_LOCK_ASSERT(lsc);
724 
725 	LACP_TRACE(NULL);
726 
727 	lsc->lsc_suppress_distributing = FALSE;
728 }
729 
730 int
731 lacp_attach(struct lagg_softc *sc)
732 {
733 	struct lacp_softc *lsc;
734 
735 	lsc = malloc(sizeof(struct lacp_softc),
736 	    M_DEVBUF, M_NOWAIT|M_ZERO);
737 	if (lsc == NULL)
738 		return (ENOMEM);
739 
740 	sc->sc_psc = (caddr_t)lsc;
741 	lsc->lsc_softc = sc;
742 
743 	lsc->lsc_hashkey = arc4random();
744 	lsc->lsc_active_aggregator = NULL;
745 	LACP_LOCK_INIT(lsc);
746 	TAILQ_INIT(&lsc->lsc_aggregators);
747 	LIST_INIT(&lsc->lsc_ports);
748 
749 	callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
750 	callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
751 
752 	/* if the lagg is already up then do the same */
753 	if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
754 		lacp_init(sc);
755 
756 	return (0);
757 }
758 
759 int
760 lacp_detach(struct lagg_softc *sc)
761 {
762 	struct lacp_softc *lsc = LACP_SOFTC(sc);
763 
764 	KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
765 	    ("aggregators still active"));
766 	KASSERT(lsc->lsc_active_aggregator == NULL,
767 	    ("aggregator still attached"));
768 
769 	sc->sc_psc = NULL;
770 	callout_drain(&lsc->lsc_transit_callout);
771 	callout_drain(&lsc->lsc_callout);
772 
773 	LACP_LOCK_DESTROY(lsc);
774 	free(lsc, M_DEVBUF);
775 	return (0);
776 }
777 
778 void
779 lacp_init(struct lagg_softc *sc)
780 {
781 	struct lacp_softc *lsc = LACP_SOFTC(sc);
782 
783 	LACP_LOCK(lsc);
784 	callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
785 	LACP_UNLOCK(lsc);
786 }
787 
788 void
789 lacp_stop(struct lagg_softc *sc)
790 {
791 	struct lacp_softc *lsc = LACP_SOFTC(sc);
792 
793 	LACP_LOCK(lsc);
794 	callout_stop(&lsc->lsc_transit_callout);
795 	callout_stop(&lsc->lsc_callout);
796 	LACP_UNLOCK(lsc);
797 }
798 
799 struct lagg_port *
800 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
801 {
802 	struct lacp_softc *lsc = LACP_SOFTC(sc);
803 	struct lacp_portmap *pm;
804 	struct lacp_port *lp;
805 	uint32_t hash;
806 
807 	if (__predict_false(lsc->lsc_suppress_distributing)) {
808 		LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
809 		return (NULL);
810 	}
811 
812 	pm = &lsc->lsc_pmap[lsc->lsc_activemap];
813 	if (pm->pm_count == 0) {
814 		LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
815 		return (NULL);
816 	}
817 
818 	if (sc->use_flowid && (m->m_flags & M_FLOWID))
819 		hash = m->m_pkthdr.flowid;
820 	else
821 		hash = lagg_hashmbuf(sc, m, lsc->lsc_hashkey);
822 	hash %= pm->pm_count;
823 	lp = pm->pm_map[hash];
824 
825 	KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
826 	    ("aggregated port is not distributing"));
827 
828 	return (lp->lp_lagg);
829 }
830 /*
831  * lacp_suppress_distributing: drop transmit packets for a while
832  * to preserve packet ordering.
833  */
834 
835 static void
836 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
837 {
838 	struct lacp_port *lp;
839 
840 	if (lsc->lsc_active_aggregator != la) {
841 		return;
842 	}
843 
844 	LACP_TRACE(NULL);
845 
846 	lsc->lsc_suppress_distributing = TRUE;
847 
848 	/* send a marker frame down each port to verify the queues are empty */
849 	LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
850 		lp->lp_flags |= LACP_PORT_MARK;
851 		lacp_xmit_marker(lp);
852 	}
853 
854 	/* set a timeout for the marker frames */
855 	callout_reset(&lsc->lsc_transit_callout,
856 	    LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
857 }
858 
859 static int
860 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
861     const struct lacp_peerinfo *b)
862 {
863 	return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
864 }
865 
866 static int
867 lacp_compare_systemid(const struct lacp_systemid *a,
868     const struct lacp_systemid *b)
869 {
870 	return (memcmp(a, b, sizeof(*a)));
871 }
872 
873 #if 0	/* unused */
874 static int
875 lacp_compare_portid(const struct lacp_portid *a,
876     const struct lacp_portid *b)
877 {
878 	return (memcmp(a, b, sizeof(*a)));
879 }
880 #endif
881 
882 static uint64_t
883 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
884 {
885 	struct lacp_port *lp;
886 	uint64_t speed;
887 
888 	lp = TAILQ_FIRST(&la->la_ports);
889 	if (lp == NULL) {
890 		return (0);
891 	}
892 
893 	speed = ifmedia_baudrate(lp->lp_media);
894 	speed *= la->la_nports;
895 	if (speed == 0) {
896 		LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
897 		    lp->lp_media, la->la_nports));
898 	}
899 
900 	return (speed);
901 }
902 
903 /*
904  * lacp_select_active_aggregator: select an aggregator to be used to transmit
905  * packets from lagg(4) interface.
906  */
907 
908 static void
909 lacp_select_active_aggregator(struct lacp_softc *lsc)
910 {
911 	struct lagg_softc *sc = lsc->lsc_softc;
912 	struct lacp_aggregator *la;
913 	struct lacp_aggregator *best_la = NULL;
914 	uint64_t best_speed = 0;
915 	char buf[LACP_LAGIDSTR_MAX+1];
916 
917 	LACP_TRACE(NULL);
918 
919 	TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
920 		uint64_t speed;
921 
922 		if (la->la_nports == 0) {
923 			continue;
924 		}
925 
926 		speed = lacp_aggregator_bandwidth(la);
927 		LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
928 		    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
929 		    speed, la->la_nports));
930 
931 		/* This aggregator is chosen if
932 		 *      the partner has a better system priority
933 		 *  or, the total aggregated speed is higher
934 		 *  or, it is already the chosen aggregator
935 		 */
936 		if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
937 		     LACP_SYS_PRI(best_la->la_partner)) ||
938 		    speed > best_speed ||
939 		    (speed == best_speed &&
940 		    la == lsc->lsc_active_aggregator)) {
941 			best_la = la;
942 			best_speed = speed;
943 		}
944 	}
945 
946 	KASSERT(best_la == NULL || best_la->la_nports > 0,
947 	    ("invalid aggregator refcnt"));
948 	KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
949 	    ("invalid aggregator list"));
950 
951 	if (lsc->lsc_active_aggregator != best_la) {
952 		LACP_DPRINTF((NULL, "active aggregator changed\n"));
953 		LACP_DPRINTF((NULL, "old %s\n",
954 		    lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
955 		    buf, sizeof(buf))));
956 	} else {
957 		LACP_DPRINTF((NULL, "active aggregator not changed\n"));
958 	}
959 	LACP_DPRINTF((NULL, "new %s\n",
960 	    lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
961 
962 	if (lsc->lsc_active_aggregator != best_la) {
963 		sc->sc_ifp->if_baudrate = best_speed;
964 		lsc->lsc_active_aggregator = best_la;
965 		lacp_update_portmap(lsc);
966 		if (best_la) {
967 			lacp_suppress_distributing(lsc, best_la);
968 		}
969 	}
970 }
971 
972 /*
973  * Updated the inactive portmap array with the new list of ports and
974  * make it live.
975  */
976 static void
977 lacp_update_portmap(struct lacp_softc *lsc)
978 {
979 	struct lacp_aggregator *la;
980 	struct lacp_portmap *p;
981 	struct lacp_port *lp;
982 	u_int newmap;
983 	int i;
984 
985 	newmap = lsc->lsc_activemap == 0 ? 1 : 0;
986 	p = &lsc->lsc_pmap[newmap];
987 	la = lsc->lsc_active_aggregator;
988 	bzero(p, sizeof(struct lacp_portmap));
989 
990 	if (la != NULL && la->la_nports > 0) {
991 		p->pm_count = la->la_nports;
992 		i = 0;
993 		TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
994 			p->pm_map[i++] = lp;
995 		KASSERT(i == p->pm_count, ("Invalid port count"));
996 	}
997 
998 	/* switch the active portmap over */
999 	atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1000 	LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1001 		    lsc->lsc_activemap,
1002 		    lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1003 }
1004 
1005 static uint16_t
1006 lacp_compose_key(struct lacp_port *lp)
1007 {
1008 	struct lagg_port *lgp = lp->lp_lagg;
1009 	struct lagg_softc *sc = lgp->lp_softc;
1010 	u_int media = lp->lp_media;
1011 	uint16_t key;
1012 
1013 	if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1014 
1015 		/*
1016 		 * non-aggregatable links should have unique keys.
1017 		 *
1018 		 * XXX this isn't really unique as if_index is 16 bit.
1019 		 */
1020 
1021 		/* bit 0..14:	(some bits of) if_index of this port */
1022 		key = lp->lp_ifp->if_index;
1023 		/* bit 15:	1 */
1024 		key |= 0x8000;
1025 	} else {
1026 		u_int subtype = IFM_SUBTYPE(media);
1027 
1028 		KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1029 		KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1030 
1031 		/* bit 0..4:	IFM_SUBTYPE */
1032 		key = subtype;
1033 		/* bit 5..14:	(some bits of) if_index of lagg device */
1034 		key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1035 		/* bit 15:	0 */
1036 	}
1037 	return (htons(key));
1038 }
1039 
1040 static void
1041 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1042 {
1043 	char buf[LACP_LAGIDSTR_MAX+1];
1044 
1045 	LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1046 	    __func__,
1047 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1048 	    buf, sizeof(buf)),
1049 	    la->la_refcnt, la->la_refcnt + 1));
1050 
1051 	KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1052 	la->la_refcnt++;
1053 	KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1054 }
1055 
1056 static void
1057 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1058 {
1059 	char buf[LACP_LAGIDSTR_MAX+1];
1060 
1061 	LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1062 	    __func__,
1063 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1064 	    buf, sizeof(buf)),
1065 	    la->la_refcnt, la->la_refcnt - 1));
1066 
1067 	KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1068 	la->la_refcnt--;
1069 	if (la->la_refcnt > 0) {
1070 		return;
1071 	}
1072 
1073 	KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1074 	KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1075 
1076 	TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1077 
1078 	free(la, M_DEVBUF);
1079 }
1080 
1081 /*
1082  * lacp_aggregator_get: allocate an aggregator.
1083  */
1084 
1085 static struct lacp_aggregator *
1086 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1087 {
1088 	struct lacp_aggregator *la;
1089 
1090 	la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1091 	if (la) {
1092 		la->la_refcnt = 1;
1093 		la->la_nports = 0;
1094 		TAILQ_INIT(&la->la_ports);
1095 		la->la_pending = 0;
1096 		TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1097 	}
1098 
1099 	return (la);
1100 }
1101 
1102 /*
1103  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1104  */
1105 
1106 static void
1107 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1108 {
1109 	lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1110 	lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1111 
1112 	la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1113 }
1114 
1115 static void
1116 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1117     const struct lacp_peerinfo *lpi_port)
1118 {
1119 	memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1120 	lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1121 	lpi_aggr->lip_key = lpi_port->lip_key;
1122 }
1123 
1124 /*
1125  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1126  */
1127 
1128 static int
1129 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1130     const struct lacp_port *lp)
1131 {
1132 	if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1133 	    !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1134 		return (0);
1135 	}
1136 
1137 	if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1138 		return (0);
1139 	}
1140 
1141 	if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1142 		return (0);
1143 	}
1144 
1145 	if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1146 		return (0);
1147 	}
1148 
1149 	return (1);
1150 }
1151 
1152 static int
1153 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1154     const struct lacp_peerinfo *b)
1155 {
1156 	if (memcmp(&a->lip_systemid, &b->lip_systemid,
1157 	    sizeof(a->lip_systemid))) {
1158 		return (0);
1159 	}
1160 
1161 	if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1162 		return (0);
1163 	}
1164 
1165 	return (1);
1166 }
1167 
1168 static void
1169 lacp_port_enable(struct lacp_port *lp)
1170 {
1171 	lp->lp_state |= LACP_STATE_AGGREGATION;
1172 }
1173 
1174 static void
1175 lacp_port_disable(struct lacp_port *lp)
1176 {
1177 	lacp_set_mux(lp, LACP_MUX_DETACHED);
1178 
1179 	lp->lp_state &= ~LACP_STATE_AGGREGATION;
1180 	lp->lp_selected = LACP_UNSELECTED;
1181 	lacp_sm_rx_record_default(lp);
1182 	lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1183 	lp->lp_state &= ~LACP_STATE_EXPIRED;
1184 }
1185 
1186 /*
1187  * lacp_select: select an aggregator.  create one if necessary.
1188  */
1189 static void
1190 lacp_select(struct lacp_port *lp)
1191 {
1192 	struct lacp_softc *lsc = lp->lp_lsc;
1193 	struct lacp_aggregator *la;
1194 	char buf[LACP_LAGIDSTR_MAX+1];
1195 
1196 	if (lp->lp_aggregator) {
1197 		return;
1198 	}
1199 
1200 	KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1201 	    ("timer_wait_while still active"));
1202 
1203 	LACP_DPRINTF((lp, "port lagid=%s\n",
1204 	    lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1205 	    buf, sizeof(buf))));
1206 
1207 	TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1208 		if (lacp_aggregator_is_compatible(la, lp)) {
1209 			break;
1210 		}
1211 	}
1212 
1213 	if (la == NULL) {
1214 		la = lacp_aggregator_get(lsc, lp);
1215 		if (la == NULL) {
1216 			LACP_DPRINTF((lp, "aggregator creation failed\n"));
1217 
1218 			/*
1219 			 * will retry on the next tick.
1220 			 */
1221 
1222 			return;
1223 		}
1224 		lacp_fill_aggregator_id(la, lp);
1225 		LACP_DPRINTF((lp, "aggregator created\n"));
1226 	} else {
1227 		LACP_DPRINTF((lp, "compatible aggregator found\n"));
1228 		if (la->la_refcnt == LACP_MAX_PORTS)
1229 			return;
1230 		lacp_aggregator_addref(lsc, la);
1231 	}
1232 
1233 	LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1234 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1235 	    buf, sizeof(buf))));
1236 
1237 	lp->lp_aggregator = la;
1238 	lp->lp_selected = LACP_SELECTED;
1239 }
1240 
1241 /*
1242  * lacp_unselect: finish unselect/detach process.
1243  */
1244 
1245 static void
1246 lacp_unselect(struct lacp_port *lp)
1247 {
1248 	struct lacp_softc *lsc = lp->lp_lsc;
1249 	struct lacp_aggregator *la = lp->lp_aggregator;
1250 
1251 	KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1252 	    ("timer_wait_while still active"));
1253 
1254 	if (la == NULL) {
1255 		return;
1256 	}
1257 
1258 	lp->lp_aggregator = NULL;
1259 	lacp_aggregator_delref(lsc, la);
1260 }
1261 
1262 /* mux machine */
1263 
1264 static void
1265 lacp_sm_mux(struct lacp_port *lp)
1266 {
1267 	enum lacp_mux_state new_state;
1268 	boolean_t p_sync =
1269 		    (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1270 	boolean_t p_collecting =
1271 	    (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1272 	enum lacp_selected selected = lp->lp_selected;
1273 	struct lacp_aggregator *la;
1274 
1275 	if (lacp_debug > 1)
1276 		lacp_dprintf(lp, "%s: state %d\n", __func__, lp->lp_mux_state);
1277 
1278 re_eval:
1279 	la = lp->lp_aggregator;
1280 	KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1281 	    ("MUX not detached"));
1282 	new_state = lp->lp_mux_state;
1283 	switch (lp->lp_mux_state) {
1284 	case LACP_MUX_DETACHED:
1285 		if (selected != LACP_UNSELECTED) {
1286 			new_state = LACP_MUX_WAITING;
1287 		}
1288 		break;
1289 	case LACP_MUX_WAITING:
1290 		KASSERT(la->la_pending > 0 ||
1291 		    !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1292 		    ("timer_wait_while still active"));
1293 		if (selected == LACP_SELECTED && la->la_pending == 0) {
1294 			new_state = LACP_MUX_ATTACHED;
1295 		} else if (selected == LACP_UNSELECTED) {
1296 			new_state = LACP_MUX_DETACHED;
1297 		}
1298 		break;
1299 	case LACP_MUX_ATTACHED:
1300 		if (selected == LACP_SELECTED && p_sync) {
1301 			new_state = LACP_MUX_COLLECTING;
1302 		} else if (selected != LACP_SELECTED) {
1303 			new_state = LACP_MUX_DETACHED;
1304 		}
1305 		break;
1306 	case LACP_MUX_COLLECTING:
1307 		if (selected == LACP_SELECTED && p_sync && p_collecting) {
1308 			new_state = LACP_MUX_DISTRIBUTING;
1309 		} else if (selected != LACP_SELECTED || !p_sync) {
1310 			new_state = LACP_MUX_ATTACHED;
1311 		}
1312 		break;
1313 	case LACP_MUX_DISTRIBUTING:
1314 		if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1315 			new_state = LACP_MUX_COLLECTING;
1316 		}
1317 		break;
1318 	default:
1319 		panic("%s: unknown state", __func__);
1320 	}
1321 
1322 	if (lp->lp_mux_state == new_state) {
1323 		return;
1324 	}
1325 
1326 	lacp_set_mux(lp, new_state);
1327 	goto re_eval;
1328 }
1329 
1330 static void
1331 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1332 {
1333 	struct lacp_aggregator *la = lp->lp_aggregator;
1334 
1335 	if (lp->lp_mux_state == new_state) {
1336 		return;
1337 	}
1338 
1339 	switch (new_state) {
1340 	case LACP_MUX_DETACHED:
1341 		lp->lp_state &= ~LACP_STATE_SYNC;
1342 		lacp_disable_distributing(lp);
1343 		lacp_disable_collecting(lp);
1344 		lacp_sm_assert_ntt(lp);
1345 		/* cancel timer */
1346 		if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1347 			KASSERT(la->la_pending > 0,
1348 			    ("timer_wait_while not active"));
1349 			la->la_pending--;
1350 		}
1351 		LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1352 		lacp_unselect(lp);
1353 		break;
1354 	case LACP_MUX_WAITING:
1355 		LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1356 		    LACP_AGGREGATE_WAIT_TIME);
1357 		la->la_pending++;
1358 		break;
1359 	case LACP_MUX_ATTACHED:
1360 		lp->lp_state |= LACP_STATE_SYNC;
1361 		lacp_disable_collecting(lp);
1362 		lacp_sm_assert_ntt(lp);
1363 		break;
1364 	case LACP_MUX_COLLECTING:
1365 		lacp_enable_collecting(lp);
1366 		lacp_disable_distributing(lp);
1367 		lacp_sm_assert_ntt(lp);
1368 		break;
1369 	case LACP_MUX_DISTRIBUTING:
1370 		lacp_enable_distributing(lp);
1371 		break;
1372 	default:
1373 		panic("%s: unknown state", __func__);
1374 	}
1375 
1376 	LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1377 
1378 	lp->lp_mux_state = new_state;
1379 }
1380 
1381 static void
1382 lacp_sm_mux_timer(struct lacp_port *lp)
1383 {
1384 	struct lacp_aggregator *la = lp->lp_aggregator;
1385 	char buf[LACP_LAGIDSTR_MAX+1];
1386 
1387 	KASSERT(la->la_pending > 0, ("no pending event"));
1388 
1389 	LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1390 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1391 	    buf, sizeof(buf)),
1392 	    la->la_pending, la->la_pending - 1));
1393 
1394 	la->la_pending--;
1395 }
1396 
1397 /* periodic transmit machine */
1398 
1399 static void
1400 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1401 {
1402 	if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1403 	    LACP_STATE_TIMEOUT)) {
1404 		return;
1405 	}
1406 
1407 	LACP_DPRINTF((lp, "partner timeout changed\n"));
1408 
1409 	/*
1410 	 * FAST_PERIODIC -> SLOW_PERIODIC
1411 	 * or
1412 	 * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1413 	 *
1414 	 * let lacp_sm_ptx_tx_schedule to update timeout.
1415 	 */
1416 
1417 	LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1418 
1419 	/*
1420 	 * if timeout has been shortened, assert NTT.
1421 	 */
1422 
1423 	if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1424 		lacp_sm_assert_ntt(lp);
1425 	}
1426 }
1427 
1428 static void
1429 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1430 {
1431 	int timeout;
1432 
1433 	if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1434 	    !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1435 
1436 		/*
1437 		 * NO_PERIODIC
1438 		 */
1439 
1440 		LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1441 		return;
1442 	}
1443 
1444 	if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1445 		return;
1446 	}
1447 
1448 	timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1449 	    LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1450 
1451 	LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1452 }
1453 
1454 static void
1455 lacp_sm_ptx_timer(struct lacp_port *lp)
1456 {
1457 	lacp_sm_assert_ntt(lp);
1458 }
1459 
1460 static void
1461 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1462 {
1463 	int timeout;
1464 
1465 	/*
1466 	 * check LACP_DISABLED first
1467 	 */
1468 
1469 	if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1470 		return;
1471 	}
1472 
1473 	/*
1474 	 * check loopback condition.
1475 	 */
1476 
1477 	if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1478 	    &lp->lp_actor.lip_systemid)) {
1479 		return;
1480 	}
1481 
1482 	/*
1483 	 * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1484 	 */
1485 
1486 	lacp_sm_rx_update_selected(lp, du);
1487 	lacp_sm_rx_update_ntt(lp, du);
1488 	lacp_sm_rx_record_pdu(lp, du);
1489 
1490 	timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1491 	    LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1492 	LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1493 
1494 	lp->lp_state &= ~LACP_STATE_EXPIRED;
1495 
1496 	/*
1497 	 * kick transmit machine without waiting the next tick.
1498 	 */
1499 
1500 	lacp_sm_tx(lp);
1501 }
1502 
1503 static void
1504 lacp_sm_rx_set_expired(struct lacp_port *lp)
1505 {
1506 	lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1507 	lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1508 	LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1509 	lp->lp_state |= LACP_STATE_EXPIRED;
1510 }
1511 
1512 static void
1513 lacp_sm_rx_timer(struct lacp_port *lp)
1514 {
1515 	if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1516 		/* CURRENT -> EXPIRED */
1517 		LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1518 		lacp_sm_rx_set_expired(lp);
1519 	} else {
1520 		/* EXPIRED -> DEFAULTED */
1521 		LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1522 		lacp_sm_rx_update_default_selected(lp);
1523 		lacp_sm_rx_record_default(lp);
1524 		lp->lp_state &= ~LACP_STATE_EXPIRED;
1525 	}
1526 }
1527 
1528 static void
1529 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1530 {
1531 	boolean_t active;
1532 	uint8_t oldpstate;
1533 	char buf[LACP_STATESTR_MAX+1];
1534 
1535 	LACP_TRACE(lp);
1536 
1537 	oldpstate = lp->lp_partner.lip_state;
1538 
1539 	active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1540 	    || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1541 	    (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1542 
1543 	lp->lp_partner = du->ldu_actor;
1544 	if (active &&
1545 	    ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1546 	    LACP_STATE_AGGREGATION) &&
1547 	    !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1548 	    || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1549 		/* XXX nothing? */
1550 	} else {
1551 		lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1552 	}
1553 
1554 	lp->lp_state &= ~LACP_STATE_DEFAULTED;
1555 
1556 	if (oldpstate != lp->lp_partner.lip_state) {
1557 		LACP_DPRINTF((lp, "old pstate %s\n",
1558 		    lacp_format_state(oldpstate, buf, sizeof(buf))));
1559 		LACP_DPRINTF((lp, "new pstate %s\n",
1560 		    lacp_format_state(lp->lp_partner.lip_state, buf,
1561 		    sizeof(buf))));
1562 	}
1563 
1564 	lacp_sm_ptx_update_timeout(lp, oldpstate);
1565 }
1566 
1567 static void
1568 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1569 {
1570 
1571 	LACP_TRACE(lp);
1572 
1573 	if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1574 	    !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1575 	    LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1576 		LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1577 		lacp_sm_assert_ntt(lp);
1578 	}
1579 }
1580 
1581 static void
1582 lacp_sm_rx_record_default(struct lacp_port *lp)
1583 {
1584 	uint8_t oldpstate;
1585 
1586 	LACP_TRACE(lp);
1587 
1588 	oldpstate = lp->lp_partner.lip_state;
1589 	lp->lp_partner = lacp_partner_admin;
1590 	lp->lp_state |= LACP_STATE_DEFAULTED;
1591 	lacp_sm_ptx_update_timeout(lp, oldpstate);
1592 }
1593 
1594 static void
1595 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1596     const struct lacp_peerinfo *info)
1597 {
1598 
1599 	LACP_TRACE(lp);
1600 
1601 	if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1602 	    !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1603 	    LACP_STATE_AGGREGATION)) {
1604 		lp->lp_selected = LACP_UNSELECTED;
1605 		/* mux machine will clean up lp->lp_aggregator */
1606 	}
1607 }
1608 
1609 static void
1610 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1611 {
1612 
1613 	LACP_TRACE(lp);
1614 
1615 	lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1616 }
1617 
1618 static void
1619 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1620 {
1621 
1622 	LACP_TRACE(lp);
1623 
1624 	lacp_sm_rx_update_selected_from_peerinfo(lp, &lacp_partner_admin);
1625 }
1626 
1627 /* transmit machine */
1628 
1629 static void
1630 lacp_sm_tx(struct lacp_port *lp)
1631 {
1632 	int error;
1633 
1634 	if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1635 #if 1
1636 	    || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1637 	    && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1638 #endif
1639 	    ) {
1640 		lp->lp_flags &= ~LACP_PORT_NTT;
1641 	}
1642 
1643 	if (!(lp->lp_flags & LACP_PORT_NTT)) {
1644 		return;
1645 	}
1646 
1647 	/* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1648 	if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1649 		    (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1650 		LACP_DPRINTF((lp, "rate limited pdu\n"));
1651 		return;
1652 	}
1653 
1654 	error = lacp_xmit_lacpdu(lp);
1655 
1656 	if (error == 0) {
1657 		lp->lp_flags &= ~LACP_PORT_NTT;
1658 	} else {
1659 		LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1660 		    error));
1661 	}
1662 }
1663 
1664 static void
1665 lacp_sm_assert_ntt(struct lacp_port *lp)
1666 {
1667 
1668 	lp->lp_flags |= LACP_PORT_NTT;
1669 }
1670 
1671 static void
1672 lacp_run_timers(struct lacp_port *lp)
1673 {
1674 	int i;
1675 
1676 	for (i = 0; i < LACP_NTIMER; i++) {
1677 		KASSERT(lp->lp_timer[i] >= 0,
1678 		    ("invalid timer value %d", lp->lp_timer[i]));
1679 		if (lp->lp_timer[i] == 0) {
1680 			continue;
1681 		} else if (--lp->lp_timer[i] <= 0) {
1682 			if (lacp_timer_funcs[i]) {
1683 				(*lacp_timer_funcs[i])(lp);
1684 			}
1685 		}
1686 	}
1687 }
1688 
1689 int
1690 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1691 {
1692 	struct lacp_softc *lsc = lp->lp_lsc;
1693 	struct lagg_port *lgp = lp->lp_lagg;
1694 	struct lacp_port *lp2;
1695 	struct markerdu *mdu;
1696 	int error = 0;
1697 	int pending = 0;
1698 
1699 	if (m->m_pkthdr.len != sizeof(*mdu)) {
1700 		goto bad;
1701 	}
1702 
1703 	if ((m->m_flags & M_MCAST) == 0) {
1704 		goto bad;
1705 	}
1706 
1707 	if (m->m_len < sizeof(*mdu)) {
1708 		m = m_pullup(m, sizeof(*mdu));
1709 		if (m == NULL) {
1710 			return (ENOMEM);
1711 		}
1712 	}
1713 
1714 	mdu = mtod(m, struct markerdu *);
1715 
1716 	if (memcmp(&mdu->mdu_eh.ether_dhost,
1717 	    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1718 		goto bad;
1719 	}
1720 
1721 	if (mdu->mdu_sph.sph_version != 1) {
1722 		goto bad;
1723 	}
1724 
1725 	switch (mdu->mdu_tlv.tlv_type) {
1726 	case MARKER_TYPE_INFO:
1727 		if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1728 		    marker_info_tlv_template, TRUE)) {
1729 			goto bad;
1730 		}
1731 		mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1732 		memcpy(&mdu->mdu_eh.ether_dhost,
1733 		    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1734 		memcpy(&mdu->mdu_eh.ether_shost,
1735 		    lgp->lp_lladdr, ETHER_ADDR_LEN);
1736 		error = lagg_enqueue(lp->lp_ifp, m);
1737 		break;
1738 
1739 	case MARKER_TYPE_RESPONSE:
1740 		if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1741 		    marker_response_tlv_template, TRUE)) {
1742 			goto bad;
1743 		}
1744 		LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1745 		    ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1746 		    ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1747 
1748 		/* Verify that it is the last marker we sent out */
1749 		if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1750 		    sizeof(struct lacp_markerinfo)))
1751 			goto bad;
1752 
1753 		LACP_LOCK(lsc);
1754 		lp->lp_flags &= ~LACP_PORT_MARK;
1755 
1756 		if (lsc->lsc_suppress_distributing) {
1757 			/* Check if any ports are waiting for a response */
1758 			LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1759 				if (lp2->lp_flags & LACP_PORT_MARK) {
1760 					pending = 1;
1761 					break;
1762 				}
1763 			}
1764 
1765 			if (pending == 0) {
1766 				/* All interface queues are clear */
1767 				LACP_DPRINTF((NULL, "queue flush complete\n"));
1768 				lsc->lsc_suppress_distributing = FALSE;
1769 			}
1770 		}
1771 		LACP_UNLOCK(lsc);
1772 		m_freem(m);
1773 		break;
1774 
1775 	default:
1776 		goto bad;
1777 	}
1778 
1779 	return (error);
1780 
1781 bad:
1782 	LACP_DPRINTF((lp, "bad marker frame\n"));
1783 	m_freem(m);
1784 	return (EINVAL);
1785 }
1786 
1787 static int
1788 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1789     const struct tlv_template *tmpl, boolean_t check_type)
1790 {
1791 	while (/* CONSTCOND */ 1) {
1792 		if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1793 			return (EINVAL);
1794 		}
1795 		if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1796 		    tlv->tlv_length != tmpl->tmpl_length) {
1797 			return (EINVAL);
1798 		}
1799 		if (tmpl->tmpl_type == 0) {
1800 			break;
1801 		}
1802 		tlv = (const struct tlvhdr *)
1803 		    ((const char *)tlv + tlv->tlv_length);
1804 		tmpl++;
1805 	}
1806 
1807 	return (0);
1808 }
1809 
1810 /* Debugging */
1811 const char *
1812 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1813 {
1814 	snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1815 	    (int)mac[0],
1816 	    (int)mac[1],
1817 	    (int)mac[2],
1818 	    (int)mac[3],
1819 	    (int)mac[4],
1820 	    (int)mac[5]);
1821 
1822 	return (buf);
1823 }
1824 
1825 const char *
1826 lacp_format_systemid(const struct lacp_systemid *sysid,
1827     char *buf, size_t buflen)
1828 {
1829 	char macbuf[LACP_MACSTR_MAX+1];
1830 
1831 	snprintf(buf, buflen, "%04X,%s",
1832 	    ntohs(sysid->lsi_prio),
1833 	    lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1834 
1835 	return (buf);
1836 }
1837 
1838 const char *
1839 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
1840 {
1841 	snprintf(buf, buflen, "%04X,%04X",
1842 	    ntohs(portid->lpi_prio),
1843 	    ntohs(portid->lpi_portno));
1844 
1845 	return (buf);
1846 }
1847 
1848 const char *
1849 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
1850 {
1851 	char sysid[LACP_SYSTEMIDSTR_MAX+1];
1852 	char portid[LACP_PORTIDSTR_MAX+1];
1853 
1854 	snprintf(buf, buflen, "(%s,%04X,%s)",
1855 	    lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
1856 	    ntohs(peer->lip_key),
1857 	    lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
1858 
1859 	return (buf);
1860 }
1861 
1862 const char *
1863 lacp_format_lagid(const struct lacp_peerinfo *a,
1864     const struct lacp_peerinfo *b, char *buf, size_t buflen)
1865 {
1866 	char astr[LACP_PARTNERSTR_MAX+1];
1867 	char bstr[LACP_PARTNERSTR_MAX+1];
1868 
1869 #if 0
1870 	/*
1871 	 * there's a convention to display small numbered peer
1872 	 * in the left.
1873 	 */
1874 
1875 	if (lacp_compare_peerinfo(a, b) > 0) {
1876 		const struct lacp_peerinfo *t;
1877 
1878 		t = a;
1879 		a = b;
1880 		b = t;
1881 	}
1882 #endif
1883 
1884 	snprintf(buf, buflen, "[%s,%s]",
1885 	    lacp_format_partner(a, astr, sizeof(astr)),
1886 	    lacp_format_partner(b, bstr, sizeof(bstr)));
1887 
1888 	return (buf);
1889 }
1890 
1891 const char *
1892 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
1893     char *buf, size_t buflen)
1894 {
1895 	if (la == NULL) {
1896 		return ("(none)");
1897 	}
1898 
1899 	return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
1900 }
1901 
1902 const char *
1903 lacp_format_state(uint8_t state, char *buf, size_t buflen)
1904 {
1905 	snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
1906 	return (buf);
1907 }
1908 
1909 static void
1910 lacp_dump_lacpdu(const struct lacpdu *du)
1911 {
1912 	char buf[LACP_PARTNERSTR_MAX+1];
1913 	char buf2[LACP_STATESTR_MAX+1];
1914 
1915 	printf("actor=%s\n",
1916 	    lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
1917 	printf("actor.state=%s\n",
1918 	    lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
1919 	printf("partner=%s\n",
1920 	    lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
1921 	printf("partner.state=%s\n",
1922 	    lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
1923 
1924 	printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
1925 }
1926 
1927 static void
1928 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
1929 {
1930 	va_list va;
1931 
1932 	if (lp) {
1933 		printf("%s: ", lp->lp_ifp->if_xname);
1934 	}
1935 
1936 	va_start(va, fmt);
1937 	vprintf(fmt, va);
1938 	va_end(va);
1939 }
1940