xref: /freebsd/sys/net/ieee8023ad_lacp.c (revision 7431dfd4580e850375fe5478d92ec770344db098)
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/eventhandler.h>
36 #include <sys/mbuf.h>
37 #include <sys/systm.h>
38 #include <sys/malloc.h>
39 #include <sys/kernel.h> /* hz */
40 #include <sys/socket.h> /* for net/if.h */
41 #include <sys/sockio.h>
42 #include <sys/sysctl.h>
43 #include <machine/stdarg.h>
44 #include <sys/lock.h>
45 #include <sys/rwlock.h>
46 #include <sys/taskqueue.h>
47 
48 #include <net/if.h>
49 #include <net/if_var.h>
50 #include <net/if_dl.h>
51 #include <net/ethernet.h>
52 #include <net/if_media.h>
53 #include <net/if_types.h>
54 
55 #include <net/if_lagg.h>
56 #include <net/ieee8023ad_lacp.h>
57 
58 /*
59  * actor system priority and port priority.
60  * XXX should be configurable.
61  */
62 
63 #define	LACP_SYSTEM_PRIO	0x8000
64 #define	LACP_PORT_PRIO		0x8000
65 
66 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
67     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
68 
69 static const struct tlv_template lacp_info_tlv_template[] = {
70 	{ LACP_TYPE_ACTORINFO,
71 	    sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
72 	{ LACP_TYPE_PARTNERINFO,
73 	    sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
74 	{ LACP_TYPE_COLLECTORINFO,
75 	    sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
76 	{ 0, 0 },
77 };
78 
79 static const struct tlv_template marker_info_tlv_template[] = {
80 	{ MARKER_TYPE_INFO,
81 	    sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
82 	{ 0, 0 },
83 };
84 
85 static const struct tlv_template marker_response_tlv_template[] = {
86 	{ MARKER_TYPE_RESPONSE,
87 	    sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
88 	{ 0, 0 },
89 };
90 
91 typedef void (*lacp_timer_func_t)(struct lacp_port *);
92 
93 static void	lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
94 static void	lacp_fill_markerinfo(struct lacp_port *,
95 		    struct lacp_markerinfo *);
96 
97 static uint64_t	lacp_aggregator_bandwidth(struct lacp_aggregator *);
98 static void	lacp_suppress_distributing(struct lacp_softc *,
99 		    struct lacp_aggregator *);
100 static void	lacp_transit_expire(void *);
101 static void	lacp_update_portmap(struct lacp_softc *);
102 static void	lacp_select_active_aggregator(struct lacp_softc *);
103 static uint16_t	lacp_compose_key(struct lacp_port *);
104 static int	tlv_check(const void *, size_t, const struct tlvhdr *,
105 		    const struct tlv_template *, boolean_t);
106 static void	lacp_tick(void *);
107 
108 static void	lacp_fill_aggregator_id(struct lacp_aggregator *,
109 		    const struct lacp_port *);
110 static void	lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
111 		    const struct lacp_peerinfo *);
112 static int	lacp_aggregator_is_compatible(const struct lacp_aggregator *,
113 		    const struct lacp_port *);
114 static int	lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
115 		    const struct lacp_peerinfo *);
116 
117 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
118 		    struct lacp_port *);
119 static void	lacp_aggregator_addref(struct lacp_softc *,
120 		    struct lacp_aggregator *);
121 static void	lacp_aggregator_delref(struct lacp_softc *,
122 		    struct lacp_aggregator *);
123 
124 /* receive machine */
125 
126 static int	lacp_pdu_input(struct lacp_port *, struct mbuf *);
127 static int	lacp_marker_input(struct lacp_port *, struct mbuf *);
128 static void	lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
129 static void	lacp_sm_rx_timer(struct lacp_port *);
130 static void	lacp_sm_rx_set_expired(struct lacp_port *);
131 static void	lacp_sm_rx_update_ntt(struct lacp_port *,
132 		    const struct lacpdu *);
133 static void	lacp_sm_rx_record_pdu(struct lacp_port *,
134 		    const struct lacpdu *);
135 static void	lacp_sm_rx_update_selected(struct lacp_port *,
136 		    const struct lacpdu *);
137 static void	lacp_sm_rx_record_default(struct lacp_port *);
138 static void	lacp_sm_rx_update_default_selected(struct lacp_port *);
139 static void	lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
140 		    const struct lacp_peerinfo *);
141 
142 /* mux machine */
143 
144 static void	lacp_sm_mux(struct lacp_port *);
145 static void	lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
146 static void	lacp_sm_mux_timer(struct lacp_port *);
147 
148 /* periodic transmit machine */
149 
150 static void	lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
151 static void	lacp_sm_ptx_tx_schedule(struct lacp_port *);
152 static void	lacp_sm_ptx_timer(struct lacp_port *);
153 
154 /* transmit machine */
155 
156 static void	lacp_sm_tx(struct lacp_port *);
157 static void	lacp_sm_assert_ntt(struct lacp_port *);
158 
159 static void	lacp_run_timers(struct lacp_port *);
160 static int	lacp_compare_peerinfo(const struct lacp_peerinfo *,
161 		    const struct lacp_peerinfo *);
162 static int	lacp_compare_systemid(const struct lacp_systemid *,
163 		    const struct lacp_systemid *);
164 static void	lacp_port_enable(struct lacp_port *);
165 static void	lacp_port_disable(struct lacp_port *);
166 static void	lacp_select(struct lacp_port *);
167 static void	lacp_unselect(struct lacp_port *);
168 static void	lacp_disable_collecting(struct lacp_port *);
169 static void	lacp_enable_collecting(struct lacp_port *);
170 static void	lacp_disable_distributing(struct lacp_port *);
171 static void	lacp_enable_distributing(struct lacp_port *);
172 static int	lacp_xmit_lacpdu(struct lacp_port *);
173 static int	lacp_xmit_marker(struct lacp_port *);
174 
175 /* Debugging */
176 
177 static void	lacp_dump_lacpdu(const struct lacpdu *);
178 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
179 		    size_t);
180 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
181 		    const struct lacp_peerinfo *, char *, size_t);
182 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
183 		    char *, size_t);
184 static const char *lacp_format_state(uint8_t, char *, size_t);
185 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
186 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
187 		    size_t);
188 static const char *lacp_format_portid(const struct lacp_portid *, char *,
189 		    size_t);
190 static void	lacp_dprintf(const struct lacp_port *, const char *, ...)
191 		    __attribute__((__format__(__printf__, 2, 3)));
192 
193 static VNET_DEFINE(int, lacp_debug);
194 #define	V_lacp_debug	VNET(lacp_debug)
195 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
196 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
197     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
198 
199 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
200 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
201 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
202 
203 /*
204  * partner administration variables.
205  * XXX should be configurable.
206  */
207 
208 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
209 	.lip_systemid = { .lsi_prio = 0xffff },
210 	.lip_portid = { .lpi_prio = 0xffff },
211 	.lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
212 	    LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
213 };
214 
215 static const struct lacp_peerinfo lacp_partner_admin_strict = {
216 	.lip_systemid = { .lsi_prio = 0xffff },
217 	.lip_portid = { .lpi_prio = 0xffff },
218 	.lip_state = 0,
219 };
220 
221 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
222 	[LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
223 	[LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
224 	[LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
225 };
226 
227 struct mbuf *
228 lacp_input(struct lagg_port *lgp, struct mbuf *m)
229 {
230 	struct lacp_port *lp = LACP_PORT(lgp);
231 	uint8_t subtype;
232 
233 	if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
234 		m_freem(m);
235 		return (NULL);
236 	}
237 
238 	m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
239 	switch (subtype) {
240 		case SLOWPROTOCOLS_SUBTYPE_LACP:
241 			lacp_pdu_input(lp, m);
242 			return (NULL);
243 
244 		case SLOWPROTOCOLS_SUBTYPE_MARKER:
245 			lacp_marker_input(lp, m);
246 			return (NULL);
247 	}
248 
249 	/* Not a subtype we are interested in */
250 	return (m);
251 }
252 
253 /*
254  * lacp_pdu_input: process lacpdu
255  */
256 static int
257 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
258 {
259 	struct lacp_softc *lsc = lp->lp_lsc;
260 	struct lacpdu *du;
261 	int error = 0;
262 
263 	if (m->m_pkthdr.len != sizeof(*du)) {
264 		goto bad;
265 	}
266 
267 	if ((m->m_flags & M_MCAST) == 0) {
268 		goto bad;
269 	}
270 
271 	if (m->m_len < sizeof(*du)) {
272 		m = m_pullup(m, sizeof(*du));
273 		if (m == NULL) {
274 			return (ENOMEM);
275 		}
276 	}
277 
278 	du = mtod(m, struct lacpdu *);
279 
280 	if (memcmp(&du->ldu_eh.ether_dhost,
281 	    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
282 		goto bad;
283 	}
284 
285 	/*
286 	 * ignore the version for compatibility with
287 	 * the future protocol revisions.
288 	 */
289 #if 0
290 	if (du->ldu_sph.sph_version != 1) {
291 		goto bad;
292 	}
293 #endif
294 
295 	/*
296 	 * ignore tlv types for compatibility with
297 	 * the future protocol revisions.
298 	 */
299 	if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
300 	    lacp_info_tlv_template, FALSE)) {
301 		goto bad;
302 	}
303 
304         if (V_lacp_debug > 0) {
305 		lacp_dprintf(lp, "lacpdu receive\n");
306 		lacp_dump_lacpdu(du);
307 	}
308 
309 	if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
310 		LACP_TPRINTF((lp, "Dropping RX PDU\n"));
311 		goto bad;
312 	}
313 
314 	LACP_LOCK(lsc);
315 	lacp_sm_rx(lp, du);
316 	LACP_UNLOCK(lsc);
317 
318 	m_freem(m);
319 	return (error);
320 
321 bad:
322 	m_freem(m);
323 	return (EINVAL);
324 }
325 
326 static void
327 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
328 {
329 	struct lagg_port *lgp = lp->lp_lagg;
330 	struct lagg_softc *sc = lgp->lp_softc;
331 
332 	info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
333 	memcpy(&info->lip_systemid.lsi_mac,
334 	    IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
335 	info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
336 	info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
337 	info->lip_state = lp->lp_state;
338 }
339 
340 static void
341 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
342 {
343 	struct ifnet *ifp = lp->lp_ifp;
344 
345 	/* Fill in the port index and system id (encoded as the MAC) */
346 	info->mi_rq_port = htons(ifp->if_index);
347 	memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
348 	info->mi_rq_xid = htonl(0);
349 }
350 
351 static int
352 lacp_xmit_lacpdu(struct lacp_port *lp)
353 {
354 	struct lagg_port *lgp = lp->lp_lagg;
355 	struct mbuf *m;
356 	struct lacpdu *du;
357 	int error;
358 
359 	LACP_LOCK_ASSERT(lp->lp_lsc);
360 
361 	m = m_gethdr(M_NOWAIT, MT_DATA);
362 	if (m == NULL) {
363 		return (ENOMEM);
364 	}
365 	m->m_len = m->m_pkthdr.len = sizeof(*du);
366 
367 	du = mtod(m, struct lacpdu *);
368 	memset(du, 0, sizeof(*du));
369 
370 	memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
371 	    ETHER_ADDR_LEN);
372 	memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
373 	du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
374 
375 	du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
376 	du->ldu_sph.sph_version = 1;
377 
378 	TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
379 	du->ldu_actor = lp->lp_actor;
380 
381 	TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
382 	    sizeof(du->ldu_partner));
383 	du->ldu_partner = lp->lp_partner;
384 
385 	TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
386 	    sizeof(du->ldu_collector));
387 	du->ldu_collector.lci_maxdelay = 0;
388 
389 	if (V_lacp_debug > 0) {
390 		lacp_dprintf(lp, "lacpdu transmit\n");
391 		lacp_dump_lacpdu(du);
392 	}
393 
394 	m->m_flags |= M_MCAST;
395 
396 	/*
397 	 * XXX should use higher priority queue.
398 	 * otherwise network congestion can break aggregation.
399 	 */
400 
401 	error = lagg_enqueue(lp->lp_ifp, m);
402 	return (error);
403 }
404 
405 static int
406 lacp_xmit_marker(struct lacp_port *lp)
407 {
408 	struct lagg_port *lgp = lp->lp_lagg;
409 	struct mbuf *m;
410 	struct markerdu *mdu;
411 	int error;
412 
413 	LACP_LOCK_ASSERT(lp->lp_lsc);
414 
415 	m = m_gethdr(M_NOWAIT, MT_DATA);
416 	if (m == NULL) {
417 		return (ENOMEM);
418 	}
419 	m->m_len = m->m_pkthdr.len = sizeof(*mdu);
420 
421 	mdu = mtod(m, struct markerdu *);
422 	memset(mdu, 0, sizeof(*mdu));
423 
424 	memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
425 	    ETHER_ADDR_LEN);
426 	memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
427 	mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
428 
429 	mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
430 	mdu->mdu_sph.sph_version = 1;
431 
432 	/* Bump the transaction id and copy over the marker info */
433 	lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
434 	TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
435 	mdu->mdu_info = lp->lp_marker;
436 
437 	LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
438 	    ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
439 	    ntohl(mdu->mdu_info.mi_rq_xid)));
440 
441 	m->m_flags |= M_MCAST;
442 	error = lagg_enqueue(lp->lp_ifp, m);
443 	return (error);
444 }
445 
446 void
447 lacp_linkstate(struct lagg_port *lgp)
448 {
449 	struct lacp_port *lp = LACP_PORT(lgp);
450 	struct lacp_softc *lsc = lp->lp_lsc;
451 	struct ifnet *ifp = lgp->lp_ifp;
452 	struct ifmediareq ifmr;
453 	int error = 0;
454 	u_int media;
455 	uint8_t old_state;
456 	uint16_t old_key;
457 
458 	bzero((char *)&ifmr, sizeof(ifmr));
459 	error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
460 	if (error != 0)
461 		return;
462 
463 	LACP_LOCK(lsc);
464 	media = ifmr.ifm_active;
465 	LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
466 	    "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
467 	    (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
468 	old_state = lp->lp_state;
469 	old_key = lp->lp_key;
470 
471 	lp->lp_media = media;
472 	/*
473 	 * If the port is not an active full duplex Ethernet link then it can
474 	 * not be aggregated.
475 	 */
476 	if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
477 	    ifp->if_link_state != LINK_STATE_UP) {
478 		lacp_port_disable(lp);
479 	} else {
480 		lacp_port_enable(lp);
481 	}
482 	lp->lp_key = lacp_compose_key(lp);
483 
484 	if (old_state != lp->lp_state || old_key != lp->lp_key) {
485 		LACP_DPRINTF((lp, "-> UNSELECTED\n"));
486 		lp->lp_selected = LACP_UNSELECTED;
487 	}
488 	LACP_UNLOCK(lsc);
489 }
490 
491 static void
492 lacp_tick(void *arg)
493 {
494 	struct lacp_softc *lsc = arg;
495 	struct lacp_port *lp;
496 
497 	LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
498 		if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
499 			continue;
500 
501 		CURVNET_SET(lp->lp_ifp->if_vnet);
502 		lacp_run_timers(lp);
503 
504 		lacp_select(lp);
505 		lacp_sm_mux(lp);
506 		lacp_sm_tx(lp);
507 		lacp_sm_ptx_tx_schedule(lp);
508 		CURVNET_RESTORE();
509 	}
510 	callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
511 }
512 
513 int
514 lacp_port_create(struct lagg_port *lgp)
515 {
516 	struct lagg_softc *sc = lgp->lp_softc;
517 	struct lacp_softc *lsc = LACP_SOFTC(sc);
518 	struct lacp_port *lp;
519 	struct ifnet *ifp = lgp->lp_ifp;
520 	struct sockaddr_dl sdl;
521 	struct ifmultiaddr *rifma = NULL;
522 	int error;
523 
524 	boolean_t active = TRUE; /* XXX should be configurable */
525 	boolean_t fast = FALSE; /* XXX should be configurable */
526 
527 	link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
528 	sdl.sdl_alen = ETHER_ADDR_LEN;
529 
530 	bcopy(&ethermulticastaddr_slowprotocols,
531 	    LLADDR(&sdl), ETHER_ADDR_LEN);
532 	error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
533 	if (error) {
534 		printf("%s: ADDMULTI failed on %s\n", __func__,
535 		    lgp->lp_ifp->if_xname);
536 		return (error);
537 	}
538 
539 	lp = malloc(sizeof(struct lacp_port),
540 	    M_DEVBUF, M_NOWAIT|M_ZERO);
541 	if (lp == NULL)
542 		return (ENOMEM);
543 
544 	LACP_LOCK(lsc);
545 	lgp->lp_psc = lp;
546 	lp->lp_ifp = ifp;
547 	lp->lp_lagg = lgp;
548 	lp->lp_lsc = lsc;
549 	lp->lp_ifma = rifma;
550 
551 	LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
552 
553 	lacp_fill_actorinfo(lp, &lp->lp_actor);
554 	lacp_fill_markerinfo(lp, &lp->lp_marker);
555 	lp->lp_state =
556 	    (active ? LACP_STATE_ACTIVITY : 0) |
557 	    (fast ? LACP_STATE_TIMEOUT : 0);
558 	lp->lp_aggregator = NULL;
559 	lacp_sm_rx_set_expired(lp);
560 	LACP_UNLOCK(lsc);
561 	lacp_linkstate(lgp);
562 
563 	return (0);
564 }
565 
566 void
567 lacp_port_destroy(struct lagg_port *lgp)
568 {
569 	struct lacp_port *lp = LACP_PORT(lgp);
570 	struct lacp_softc *lsc = lp->lp_lsc;
571 	int i;
572 
573 	LACP_LOCK(lsc);
574 	for (i = 0; i < LACP_NTIMER; i++) {
575 		LACP_TIMER_DISARM(lp, i);
576 	}
577 
578 	lacp_disable_collecting(lp);
579 	lacp_disable_distributing(lp);
580 	lacp_unselect(lp);
581 
582 	LIST_REMOVE(lp, lp_next);
583 	LACP_UNLOCK(lsc);
584 
585 	/* The address may have already been removed by if_purgemaddrs() */
586 	if (!lgp->lp_detaching)
587 		if_delmulti_ifma(lp->lp_ifma);
588 
589 	free(lp, M_DEVBUF);
590 }
591 
592 void
593 lacp_req(struct lagg_softc *sc, void *data)
594 {
595 	struct lacp_opreq *req = (struct lacp_opreq *)data;
596 	struct lacp_softc *lsc = LACP_SOFTC(sc);
597 	struct lacp_aggregator *la;
598 
599 	bzero(req, sizeof(struct lacp_opreq));
600 
601 	/*
602 	 * If the LACP softc is NULL, return with the opreq structure full of
603 	 * zeros.  It is normal for the softc to be NULL while the lagg is
604 	 * being destroyed.
605 	 */
606 	if (NULL == lsc)
607 		return;
608 
609 	la = lsc->lsc_active_aggregator;
610 	LACP_LOCK(lsc);
611 	if (la != NULL) {
612 		req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
613 		memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
614 		    ETHER_ADDR_LEN);
615 		req->actor_key = ntohs(la->la_actor.lip_key);
616 		req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
617 		req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
618 		req->actor_state = la->la_actor.lip_state;
619 
620 		req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
621 		memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
622 		    ETHER_ADDR_LEN);
623 		req->partner_key = ntohs(la->la_partner.lip_key);
624 		req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
625 		req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
626 		req->partner_state = la->la_partner.lip_state;
627 	}
628 	LACP_UNLOCK(lsc);
629 }
630 
631 void
632 lacp_portreq(struct lagg_port *lgp, void *data)
633 {
634 	struct lacp_opreq *req = (struct lacp_opreq *)data;
635 	struct lacp_port *lp = LACP_PORT(lgp);
636 	struct lacp_softc *lsc = lp->lp_lsc;
637 
638 	LACP_LOCK(lsc);
639 	req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
640 	memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
641 	    ETHER_ADDR_LEN);
642 	req->actor_key = ntohs(lp->lp_actor.lip_key);
643 	req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
644 	req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
645 	req->actor_state = lp->lp_actor.lip_state;
646 
647 	req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
648 	memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
649 	    ETHER_ADDR_LEN);
650 	req->partner_key = ntohs(lp->lp_partner.lip_key);
651 	req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
652 	req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
653 	req->partner_state = lp->lp_partner.lip_state;
654 	LACP_UNLOCK(lsc);
655 }
656 
657 static void
658 lacp_disable_collecting(struct lacp_port *lp)
659 {
660 	LACP_DPRINTF((lp, "collecting disabled\n"));
661 	lp->lp_state &= ~LACP_STATE_COLLECTING;
662 }
663 
664 static void
665 lacp_enable_collecting(struct lacp_port *lp)
666 {
667 	LACP_DPRINTF((lp, "collecting enabled\n"));
668 	lp->lp_state |= LACP_STATE_COLLECTING;
669 }
670 
671 static void
672 lacp_disable_distributing(struct lacp_port *lp)
673 {
674 	struct lacp_aggregator *la = lp->lp_aggregator;
675 	struct lacp_softc *lsc = lp->lp_lsc;
676 	struct lagg_softc *sc = lsc->lsc_softc;
677 	char buf[LACP_LAGIDSTR_MAX+1];
678 
679 	LACP_LOCK_ASSERT(lsc);
680 
681 	if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
682 		return;
683 	}
684 
685 	KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
686 	KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
687 	KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
688 
689 	LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
690 	    "nports %d -> %d\n",
691 	    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
692 	    la->la_nports, la->la_nports - 1));
693 
694 	TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
695 	la->la_nports--;
696 	sc->sc_active = la->la_nports;
697 
698 	if (lsc->lsc_active_aggregator == la) {
699 		lacp_suppress_distributing(lsc, la);
700 		lacp_select_active_aggregator(lsc);
701 		/* regenerate the port map, the active aggregator has changed */
702 		lacp_update_portmap(lsc);
703 	}
704 
705 	lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
706 }
707 
708 static void
709 lacp_enable_distributing(struct lacp_port *lp)
710 {
711 	struct lacp_aggregator *la = lp->lp_aggregator;
712 	struct lacp_softc *lsc = lp->lp_lsc;
713 	struct lagg_softc *sc = lsc->lsc_softc;
714 	char buf[LACP_LAGIDSTR_MAX+1];
715 
716 	LACP_LOCK_ASSERT(lsc);
717 
718 	if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
719 		return;
720 	}
721 
722 	LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
723 	    "nports %d -> %d\n",
724 	    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
725 	    la->la_nports, la->la_nports + 1));
726 
727 	KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
728 	TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
729 	la->la_nports++;
730 	sc->sc_active = la->la_nports;
731 
732 	lp->lp_state |= LACP_STATE_DISTRIBUTING;
733 
734 	if (lsc->lsc_active_aggregator == la) {
735 		lacp_suppress_distributing(lsc, la);
736 		lacp_update_portmap(lsc);
737 	} else
738 		/* try to become the active aggregator */
739 		lacp_select_active_aggregator(lsc);
740 }
741 
742 static void
743 lacp_transit_expire(void *vp)
744 {
745 	struct lacp_softc *lsc = vp;
746 
747 	LACP_LOCK_ASSERT(lsc);
748 
749 	CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
750 	LACP_TRACE(NULL);
751 	CURVNET_RESTORE();
752 
753 	lsc->lsc_suppress_distributing = FALSE;
754 }
755 
756 void
757 lacp_attach(struct lagg_softc *sc)
758 {
759 	struct lacp_softc *lsc;
760 
761 	lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
762 
763 	sc->sc_psc = lsc;
764 	lsc->lsc_softc = sc;
765 
766 	lsc->lsc_hashkey = arc4random();
767 	lsc->lsc_active_aggregator = NULL;
768 	lsc->lsc_strict_mode = 1;
769 	LACP_LOCK_INIT(lsc);
770 	TAILQ_INIT(&lsc->lsc_aggregators);
771 	LIST_INIT(&lsc->lsc_ports);
772 
773 	callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
774 	callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
775 
776 	/* if the lagg is already up then do the same */
777 	if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
778 		lacp_init(sc);
779 }
780 
781 void
782 lacp_detach(void *psc)
783 {
784 	struct lacp_softc *lsc = (struct lacp_softc *)psc;
785 
786 	KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
787 	    ("aggregators still active"));
788 	KASSERT(lsc->lsc_active_aggregator == NULL,
789 	    ("aggregator still attached"));
790 
791 	callout_drain(&lsc->lsc_transit_callout);
792 	callout_drain(&lsc->lsc_callout);
793 
794 	LACP_LOCK_DESTROY(lsc);
795 	free(lsc, M_DEVBUF);
796 }
797 
798 void
799 lacp_init(struct lagg_softc *sc)
800 {
801 	struct lacp_softc *lsc = LACP_SOFTC(sc);
802 
803 	LACP_LOCK(lsc);
804 	callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
805 	LACP_UNLOCK(lsc);
806 }
807 
808 void
809 lacp_stop(struct lagg_softc *sc)
810 {
811 	struct lacp_softc *lsc = LACP_SOFTC(sc);
812 
813 	LACP_LOCK(lsc);
814 	callout_stop(&lsc->lsc_transit_callout);
815 	callout_stop(&lsc->lsc_callout);
816 	LACP_UNLOCK(lsc);
817 }
818 
819 struct lagg_port *
820 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
821 {
822 	struct lacp_softc *lsc = LACP_SOFTC(sc);
823 	struct lacp_portmap *pm;
824 	struct lacp_port *lp;
825 	uint32_t hash;
826 
827 	if (__predict_false(lsc->lsc_suppress_distributing)) {
828 		LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
829 		return (NULL);
830 	}
831 
832 	pm = &lsc->lsc_pmap[lsc->lsc_activemap];
833 	if (pm->pm_count == 0) {
834 		LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
835 		return (NULL);
836 	}
837 
838 	if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) && (m->m_flags & M_FLOWID))
839 		hash = m->m_pkthdr.flowid >> sc->flowid_shift;
840 	else
841 		hash = lagg_hashmbuf(sc, m, lsc->lsc_hashkey);
842 	hash %= pm->pm_count;
843 	lp = pm->pm_map[hash];
844 
845 	KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
846 	    ("aggregated port is not distributing"));
847 
848 	return (lp->lp_lagg);
849 }
850 /*
851  * lacp_suppress_distributing: drop transmit packets for a while
852  * to preserve packet ordering.
853  */
854 
855 static void
856 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
857 {
858 	struct lacp_port *lp;
859 
860 	if (lsc->lsc_active_aggregator != la) {
861 		return;
862 	}
863 
864 	LACP_TRACE(NULL);
865 
866 	lsc->lsc_suppress_distributing = TRUE;
867 
868 	/* send a marker frame down each port to verify the queues are empty */
869 	LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
870 		lp->lp_flags |= LACP_PORT_MARK;
871 		lacp_xmit_marker(lp);
872 	}
873 
874 	/* set a timeout for the marker frames */
875 	callout_reset(&lsc->lsc_transit_callout,
876 	    LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
877 }
878 
879 static int
880 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
881     const struct lacp_peerinfo *b)
882 {
883 	return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
884 }
885 
886 static int
887 lacp_compare_systemid(const struct lacp_systemid *a,
888     const struct lacp_systemid *b)
889 {
890 	return (memcmp(a, b, sizeof(*a)));
891 }
892 
893 #if 0	/* unused */
894 static int
895 lacp_compare_portid(const struct lacp_portid *a,
896     const struct lacp_portid *b)
897 {
898 	return (memcmp(a, b, sizeof(*a)));
899 }
900 #endif
901 
902 static uint64_t
903 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
904 {
905 	struct lacp_port *lp;
906 	uint64_t speed;
907 
908 	lp = TAILQ_FIRST(&la->la_ports);
909 	if (lp == NULL) {
910 		return (0);
911 	}
912 
913 	speed = ifmedia_baudrate(lp->lp_media);
914 	speed *= la->la_nports;
915 	if (speed == 0) {
916 		LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
917 		    lp->lp_media, la->la_nports));
918 	}
919 
920 	return (speed);
921 }
922 
923 /*
924  * lacp_select_active_aggregator: select an aggregator to be used to transmit
925  * packets from lagg(4) interface.
926  */
927 
928 static void
929 lacp_select_active_aggregator(struct lacp_softc *lsc)
930 {
931 	struct lacp_aggregator *la;
932 	struct lacp_aggregator *best_la = NULL;
933 	uint64_t best_speed = 0;
934 	char buf[LACP_LAGIDSTR_MAX+1];
935 
936 	LACP_TRACE(NULL);
937 
938 	TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
939 		uint64_t speed;
940 
941 		if (la->la_nports == 0) {
942 			continue;
943 		}
944 
945 		speed = lacp_aggregator_bandwidth(la);
946 		LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
947 		    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
948 		    speed, la->la_nports));
949 
950 		/*
951 		 * This aggregator is chosen if the partner has a better
952 		 * system priority or, the total aggregated speed is higher
953 		 * or, it is already the chosen aggregator
954 		 */
955 		if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
956 		    LACP_SYS_PRI(best_la->la_partner)) ||
957 		    speed > best_speed ||
958 		    (speed == best_speed &&
959 		    la == lsc->lsc_active_aggregator)) {
960 			best_la = la;
961 			best_speed = speed;
962 		}
963 	}
964 
965 	KASSERT(best_la == NULL || best_la->la_nports > 0,
966 	    ("invalid aggregator refcnt"));
967 	KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
968 	    ("invalid aggregator list"));
969 
970 	if (lsc->lsc_active_aggregator != best_la) {
971 		LACP_DPRINTF((NULL, "active aggregator changed\n"));
972 		LACP_DPRINTF((NULL, "old %s\n",
973 		    lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
974 		    buf, sizeof(buf))));
975 	} else {
976 		LACP_DPRINTF((NULL, "active aggregator not changed\n"));
977 	}
978 	LACP_DPRINTF((NULL, "new %s\n",
979 	    lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
980 
981 	if (lsc->lsc_active_aggregator != best_la) {
982 		lsc->lsc_active_aggregator = best_la;
983 		lacp_update_portmap(lsc);
984 		if (best_la) {
985 			lacp_suppress_distributing(lsc, best_la);
986 		}
987 	}
988 }
989 
990 /*
991  * Updated the inactive portmap array with the new list of ports and
992  * make it live.
993  */
994 static void
995 lacp_update_portmap(struct lacp_softc *lsc)
996 {
997 	struct lagg_softc *sc = lsc->lsc_softc;
998 	struct lacp_aggregator *la;
999 	struct lacp_portmap *p;
1000 	struct lacp_port *lp;
1001 	uint64_t speed;
1002 	u_int newmap;
1003 	int i;
1004 
1005 	newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1006 	p = &lsc->lsc_pmap[newmap];
1007 	la = lsc->lsc_active_aggregator;
1008 	speed = 0;
1009 	bzero(p, sizeof(struct lacp_portmap));
1010 
1011 	if (la != NULL && la->la_nports > 0) {
1012 		p->pm_count = la->la_nports;
1013 		i = 0;
1014 		TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1015 			p->pm_map[i++] = lp;
1016 		KASSERT(i == p->pm_count, ("Invalid port count"));
1017 		speed = lacp_aggregator_bandwidth(la);
1018 	}
1019 	sc->sc_ifp->if_baudrate = speed;
1020 
1021 	/* switch the active portmap over */
1022 	atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1023 	LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1024 		    lsc->lsc_activemap,
1025 		    lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1026 }
1027 
1028 static uint16_t
1029 lacp_compose_key(struct lacp_port *lp)
1030 {
1031 	struct lagg_port *lgp = lp->lp_lagg;
1032 	struct lagg_softc *sc = lgp->lp_softc;
1033 	u_int media = lp->lp_media;
1034 	uint16_t key;
1035 
1036 	if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1037 
1038 		/*
1039 		 * non-aggregatable links should have unique keys.
1040 		 *
1041 		 * XXX this isn't really unique as if_index is 16 bit.
1042 		 */
1043 
1044 		/* bit 0..14:	(some bits of) if_index of this port */
1045 		key = lp->lp_ifp->if_index;
1046 		/* bit 15:	1 */
1047 		key |= 0x8000;
1048 	} else {
1049 		u_int subtype = IFM_SUBTYPE(media);
1050 
1051 		KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1052 		KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1053 
1054 		/* bit 0..4:	IFM_SUBTYPE modulo speed */
1055 		switch (subtype) {
1056 		case IFM_10_T:
1057 		case IFM_10_2:
1058 		case IFM_10_5:
1059 		case IFM_10_STP:
1060 		case IFM_10_FL:
1061 			key = IFM_10_T;
1062 			break;
1063 		case IFM_100_TX:
1064 		case IFM_100_FX:
1065 		case IFM_100_T4:
1066 		case IFM_100_VG:
1067 		case IFM_100_T2:
1068 			key = IFM_100_TX;
1069 			break;
1070 		case IFM_1000_SX:
1071 		case IFM_1000_LX:
1072 		case IFM_1000_CX:
1073 		case IFM_1000_T:
1074 			key = IFM_1000_SX;
1075 			break;
1076 		case IFM_10G_LR:
1077 		case IFM_10G_SR:
1078 		case IFM_10G_CX4:
1079 		case IFM_10G_TWINAX:
1080 		case IFM_10G_TWINAX_LONG:
1081 		case IFM_10G_LRM:
1082 		case IFM_10G_T:
1083 			key = IFM_10G_LR;
1084 			break;
1085 		case IFM_40G_CR4:
1086 		case IFM_40G_SR4:
1087 		case IFM_40G_LR4:
1088 			key = IFM_40G_CR4;
1089 			break;
1090 		default:
1091 			key = subtype;
1092 		}
1093 		/* bit 5..14:	(some bits of) if_index of lagg device */
1094 		key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1095 		/* bit 15:	0 */
1096 	}
1097 	return (htons(key));
1098 }
1099 
1100 static void
1101 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1102 {
1103 	char buf[LACP_LAGIDSTR_MAX+1];
1104 
1105 	LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1106 	    __func__,
1107 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1108 	    buf, sizeof(buf)),
1109 	    la->la_refcnt, la->la_refcnt + 1));
1110 
1111 	KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1112 	la->la_refcnt++;
1113 	KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1114 }
1115 
1116 static void
1117 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1118 {
1119 	char buf[LACP_LAGIDSTR_MAX+1];
1120 
1121 	LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1122 	    __func__,
1123 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1124 	    buf, sizeof(buf)),
1125 	    la->la_refcnt, la->la_refcnt - 1));
1126 
1127 	KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1128 	la->la_refcnt--;
1129 	if (la->la_refcnt > 0) {
1130 		return;
1131 	}
1132 
1133 	KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1134 	KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1135 
1136 	TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1137 
1138 	free(la, M_DEVBUF);
1139 }
1140 
1141 /*
1142  * lacp_aggregator_get: allocate an aggregator.
1143  */
1144 
1145 static struct lacp_aggregator *
1146 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1147 {
1148 	struct lacp_aggregator *la;
1149 
1150 	la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1151 	if (la) {
1152 		la->la_refcnt = 1;
1153 		la->la_nports = 0;
1154 		TAILQ_INIT(&la->la_ports);
1155 		la->la_pending = 0;
1156 		TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1157 	}
1158 
1159 	return (la);
1160 }
1161 
1162 /*
1163  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1164  */
1165 
1166 static void
1167 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1168 {
1169 	lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1170 	lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1171 
1172 	la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1173 }
1174 
1175 static void
1176 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1177     const struct lacp_peerinfo *lpi_port)
1178 {
1179 	memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1180 	lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1181 	lpi_aggr->lip_key = lpi_port->lip_key;
1182 }
1183 
1184 /*
1185  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1186  */
1187 
1188 static int
1189 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1190     const struct lacp_port *lp)
1191 {
1192 	if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1193 	    !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1194 		return (0);
1195 	}
1196 
1197 	if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1198 		return (0);
1199 	}
1200 
1201 	if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1202 		return (0);
1203 	}
1204 
1205 	if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1206 		return (0);
1207 	}
1208 
1209 	return (1);
1210 }
1211 
1212 static int
1213 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1214     const struct lacp_peerinfo *b)
1215 {
1216 	if (memcmp(&a->lip_systemid, &b->lip_systemid,
1217 	    sizeof(a->lip_systemid))) {
1218 		return (0);
1219 	}
1220 
1221 	if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1222 		return (0);
1223 	}
1224 
1225 	return (1);
1226 }
1227 
1228 static void
1229 lacp_port_enable(struct lacp_port *lp)
1230 {
1231 	lp->lp_state |= LACP_STATE_AGGREGATION;
1232 }
1233 
1234 static void
1235 lacp_port_disable(struct lacp_port *lp)
1236 {
1237 	lacp_set_mux(lp, LACP_MUX_DETACHED);
1238 
1239 	lp->lp_state &= ~LACP_STATE_AGGREGATION;
1240 	lp->lp_selected = LACP_UNSELECTED;
1241 	lacp_sm_rx_record_default(lp);
1242 	lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1243 	lp->lp_state &= ~LACP_STATE_EXPIRED;
1244 }
1245 
1246 /*
1247  * lacp_select: select an aggregator.  create one if necessary.
1248  */
1249 static void
1250 lacp_select(struct lacp_port *lp)
1251 {
1252 	struct lacp_softc *lsc = lp->lp_lsc;
1253 	struct lacp_aggregator *la;
1254 	char buf[LACP_LAGIDSTR_MAX+1];
1255 
1256 	if (lp->lp_aggregator) {
1257 		return;
1258 	}
1259 
1260 	KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1261 	    ("timer_wait_while still active"));
1262 
1263 	LACP_DPRINTF((lp, "port lagid=%s\n",
1264 	    lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1265 	    buf, sizeof(buf))));
1266 
1267 	TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1268 		if (lacp_aggregator_is_compatible(la, lp)) {
1269 			break;
1270 		}
1271 	}
1272 
1273 	if (la == NULL) {
1274 		la = lacp_aggregator_get(lsc, lp);
1275 		if (la == NULL) {
1276 			LACP_DPRINTF((lp, "aggregator creation failed\n"));
1277 
1278 			/*
1279 			 * will retry on the next tick.
1280 			 */
1281 
1282 			return;
1283 		}
1284 		lacp_fill_aggregator_id(la, lp);
1285 		LACP_DPRINTF((lp, "aggregator created\n"));
1286 	} else {
1287 		LACP_DPRINTF((lp, "compatible aggregator found\n"));
1288 		if (la->la_refcnt == LACP_MAX_PORTS)
1289 			return;
1290 		lacp_aggregator_addref(lsc, la);
1291 	}
1292 
1293 	LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1294 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1295 	    buf, sizeof(buf))));
1296 
1297 	lp->lp_aggregator = la;
1298 	lp->lp_selected = LACP_SELECTED;
1299 }
1300 
1301 /*
1302  * lacp_unselect: finish unselect/detach process.
1303  */
1304 
1305 static void
1306 lacp_unselect(struct lacp_port *lp)
1307 {
1308 	struct lacp_softc *lsc = lp->lp_lsc;
1309 	struct lacp_aggregator *la = lp->lp_aggregator;
1310 
1311 	KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1312 	    ("timer_wait_while still active"));
1313 
1314 	if (la == NULL) {
1315 		return;
1316 	}
1317 
1318 	lp->lp_aggregator = NULL;
1319 	lacp_aggregator_delref(lsc, la);
1320 }
1321 
1322 /* mux machine */
1323 
1324 static void
1325 lacp_sm_mux(struct lacp_port *lp)
1326 {
1327 	struct lagg_port *lgp = lp->lp_lagg;
1328 	struct lagg_softc *sc = lgp->lp_softc;
1329 	enum lacp_mux_state new_state;
1330 	boolean_t p_sync =
1331 		    (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1332 	boolean_t p_collecting =
1333 	    (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1334 	enum lacp_selected selected = lp->lp_selected;
1335 	struct lacp_aggregator *la;
1336 
1337 	if (V_lacp_debug > 1)
1338 		lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1339 		    "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1340 		    lp->lp_mux_state, selected, p_sync, p_collecting);
1341 
1342 re_eval:
1343 	la = lp->lp_aggregator;
1344 	KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1345 	    ("MUX not detached"));
1346 	new_state = lp->lp_mux_state;
1347 	switch (lp->lp_mux_state) {
1348 	case LACP_MUX_DETACHED:
1349 		if (selected != LACP_UNSELECTED) {
1350 			new_state = LACP_MUX_WAITING;
1351 		}
1352 		break;
1353 	case LACP_MUX_WAITING:
1354 		KASSERT(la->la_pending > 0 ||
1355 		    !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1356 		    ("timer_wait_while still active"));
1357 		if (selected == LACP_SELECTED && la->la_pending == 0) {
1358 			new_state = LACP_MUX_ATTACHED;
1359 		} else if (selected == LACP_UNSELECTED) {
1360 			new_state = LACP_MUX_DETACHED;
1361 		}
1362 		break;
1363 	case LACP_MUX_ATTACHED:
1364 		if (selected == LACP_SELECTED && p_sync) {
1365 			new_state = LACP_MUX_COLLECTING;
1366 		} else if (selected != LACP_SELECTED) {
1367 			new_state = LACP_MUX_DETACHED;
1368 		}
1369 		break;
1370 	case LACP_MUX_COLLECTING:
1371 		if (selected == LACP_SELECTED && p_sync && p_collecting) {
1372 			new_state = LACP_MUX_DISTRIBUTING;
1373 		} else if (selected != LACP_SELECTED || !p_sync) {
1374 			new_state = LACP_MUX_ATTACHED;
1375 		}
1376 		break;
1377 	case LACP_MUX_DISTRIBUTING:
1378 		if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1379 			new_state = LACP_MUX_COLLECTING;
1380 			lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1381 			sc->sc_flapping++;
1382 		}
1383 		break;
1384 	default:
1385 		panic("%s: unknown state", __func__);
1386 	}
1387 
1388 	if (lp->lp_mux_state == new_state) {
1389 		return;
1390 	}
1391 
1392 	lacp_set_mux(lp, new_state);
1393 	goto re_eval;
1394 }
1395 
1396 static void
1397 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1398 {
1399 	struct lacp_aggregator *la = lp->lp_aggregator;
1400 
1401 	if (lp->lp_mux_state == new_state) {
1402 		return;
1403 	}
1404 
1405 	switch (new_state) {
1406 	case LACP_MUX_DETACHED:
1407 		lp->lp_state &= ~LACP_STATE_SYNC;
1408 		lacp_disable_distributing(lp);
1409 		lacp_disable_collecting(lp);
1410 		lacp_sm_assert_ntt(lp);
1411 		/* cancel timer */
1412 		if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1413 			KASSERT(la->la_pending > 0,
1414 			    ("timer_wait_while not active"));
1415 			la->la_pending--;
1416 		}
1417 		LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1418 		lacp_unselect(lp);
1419 		break;
1420 	case LACP_MUX_WAITING:
1421 		LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1422 		    LACP_AGGREGATE_WAIT_TIME);
1423 		la->la_pending++;
1424 		break;
1425 	case LACP_MUX_ATTACHED:
1426 		lp->lp_state |= LACP_STATE_SYNC;
1427 		lacp_disable_collecting(lp);
1428 		lacp_sm_assert_ntt(lp);
1429 		break;
1430 	case LACP_MUX_COLLECTING:
1431 		lacp_enable_collecting(lp);
1432 		lacp_disable_distributing(lp);
1433 		lacp_sm_assert_ntt(lp);
1434 		break;
1435 	case LACP_MUX_DISTRIBUTING:
1436 		lacp_enable_distributing(lp);
1437 		break;
1438 	default:
1439 		panic("%s: unknown state", __func__);
1440 	}
1441 
1442 	LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1443 
1444 	lp->lp_mux_state = new_state;
1445 }
1446 
1447 static void
1448 lacp_sm_mux_timer(struct lacp_port *lp)
1449 {
1450 	struct lacp_aggregator *la = lp->lp_aggregator;
1451 	char buf[LACP_LAGIDSTR_MAX+1];
1452 
1453 	KASSERT(la->la_pending > 0, ("no pending event"));
1454 
1455 	LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1456 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1457 	    buf, sizeof(buf)),
1458 	    la->la_pending, la->la_pending - 1));
1459 
1460 	la->la_pending--;
1461 }
1462 
1463 /* periodic transmit machine */
1464 
1465 static void
1466 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1467 {
1468 	if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1469 	    LACP_STATE_TIMEOUT)) {
1470 		return;
1471 	}
1472 
1473 	LACP_DPRINTF((lp, "partner timeout changed\n"));
1474 
1475 	/*
1476 	 * FAST_PERIODIC -> SLOW_PERIODIC
1477 	 * or
1478 	 * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1479 	 *
1480 	 * let lacp_sm_ptx_tx_schedule to update timeout.
1481 	 */
1482 
1483 	LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1484 
1485 	/*
1486 	 * if timeout has been shortened, assert NTT.
1487 	 */
1488 
1489 	if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1490 		lacp_sm_assert_ntt(lp);
1491 	}
1492 }
1493 
1494 static void
1495 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1496 {
1497 	int timeout;
1498 
1499 	if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1500 	    !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1501 
1502 		/*
1503 		 * NO_PERIODIC
1504 		 */
1505 
1506 		LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1507 		return;
1508 	}
1509 
1510 	if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1511 		return;
1512 	}
1513 
1514 	timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1515 	    LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1516 
1517 	LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1518 }
1519 
1520 static void
1521 lacp_sm_ptx_timer(struct lacp_port *lp)
1522 {
1523 	lacp_sm_assert_ntt(lp);
1524 }
1525 
1526 static void
1527 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1528 {
1529 	int timeout;
1530 
1531 	/*
1532 	 * check LACP_DISABLED first
1533 	 */
1534 
1535 	if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1536 		return;
1537 	}
1538 
1539 	/*
1540 	 * check loopback condition.
1541 	 */
1542 
1543 	if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1544 	    &lp->lp_actor.lip_systemid)) {
1545 		return;
1546 	}
1547 
1548 	/*
1549 	 * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1550 	 */
1551 
1552 	lacp_sm_rx_update_selected(lp, du);
1553 	lacp_sm_rx_update_ntt(lp, du);
1554 	lacp_sm_rx_record_pdu(lp, du);
1555 
1556 	timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1557 	    LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1558 	LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1559 
1560 	lp->lp_state &= ~LACP_STATE_EXPIRED;
1561 
1562 	/*
1563 	 * kick transmit machine without waiting the next tick.
1564 	 */
1565 
1566 	lacp_sm_tx(lp);
1567 }
1568 
1569 static void
1570 lacp_sm_rx_set_expired(struct lacp_port *lp)
1571 {
1572 	lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1573 	lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1574 	LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1575 	lp->lp_state |= LACP_STATE_EXPIRED;
1576 }
1577 
1578 static void
1579 lacp_sm_rx_timer(struct lacp_port *lp)
1580 {
1581 	if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1582 		/* CURRENT -> EXPIRED */
1583 		LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1584 		lacp_sm_rx_set_expired(lp);
1585 	} else {
1586 		/* EXPIRED -> DEFAULTED */
1587 		LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1588 		lacp_sm_rx_update_default_selected(lp);
1589 		lacp_sm_rx_record_default(lp);
1590 		lp->lp_state &= ~LACP_STATE_EXPIRED;
1591 	}
1592 }
1593 
1594 static void
1595 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1596 {
1597 	boolean_t active;
1598 	uint8_t oldpstate;
1599 	char buf[LACP_STATESTR_MAX+1];
1600 
1601 	LACP_TRACE(lp);
1602 
1603 	oldpstate = lp->lp_partner.lip_state;
1604 
1605 	active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1606 	    || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1607 	    (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1608 
1609 	lp->lp_partner = du->ldu_actor;
1610 	if (active &&
1611 	    ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1612 	    LACP_STATE_AGGREGATION) &&
1613 	    !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1614 	    || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1615 		/* XXX nothing? */
1616 	} else {
1617 		lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1618 	}
1619 
1620 	lp->lp_state &= ~LACP_STATE_DEFAULTED;
1621 
1622 	if (oldpstate != lp->lp_partner.lip_state) {
1623 		LACP_DPRINTF((lp, "old pstate %s\n",
1624 		    lacp_format_state(oldpstate, buf, sizeof(buf))));
1625 		LACP_DPRINTF((lp, "new pstate %s\n",
1626 		    lacp_format_state(lp->lp_partner.lip_state, buf,
1627 		    sizeof(buf))));
1628 	}
1629 
1630 	/* XXX Hack, still need to implement 5.4.9 para 2,3,4 */
1631 	if (lp->lp_lsc->lsc_strict_mode)
1632 		lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1633 
1634 	lacp_sm_ptx_update_timeout(lp, oldpstate);
1635 }
1636 
1637 static void
1638 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1639 {
1640 
1641 	LACP_TRACE(lp);
1642 
1643 	if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1644 	    !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1645 	    LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1646 		LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1647 		lacp_sm_assert_ntt(lp);
1648 	}
1649 }
1650 
1651 static void
1652 lacp_sm_rx_record_default(struct lacp_port *lp)
1653 {
1654 	uint8_t oldpstate;
1655 
1656 	LACP_TRACE(lp);
1657 
1658 	oldpstate = lp->lp_partner.lip_state;
1659 	if (lp->lp_lsc->lsc_strict_mode)
1660 		lp->lp_partner = lacp_partner_admin_strict;
1661 	else
1662 		lp->lp_partner = lacp_partner_admin_optimistic;;
1663 	lp->lp_state |= LACP_STATE_DEFAULTED;
1664 	lacp_sm_ptx_update_timeout(lp, oldpstate);
1665 }
1666 
1667 static void
1668 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1669     const struct lacp_peerinfo *info)
1670 {
1671 
1672 	LACP_TRACE(lp);
1673 
1674 	if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1675 	    !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1676 	    LACP_STATE_AGGREGATION)) {
1677 		lp->lp_selected = LACP_UNSELECTED;
1678 		/* mux machine will clean up lp->lp_aggregator */
1679 	}
1680 }
1681 
1682 static void
1683 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1684 {
1685 
1686 	LACP_TRACE(lp);
1687 
1688 	lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1689 }
1690 
1691 static void
1692 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1693 {
1694 
1695 	LACP_TRACE(lp);
1696 
1697 	if (lp->lp_lsc->lsc_strict_mode)
1698 		lacp_sm_rx_update_selected_from_peerinfo(lp,
1699 		    &lacp_partner_admin_strict);
1700 	else
1701 		lacp_sm_rx_update_selected_from_peerinfo(lp,
1702 		    &lacp_partner_admin_optimistic);
1703 }
1704 
1705 /* transmit machine */
1706 
1707 static void
1708 lacp_sm_tx(struct lacp_port *lp)
1709 {
1710 	int error = 0;
1711 
1712 	if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1713 #if 1
1714 	    || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1715 	    && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1716 #endif
1717 	    ) {
1718 		lp->lp_flags &= ~LACP_PORT_NTT;
1719 	}
1720 
1721 	if (!(lp->lp_flags & LACP_PORT_NTT)) {
1722 		return;
1723 	}
1724 
1725 	/* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1726 	if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1727 		    (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1728 		LACP_DPRINTF((lp, "rate limited pdu\n"));
1729 		return;
1730 	}
1731 
1732 	if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1733 		error = lacp_xmit_lacpdu(lp);
1734 	} else {
1735 		LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1736 	}
1737 
1738 	if (error == 0) {
1739 		lp->lp_flags &= ~LACP_PORT_NTT;
1740 	} else {
1741 		LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1742 		    error));
1743 	}
1744 }
1745 
1746 static void
1747 lacp_sm_assert_ntt(struct lacp_port *lp)
1748 {
1749 
1750 	lp->lp_flags |= LACP_PORT_NTT;
1751 }
1752 
1753 static void
1754 lacp_run_timers(struct lacp_port *lp)
1755 {
1756 	int i;
1757 
1758 	for (i = 0; i < LACP_NTIMER; i++) {
1759 		KASSERT(lp->lp_timer[i] >= 0,
1760 		    ("invalid timer value %d", lp->lp_timer[i]));
1761 		if (lp->lp_timer[i] == 0) {
1762 			continue;
1763 		} else if (--lp->lp_timer[i] <= 0) {
1764 			if (lacp_timer_funcs[i]) {
1765 				(*lacp_timer_funcs[i])(lp);
1766 			}
1767 		}
1768 	}
1769 }
1770 
1771 int
1772 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1773 {
1774 	struct lacp_softc *lsc = lp->lp_lsc;
1775 	struct lagg_port *lgp = lp->lp_lagg;
1776 	struct lacp_port *lp2;
1777 	struct markerdu *mdu;
1778 	int error = 0;
1779 	int pending = 0;
1780 
1781 	if (m->m_pkthdr.len != sizeof(*mdu)) {
1782 		goto bad;
1783 	}
1784 
1785 	if ((m->m_flags & M_MCAST) == 0) {
1786 		goto bad;
1787 	}
1788 
1789 	if (m->m_len < sizeof(*mdu)) {
1790 		m = m_pullup(m, sizeof(*mdu));
1791 		if (m == NULL) {
1792 			return (ENOMEM);
1793 		}
1794 	}
1795 
1796 	mdu = mtod(m, struct markerdu *);
1797 
1798 	if (memcmp(&mdu->mdu_eh.ether_dhost,
1799 	    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1800 		goto bad;
1801 	}
1802 
1803 	if (mdu->mdu_sph.sph_version != 1) {
1804 		goto bad;
1805 	}
1806 
1807 	switch (mdu->mdu_tlv.tlv_type) {
1808 	case MARKER_TYPE_INFO:
1809 		if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1810 		    marker_info_tlv_template, TRUE)) {
1811 			goto bad;
1812 		}
1813 		mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1814 		memcpy(&mdu->mdu_eh.ether_dhost,
1815 		    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1816 		memcpy(&mdu->mdu_eh.ether_shost,
1817 		    lgp->lp_lladdr, ETHER_ADDR_LEN);
1818 		error = lagg_enqueue(lp->lp_ifp, m);
1819 		break;
1820 
1821 	case MARKER_TYPE_RESPONSE:
1822 		if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1823 		    marker_response_tlv_template, TRUE)) {
1824 			goto bad;
1825 		}
1826 		LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1827 		    ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1828 		    ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1829 
1830 		/* Verify that it is the last marker we sent out */
1831 		if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1832 		    sizeof(struct lacp_markerinfo)))
1833 			goto bad;
1834 
1835 		LACP_LOCK(lsc);
1836 		lp->lp_flags &= ~LACP_PORT_MARK;
1837 
1838 		if (lsc->lsc_suppress_distributing) {
1839 			/* Check if any ports are waiting for a response */
1840 			LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1841 				if (lp2->lp_flags & LACP_PORT_MARK) {
1842 					pending = 1;
1843 					break;
1844 				}
1845 			}
1846 
1847 			if (pending == 0) {
1848 				/* All interface queues are clear */
1849 				LACP_DPRINTF((NULL, "queue flush complete\n"));
1850 				lsc->lsc_suppress_distributing = FALSE;
1851 			}
1852 		}
1853 		LACP_UNLOCK(lsc);
1854 		m_freem(m);
1855 		break;
1856 
1857 	default:
1858 		goto bad;
1859 	}
1860 
1861 	return (error);
1862 
1863 bad:
1864 	LACP_DPRINTF((lp, "bad marker frame\n"));
1865 	m_freem(m);
1866 	return (EINVAL);
1867 }
1868 
1869 static int
1870 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1871     const struct tlv_template *tmpl, boolean_t check_type)
1872 {
1873 	while (/* CONSTCOND */ 1) {
1874 		if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1875 			return (EINVAL);
1876 		}
1877 		if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1878 		    tlv->tlv_length != tmpl->tmpl_length) {
1879 			return (EINVAL);
1880 		}
1881 		if (tmpl->tmpl_type == 0) {
1882 			break;
1883 		}
1884 		tlv = (const struct tlvhdr *)
1885 		    ((const char *)tlv + tlv->tlv_length);
1886 		tmpl++;
1887 	}
1888 
1889 	return (0);
1890 }
1891 
1892 /* Debugging */
1893 const char *
1894 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1895 {
1896 	snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1897 	    (int)mac[0],
1898 	    (int)mac[1],
1899 	    (int)mac[2],
1900 	    (int)mac[3],
1901 	    (int)mac[4],
1902 	    (int)mac[5]);
1903 
1904 	return (buf);
1905 }
1906 
1907 const char *
1908 lacp_format_systemid(const struct lacp_systemid *sysid,
1909     char *buf, size_t buflen)
1910 {
1911 	char macbuf[LACP_MACSTR_MAX+1];
1912 
1913 	snprintf(buf, buflen, "%04X,%s",
1914 	    ntohs(sysid->lsi_prio),
1915 	    lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1916 
1917 	return (buf);
1918 }
1919 
1920 const char *
1921 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
1922 {
1923 	snprintf(buf, buflen, "%04X,%04X",
1924 	    ntohs(portid->lpi_prio),
1925 	    ntohs(portid->lpi_portno));
1926 
1927 	return (buf);
1928 }
1929 
1930 const char *
1931 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
1932 {
1933 	char sysid[LACP_SYSTEMIDSTR_MAX+1];
1934 	char portid[LACP_PORTIDSTR_MAX+1];
1935 
1936 	snprintf(buf, buflen, "(%s,%04X,%s)",
1937 	    lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
1938 	    ntohs(peer->lip_key),
1939 	    lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
1940 
1941 	return (buf);
1942 }
1943 
1944 const char *
1945 lacp_format_lagid(const struct lacp_peerinfo *a,
1946     const struct lacp_peerinfo *b, char *buf, size_t buflen)
1947 {
1948 	char astr[LACP_PARTNERSTR_MAX+1];
1949 	char bstr[LACP_PARTNERSTR_MAX+1];
1950 
1951 #if 0
1952 	/*
1953 	 * there's a convention to display small numbered peer
1954 	 * in the left.
1955 	 */
1956 
1957 	if (lacp_compare_peerinfo(a, b) > 0) {
1958 		const struct lacp_peerinfo *t;
1959 
1960 		t = a;
1961 		a = b;
1962 		b = t;
1963 	}
1964 #endif
1965 
1966 	snprintf(buf, buflen, "[%s,%s]",
1967 	    lacp_format_partner(a, astr, sizeof(astr)),
1968 	    lacp_format_partner(b, bstr, sizeof(bstr)));
1969 
1970 	return (buf);
1971 }
1972 
1973 const char *
1974 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
1975     char *buf, size_t buflen)
1976 {
1977 	if (la == NULL) {
1978 		return ("(none)");
1979 	}
1980 
1981 	return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
1982 }
1983 
1984 const char *
1985 lacp_format_state(uint8_t state, char *buf, size_t buflen)
1986 {
1987 	snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
1988 	return (buf);
1989 }
1990 
1991 static void
1992 lacp_dump_lacpdu(const struct lacpdu *du)
1993 {
1994 	char buf[LACP_PARTNERSTR_MAX+1];
1995 	char buf2[LACP_STATESTR_MAX+1];
1996 
1997 	printf("actor=%s\n",
1998 	    lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
1999 	printf("actor.state=%s\n",
2000 	    lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2001 	printf("partner=%s\n",
2002 	    lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2003 	printf("partner.state=%s\n",
2004 	    lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2005 
2006 	printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2007 }
2008 
2009 static void
2010 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2011 {
2012 	va_list va;
2013 
2014 	if (lp) {
2015 		printf("%s: ", lp->lp_ifp->if_xname);
2016 	}
2017 
2018 	va_start(va, fmt);
2019 	vprintf(fmt, va);
2020 	va_end(va);
2021 }
2022