xref: /freebsd/sys/net/ieee8023ad_lacp.c (revision 0d972b25f64dc1f52aff3fe09bc62cbaf332df83)
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/fnv_hash.h>
39 #include <sys/malloc.h>
40 #include <sys/kernel.h> /* hz */
41 #include <sys/socket.h> /* for net/if.h */
42 #include <sys/sockio.h>
43 #include <sys/sysctl.h>
44 #include <machine/stdarg.h>
45 #include <sys/lock.h>
46 #include <sys/rwlock.h>
47 #include <sys/taskqueue.h>
48 
49 #include <net/if.h>
50 #include <net/if_var.h>
51 #include <net/if_dl.h>
52 #include <net/ethernet.h>
53 #include <net/if_media.h>
54 #include <net/if_types.h>
55 
56 #include <net/if_lagg.h>
57 #include <net/ieee8023ad_lacp.h>
58 
59 /*
60  * actor system priority and port priority.
61  * XXX should be configurable.
62  */
63 
64 #define	LACP_SYSTEM_PRIO	0x8000
65 #define	LACP_PORT_PRIO		0x8000
66 
67 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
68     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
69 
70 static const struct tlv_template lacp_info_tlv_template[] = {
71 	{ LACP_TYPE_ACTORINFO,
72 	    sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
73 	{ LACP_TYPE_PARTNERINFO,
74 	    sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
75 	{ LACP_TYPE_COLLECTORINFO,
76 	    sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
77 	{ 0, 0 },
78 };
79 
80 static const struct tlv_template marker_info_tlv_template[] = {
81 	{ MARKER_TYPE_INFO,
82 	    sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
83 	{ 0, 0 },
84 };
85 
86 static const struct tlv_template marker_response_tlv_template[] = {
87 	{ MARKER_TYPE_RESPONSE,
88 	    sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
89 	{ 0, 0 },
90 };
91 
92 typedef void (*lacp_timer_func_t)(struct lacp_port *);
93 
94 static void	lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
95 static void	lacp_fill_markerinfo(struct lacp_port *,
96 		    struct lacp_markerinfo *);
97 
98 static uint64_t	lacp_aggregator_bandwidth(struct lacp_aggregator *);
99 static void	lacp_suppress_distributing(struct lacp_softc *,
100 		    struct lacp_aggregator *);
101 static void	lacp_transit_expire(void *);
102 static void	lacp_update_portmap(struct lacp_softc *);
103 static void	lacp_select_active_aggregator(struct lacp_softc *);
104 static uint16_t	lacp_compose_key(struct lacp_port *);
105 static int	tlv_check(const void *, size_t, const struct tlvhdr *,
106 		    const struct tlv_template *, boolean_t);
107 static void	lacp_tick(void *);
108 
109 static void	lacp_fill_aggregator_id(struct lacp_aggregator *,
110 		    const struct lacp_port *);
111 static void	lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
112 		    const struct lacp_peerinfo *);
113 static int	lacp_aggregator_is_compatible(const struct lacp_aggregator *,
114 		    const struct lacp_port *);
115 static int	lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
116 		    const struct lacp_peerinfo *);
117 
118 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
119 		    struct lacp_port *);
120 static void	lacp_aggregator_addref(struct lacp_softc *,
121 		    struct lacp_aggregator *);
122 static void	lacp_aggregator_delref(struct lacp_softc *,
123 		    struct lacp_aggregator *);
124 
125 /* receive machine */
126 
127 static int	lacp_pdu_input(struct lacp_port *, struct mbuf *);
128 static int	lacp_marker_input(struct lacp_port *, struct mbuf *);
129 static void	lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
130 static void	lacp_sm_rx_timer(struct lacp_port *);
131 static void	lacp_sm_rx_set_expired(struct lacp_port *);
132 static void	lacp_sm_rx_update_ntt(struct lacp_port *,
133 		    const struct lacpdu *);
134 static void	lacp_sm_rx_record_pdu(struct lacp_port *,
135 		    const struct lacpdu *);
136 static void	lacp_sm_rx_update_selected(struct lacp_port *,
137 		    const struct lacpdu *);
138 static void	lacp_sm_rx_record_default(struct lacp_port *);
139 static void	lacp_sm_rx_update_default_selected(struct lacp_port *);
140 static void	lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
141 		    const struct lacp_peerinfo *);
142 
143 /* mux machine */
144 
145 static void	lacp_sm_mux(struct lacp_port *);
146 static void	lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
147 static void	lacp_sm_mux_timer(struct lacp_port *);
148 
149 /* periodic transmit machine */
150 
151 static void	lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
152 static void	lacp_sm_ptx_tx_schedule(struct lacp_port *);
153 static void	lacp_sm_ptx_timer(struct lacp_port *);
154 
155 /* transmit machine */
156 
157 static void	lacp_sm_tx(struct lacp_port *);
158 static void	lacp_sm_assert_ntt(struct lacp_port *);
159 
160 static void	lacp_run_timers(struct lacp_port *);
161 static int	lacp_compare_peerinfo(const struct lacp_peerinfo *,
162 		    const struct lacp_peerinfo *);
163 static int	lacp_compare_systemid(const struct lacp_systemid *,
164 		    const struct lacp_systemid *);
165 static void	lacp_port_enable(struct lacp_port *);
166 static void	lacp_port_disable(struct lacp_port *);
167 static void	lacp_select(struct lacp_port *);
168 static void	lacp_unselect(struct lacp_port *);
169 static void	lacp_disable_collecting(struct lacp_port *);
170 static void	lacp_enable_collecting(struct lacp_port *);
171 static void	lacp_disable_distributing(struct lacp_port *);
172 static void	lacp_enable_distributing(struct lacp_port *);
173 static int	lacp_xmit_lacpdu(struct lacp_port *);
174 static int	lacp_xmit_marker(struct lacp_port *);
175 
176 /* Debugging */
177 
178 static void	lacp_dump_lacpdu(const struct lacpdu *);
179 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
180 		    size_t);
181 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
182 		    const struct lacp_peerinfo *, char *, size_t);
183 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
184 		    char *, size_t);
185 static const char *lacp_format_state(uint8_t, char *, size_t);
186 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
187 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
188 		    size_t);
189 static const char *lacp_format_portid(const struct lacp_portid *, char *,
190 		    size_t);
191 static void	lacp_dprintf(const struct lacp_port *, const char *, ...)
192 		    __attribute__((__format__(__printf__, 2, 3)));
193 
194 static VNET_DEFINE(int, lacp_debug);
195 #define	V_lacp_debug	VNET(lacp_debug)
196 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
197 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
198     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
199 
200 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
201 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
202 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
203 
204 /*
205  * partner administration variables.
206  * XXX should be configurable.
207  */
208 
209 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
210 	.lip_systemid = { .lsi_prio = 0xffff },
211 	.lip_portid = { .lpi_prio = 0xffff },
212 	.lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
213 	    LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
214 };
215 
216 static const struct lacp_peerinfo lacp_partner_admin_strict = {
217 	.lip_systemid = { .lsi_prio = 0xffff },
218 	.lip_portid = { .lpi_prio = 0xffff },
219 	.lip_state = 0,
220 };
221 
222 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
223 	[LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
224 	[LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
225 	[LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
226 };
227 
228 struct mbuf *
229 lacp_input(struct lagg_port *lgp, struct mbuf *m)
230 {
231 	struct lacp_port *lp = LACP_PORT(lgp);
232 	uint8_t subtype;
233 
234 	if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
235 		m_freem(m);
236 		return (NULL);
237 	}
238 
239 	m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
240 	switch (subtype) {
241 		case SLOWPROTOCOLS_SUBTYPE_LACP:
242 			lacp_pdu_input(lp, m);
243 			return (NULL);
244 
245 		case SLOWPROTOCOLS_SUBTYPE_MARKER:
246 			lacp_marker_input(lp, m);
247 			return (NULL);
248 	}
249 
250 	/* Not a subtype we are interested in */
251 	return (m);
252 }
253 
254 /*
255  * lacp_pdu_input: process lacpdu
256  */
257 static int
258 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
259 {
260 	struct lacp_softc *lsc = lp->lp_lsc;
261 	struct lacpdu *du;
262 	int error = 0;
263 
264 	if (m->m_pkthdr.len != sizeof(*du)) {
265 		goto bad;
266 	}
267 
268 	if ((m->m_flags & M_MCAST) == 0) {
269 		goto bad;
270 	}
271 
272 	if (m->m_len < sizeof(*du)) {
273 		m = m_pullup(m, sizeof(*du));
274 		if (m == NULL) {
275 			return (ENOMEM);
276 		}
277 	}
278 
279 	du = mtod(m, struct lacpdu *);
280 
281 	if (memcmp(&du->ldu_eh.ether_dhost,
282 	    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
283 		goto bad;
284 	}
285 
286 	/*
287 	 * ignore the version for compatibility with
288 	 * the future protocol revisions.
289 	 */
290 #if 0
291 	if (du->ldu_sph.sph_version != 1) {
292 		goto bad;
293 	}
294 #endif
295 
296 	/*
297 	 * ignore tlv types for compatibility with
298 	 * the future protocol revisions.
299 	 */
300 	if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
301 	    lacp_info_tlv_template, FALSE)) {
302 		goto bad;
303 	}
304 
305         if (V_lacp_debug > 0) {
306 		lacp_dprintf(lp, "lacpdu receive\n");
307 		lacp_dump_lacpdu(du);
308 	}
309 
310 	if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
311 		LACP_TPRINTF((lp, "Dropping RX PDU\n"));
312 		goto bad;
313 	}
314 
315 	LACP_LOCK(lsc);
316 	lacp_sm_rx(lp, du);
317 	LACP_UNLOCK(lsc);
318 
319 	m_freem(m);
320 	return (error);
321 
322 bad:
323 	m_freem(m);
324 	return (EINVAL);
325 }
326 
327 static void
328 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
329 {
330 	struct lagg_port *lgp = lp->lp_lagg;
331 	struct lagg_softc *sc = lgp->lp_softc;
332 
333 	info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
334 	memcpy(&info->lip_systemid.lsi_mac,
335 	    IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
336 	info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
337 	info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
338 	info->lip_state = lp->lp_state;
339 }
340 
341 static void
342 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
343 {
344 	struct ifnet *ifp = lp->lp_ifp;
345 
346 	/* Fill in the port index and system id (encoded as the MAC) */
347 	info->mi_rq_port = htons(ifp->if_index);
348 	memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
349 	info->mi_rq_xid = htonl(0);
350 }
351 
352 static int
353 lacp_xmit_lacpdu(struct lacp_port *lp)
354 {
355 	struct lagg_port *lgp = lp->lp_lagg;
356 	struct mbuf *m;
357 	struct lacpdu *du;
358 	int error;
359 
360 	LACP_LOCK_ASSERT(lp->lp_lsc);
361 
362 	m = m_gethdr(M_NOWAIT, MT_DATA);
363 	if (m == NULL) {
364 		return (ENOMEM);
365 	}
366 	m->m_len = m->m_pkthdr.len = sizeof(*du);
367 
368 	du = mtod(m, struct lacpdu *);
369 	memset(du, 0, sizeof(*du));
370 
371 	memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
372 	    ETHER_ADDR_LEN);
373 	memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
374 	du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
375 
376 	du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
377 	du->ldu_sph.sph_version = 1;
378 
379 	TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
380 	du->ldu_actor = lp->lp_actor;
381 
382 	TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
383 	    sizeof(du->ldu_partner));
384 	du->ldu_partner = lp->lp_partner;
385 
386 	TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
387 	    sizeof(du->ldu_collector));
388 	du->ldu_collector.lci_maxdelay = 0;
389 
390 	if (V_lacp_debug > 0) {
391 		lacp_dprintf(lp, "lacpdu transmit\n");
392 		lacp_dump_lacpdu(du);
393 	}
394 
395 	m->m_flags |= M_MCAST;
396 
397 	/*
398 	 * XXX should use higher priority queue.
399 	 * otherwise network congestion can break aggregation.
400 	 */
401 
402 	error = lagg_enqueue(lp->lp_ifp, m);
403 	return (error);
404 }
405 
406 static int
407 lacp_xmit_marker(struct lacp_port *lp)
408 {
409 	struct lagg_port *lgp = lp->lp_lagg;
410 	struct mbuf *m;
411 	struct markerdu *mdu;
412 	int error;
413 
414 	LACP_LOCK_ASSERT(lp->lp_lsc);
415 
416 	m = m_gethdr(M_NOWAIT, MT_DATA);
417 	if (m == NULL) {
418 		return (ENOMEM);
419 	}
420 	m->m_len = m->m_pkthdr.len = sizeof(*mdu);
421 
422 	mdu = mtod(m, struct markerdu *);
423 	memset(mdu, 0, sizeof(*mdu));
424 
425 	memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
426 	    ETHER_ADDR_LEN);
427 	memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
428 	mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
429 
430 	mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
431 	mdu->mdu_sph.sph_version = 1;
432 
433 	/* Bump the transaction id and copy over the marker info */
434 	lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
435 	TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
436 	mdu->mdu_info = lp->lp_marker;
437 
438 	LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
439 	    ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
440 	    ntohl(mdu->mdu_info.mi_rq_xid)));
441 
442 	m->m_flags |= M_MCAST;
443 	error = lagg_enqueue(lp->lp_ifp, m);
444 	return (error);
445 }
446 
447 void
448 lacp_linkstate(struct lagg_port *lgp)
449 {
450 	struct lacp_port *lp = LACP_PORT(lgp);
451 	struct lacp_softc *lsc = lp->lp_lsc;
452 	struct ifnet *ifp = lgp->lp_ifp;
453 	struct ifmediareq ifmr;
454 	int error = 0;
455 	u_int media;
456 	uint8_t old_state;
457 	uint16_t old_key;
458 
459 	bzero((char *)&ifmr, sizeof(ifmr));
460 	error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
461 	if (error != 0)
462 		return;
463 
464 	LACP_LOCK(lsc);
465 	media = ifmr.ifm_active;
466 	LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
467 	    "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
468 	    (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
469 	old_state = lp->lp_state;
470 	old_key = lp->lp_key;
471 
472 	lp->lp_media = media;
473 	/*
474 	 * If the port is not an active full duplex Ethernet link then it can
475 	 * not be aggregated.
476 	 */
477 	if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
478 	    ifp->if_link_state != LINK_STATE_UP) {
479 		lacp_port_disable(lp);
480 	} else {
481 		lacp_port_enable(lp);
482 	}
483 	lp->lp_key = lacp_compose_key(lp);
484 
485 	if (old_state != lp->lp_state || old_key != lp->lp_key) {
486 		LACP_DPRINTF((lp, "-> UNSELECTED\n"));
487 		lp->lp_selected = LACP_UNSELECTED;
488 	}
489 	LACP_UNLOCK(lsc);
490 }
491 
492 static void
493 lacp_tick(void *arg)
494 {
495 	struct lacp_softc *lsc = arg;
496 	struct lacp_port *lp;
497 
498 	LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
499 		if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
500 			continue;
501 
502 		CURVNET_SET(lp->lp_ifp->if_vnet);
503 		lacp_run_timers(lp);
504 
505 		lacp_select(lp);
506 		lacp_sm_mux(lp);
507 		lacp_sm_tx(lp);
508 		lacp_sm_ptx_tx_schedule(lp);
509 		CURVNET_RESTORE();
510 	}
511 	callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
512 }
513 
514 int
515 lacp_port_create(struct lagg_port *lgp)
516 {
517 	struct lagg_softc *sc = lgp->lp_softc;
518 	struct lacp_softc *lsc = LACP_SOFTC(sc);
519 	struct lacp_port *lp;
520 	struct ifnet *ifp = lgp->lp_ifp;
521 	struct sockaddr_dl sdl;
522 	struct ifmultiaddr *rifma = NULL;
523 	int error;
524 
525 	boolean_t active = TRUE; /* XXX should be configurable */
526 	boolean_t fast = FALSE; /* XXX should be configurable */
527 
528 	link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
529 	sdl.sdl_alen = ETHER_ADDR_LEN;
530 
531 	bcopy(&ethermulticastaddr_slowprotocols,
532 	    LLADDR(&sdl), ETHER_ADDR_LEN);
533 	error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
534 	if (error) {
535 		printf("%s: ADDMULTI failed on %s\n", __func__,
536 		    lgp->lp_ifp->if_xname);
537 		return (error);
538 	}
539 
540 	lp = malloc(sizeof(struct lacp_port),
541 	    M_DEVBUF, M_NOWAIT|M_ZERO);
542 	if (lp == NULL)
543 		return (ENOMEM);
544 
545 	LACP_LOCK(lsc);
546 	lgp->lp_psc = lp;
547 	lp->lp_ifp = ifp;
548 	lp->lp_lagg = lgp;
549 	lp->lp_lsc = lsc;
550 	lp->lp_ifma = rifma;
551 
552 	LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
553 
554 	lacp_fill_actorinfo(lp, &lp->lp_actor);
555 	lacp_fill_markerinfo(lp, &lp->lp_marker);
556 	lp->lp_state =
557 	    (active ? LACP_STATE_ACTIVITY : 0) |
558 	    (fast ? LACP_STATE_TIMEOUT : 0);
559 	lp->lp_aggregator = NULL;
560 	lacp_sm_rx_set_expired(lp);
561 	LACP_UNLOCK(lsc);
562 	lacp_linkstate(lgp);
563 
564 	return (0);
565 }
566 
567 void
568 lacp_port_destroy(struct lagg_port *lgp)
569 {
570 	struct lacp_port *lp = LACP_PORT(lgp);
571 	struct lacp_softc *lsc = lp->lp_lsc;
572 	int i;
573 
574 	LACP_LOCK(lsc);
575 	for (i = 0; i < LACP_NTIMER; i++) {
576 		LACP_TIMER_DISARM(lp, i);
577 	}
578 
579 	lacp_disable_collecting(lp);
580 	lacp_disable_distributing(lp);
581 	lacp_unselect(lp);
582 
583 	LIST_REMOVE(lp, lp_next);
584 	LACP_UNLOCK(lsc);
585 
586 	/* The address may have already been removed by if_purgemaddrs() */
587 	if (!lgp->lp_detaching)
588 		if_delmulti_ifma(lp->lp_ifma);
589 
590 	free(lp, M_DEVBUF);
591 }
592 
593 void
594 lacp_req(struct lagg_softc *sc, void *data)
595 {
596 	struct lacp_opreq *req = (struct lacp_opreq *)data;
597 	struct lacp_softc *lsc = LACP_SOFTC(sc);
598 	struct lacp_aggregator *la;
599 
600 	bzero(req, sizeof(struct lacp_opreq));
601 
602 	/*
603 	 * If the LACP softc is NULL, return with the opreq structure full of
604 	 * zeros.  It is normal for the softc to be NULL while the lagg is
605 	 * being destroyed.
606 	 */
607 	if (NULL == lsc)
608 		return;
609 
610 	la = lsc->lsc_active_aggregator;
611 	LACP_LOCK(lsc);
612 	if (la != NULL) {
613 		req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
614 		memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
615 		    ETHER_ADDR_LEN);
616 		req->actor_key = ntohs(la->la_actor.lip_key);
617 		req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
618 		req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
619 		req->actor_state = la->la_actor.lip_state;
620 
621 		req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
622 		memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
623 		    ETHER_ADDR_LEN);
624 		req->partner_key = ntohs(la->la_partner.lip_key);
625 		req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
626 		req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
627 		req->partner_state = la->la_partner.lip_state;
628 	}
629 	LACP_UNLOCK(lsc);
630 }
631 
632 void
633 lacp_portreq(struct lagg_port *lgp, void *data)
634 {
635 	struct lacp_opreq *req = (struct lacp_opreq *)data;
636 	struct lacp_port *lp = LACP_PORT(lgp);
637 	struct lacp_softc *lsc = lp->lp_lsc;
638 
639 	LACP_LOCK(lsc);
640 	req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
641 	memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
642 	    ETHER_ADDR_LEN);
643 	req->actor_key = ntohs(lp->lp_actor.lip_key);
644 	req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
645 	req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
646 	req->actor_state = lp->lp_actor.lip_state;
647 
648 	req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
649 	memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
650 	    ETHER_ADDR_LEN);
651 	req->partner_key = ntohs(lp->lp_partner.lip_key);
652 	req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
653 	req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
654 	req->partner_state = lp->lp_partner.lip_state;
655 	LACP_UNLOCK(lsc);
656 }
657 
658 static void
659 lacp_disable_collecting(struct lacp_port *lp)
660 {
661 	LACP_DPRINTF((lp, "collecting disabled\n"));
662 	lp->lp_state &= ~LACP_STATE_COLLECTING;
663 }
664 
665 static void
666 lacp_enable_collecting(struct lacp_port *lp)
667 {
668 	LACP_DPRINTF((lp, "collecting enabled\n"));
669 	lp->lp_state |= LACP_STATE_COLLECTING;
670 }
671 
672 static void
673 lacp_disable_distributing(struct lacp_port *lp)
674 {
675 	struct lacp_aggregator *la = lp->lp_aggregator;
676 	struct lacp_softc *lsc = lp->lp_lsc;
677 	struct lagg_softc *sc = lsc->lsc_softc;
678 	char buf[LACP_LAGIDSTR_MAX+1];
679 
680 	LACP_LOCK_ASSERT(lsc);
681 
682 	if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
683 		return;
684 	}
685 
686 	KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
687 	KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
688 	KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
689 
690 	LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
691 	    "nports %d -> %d\n",
692 	    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
693 	    la->la_nports, la->la_nports - 1));
694 
695 	TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
696 	la->la_nports--;
697 	sc->sc_active = la->la_nports;
698 
699 	if (lsc->lsc_active_aggregator == la) {
700 		lacp_suppress_distributing(lsc, la);
701 		lacp_select_active_aggregator(lsc);
702 		/* regenerate the port map, the active aggregator has changed */
703 		lacp_update_portmap(lsc);
704 	}
705 
706 	lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
707 }
708 
709 static void
710 lacp_enable_distributing(struct lacp_port *lp)
711 {
712 	struct lacp_aggregator *la = lp->lp_aggregator;
713 	struct lacp_softc *lsc = lp->lp_lsc;
714 	struct lagg_softc *sc = lsc->lsc_softc;
715 	char buf[LACP_LAGIDSTR_MAX+1];
716 
717 	LACP_LOCK_ASSERT(lsc);
718 
719 	if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
720 		return;
721 	}
722 
723 	LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
724 	    "nports %d -> %d\n",
725 	    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
726 	    la->la_nports, la->la_nports + 1));
727 
728 	KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
729 	TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
730 	la->la_nports++;
731 	sc->sc_active = la->la_nports;
732 
733 	lp->lp_state |= LACP_STATE_DISTRIBUTING;
734 
735 	if (lsc->lsc_active_aggregator == la) {
736 		lacp_suppress_distributing(lsc, la);
737 		lacp_update_portmap(lsc);
738 	} else
739 		/* try to become the active aggregator */
740 		lacp_select_active_aggregator(lsc);
741 }
742 
743 static void
744 lacp_transit_expire(void *vp)
745 {
746 	struct lacp_softc *lsc = vp;
747 
748 	LACP_LOCK_ASSERT(lsc);
749 
750 	CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
751 	LACP_TRACE(NULL);
752 	CURVNET_RESTORE();
753 
754 	lsc->lsc_suppress_distributing = FALSE;
755 }
756 
757 void
758 lacp_attach(struct lagg_softc *sc)
759 {
760 	struct lacp_softc *lsc;
761 	uint32_t seed;
762 
763 	lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
764 
765 	sc->sc_psc = lsc;
766 	lsc->lsc_softc = sc;
767 
768 	seed = arc4random();
769 	lsc->lsc_hashkey = FNV1_32_INIT;
770 	lsc->lsc_hashkey = fnv_32_buf(&seed, sizeof(seed), lsc->lsc_hashkey);
771 	lsc->lsc_active_aggregator = NULL;
772 	lsc->lsc_strict_mode = 1;
773 	LACP_LOCK_INIT(lsc);
774 	TAILQ_INIT(&lsc->lsc_aggregators);
775 	LIST_INIT(&lsc->lsc_ports);
776 
777 	callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
778 	callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
779 
780 	/* if the lagg is already up then do the same */
781 	if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
782 		lacp_init(sc);
783 }
784 
785 void
786 lacp_detach(void *psc)
787 {
788 	struct lacp_softc *lsc = (struct lacp_softc *)psc;
789 
790 	KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
791 	    ("aggregators still active"));
792 	KASSERT(lsc->lsc_active_aggregator == NULL,
793 	    ("aggregator still attached"));
794 
795 	callout_drain(&lsc->lsc_transit_callout);
796 	callout_drain(&lsc->lsc_callout);
797 
798 	LACP_LOCK_DESTROY(lsc);
799 	free(lsc, M_DEVBUF);
800 }
801 
802 void
803 lacp_init(struct lagg_softc *sc)
804 {
805 	struct lacp_softc *lsc = LACP_SOFTC(sc);
806 
807 	LACP_LOCK(lsc);
808 	callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
809 	LACP_UNLOCK(lsc);
810 }
811 
812 void
813 lacp_stop(struct lagg_softc *sc)
814 {
815 	struct lacp_softc *lsc = LACP_SOFTC(sc);
816 
817 	LACP_LOCK(lsc);
818 	callout_stop(&lsc->lsc_transit_callout);
819 	callout_stop(&lsc->lsc_callout);
820 	LACP_UNLOCK(lsc);
821 }
822 
823 struct lagg_port *
824 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
825 {
826 	struct lacp_softc *lsc = LACP_SOFTC(sc);
827 	struct lacp_portmap *pm;
828 	struct lacp_port *lp;
829 	uint32_t hash;
830 
831 	if (__predict_false(lsc->lsc_suppress_distributing)) {
832 		LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
833 		return (NULL);
834 	}
835 
836 	pm = &lsc->lsc_pmap[lsc->lsc_activemap];
837 	if (pm->pm_count == 0) {
838 		LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
839 		return (NULL);
840 	}
841 
842 	if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
843 	    M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
844 		hash = m->m_pkthdr.flowid >> sc->flowid_shift;
845 	else
846 		hash = lagg_hashmbuf(sc, m, lsc->lsc_hashkey);
847 	hash %= pm->pm_count;
848 	lp = pm->pm_map[hash];
849 
850 	KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
851 	    ("aggregated port is not distributing"));
852 
853 	return (lp->lp_lagg);
854 }
855 /*
856  * lacp_suppress_distributing: drop transmit packets for a while
857  * to preserve packet ordering.
858  */
859 
860 static void
861 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
862 {
863 	struct lacp_port *lp;
864 
865 	if (lsc->lsc_active_aggregator != la) {
866 		return;
867 	}
868 
869 	LACP_TRACE(NULL);
870 
871 	lsc->lsc_suppress_distributing = TRUE;
872 
873 	/* send a marker frame down each port to verify the queues are empty */
874 	LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
875 		lp->lp_flags |= LACP_PORT_MARK;
876 		lacp_xmit_marker(lp);
877 	}
878 
879 	/* set a timeout for the marker frames */
880 	callout_reset(&lsc->lsc_transit_callout,
881 	    LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
882 }
883 
884 static int
885 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
886     const struct lacp_peerinfo *b)
887 {
888 	return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
889 }
890 
891 static int
892 lacp_compare_systemid(const struct lacp_systemid *a,
893     const struct lacp_systemid *b)
894 {
895 	return (memcmp(a, b, sizeof(*a)));
896 }
897 
898 #if 0	/* unused */
899 static int
900 lacp_compare_portid(const struct lacp_portid *a,
901     const struct lacp_portid *b)
902 {
903 	return (memcmp(a, b, sizeof(*a)));
904 }
905 #endif
906 
907 static uint64_t
908 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
909 {
910 	struct lacp_port *lp;
911 	uint64_t speed;
912 
913 	lp = TAILQ_FIRST(&la->la_ports);
914 	if (lp == NULL) {
915 		return (0);
916 	}
917 
918 	speed = ifmedia_baudrate(lp->lp_media);
919 	speed *= la->la_nports;
920 	if (speed == 0) {
921 		LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
922 		    lp->lp_media, la->la_nports));
923 	}
924 
925 	return (speed);
926 }
927 
928 /*
929  * lacp_select_active_aggregator: select an aggregator to be used to transmit
930  * packets from lagg(4) interface.
931  */
932 
933 static void
934 lacp_select_active_aggregator(struct lacp_softc *lsc)
935 {
936 	struct lacp_aggregator *la;
937 	struct lacp_aggregator *best_la = NULL;
938 	uint64_t best_speed = 0;
939 	char buf[LACP_LAGIDSTR_MAX+1];
940 
941 	LACP_TRACE(NULL);
942 
943 	TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
944 		uint64_t speed;
945 
946 		if (la->la_nports == 0) {
947 			continue;
948 		}
949 
950 		speed = lacp_aggregator_bandwidth(la);
951 		LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
952 		    lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
953 		    speed, la->la_nports));
954 
955 		/*
956 		 * This aggregator is chosen if the partner has a better
957 		 * system priority or, the total aggregated speed is higher
958 		 * or, it is already the chosen aggregator
959 		 */
960 		if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
961 		    LACP_SYS_PRI(best_la->la_partner)) ||
962 		    speed > best_speed ||
963 		    (speed == best_speed &&
964 		    la == lsc->lsc_active_aggregator)) {
965 			best_la = la;
966 			best_speed = speed;
967 		}
968 	}
969 
970 	KASSERT(best_la == NULL || best_la->la_nports > 0,
971 	    ("invalid aggregator refcnt"));
972 	KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
973 	    ("invalid aggregator list"));
974 
975 	if (lsc->lsc_active_aggregator != best_la) {
976 		LACP_DPRINTF((NULL, "active aggregator changed\n"));
977 		LACP_DPRINTF((NULL, "old %s\n",
978 		    lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
979 		    buf, sizeof(buf))));
980 	} else {
981 		LACP_DPRINTF((NULL, "active aggregator not changed\n"));
982 	}
983 	LACP_DPRINTF((NULL, "new %s\n",
984 	    lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
985 
986 	if (lsc->lsc_active_aggregator != best_la) {
987 		lsc->lsc_active_aggregator = best_la;
988 		lacp_update_portmap(lsc);
989 		if (best_la) {
990 			lacp_suppress_distributing(lsc, best_la);
991 		}
992 	}
993 }
994 
995 /*
996  * Updated the inactive portmap array with the new list of ports and
997  * make it live.
998  */
999 static void
1000 lacp_update_portmap(struct lacp_softc *lsc)
1001 {
1002 	struct lagg_softc *sc = lsc->lsc_softc;
1003 	struct lacp_aggregator *la;
1004 	struct lacp_portmap *p;
1005 	struct lacp_port *lp;
1006 	uint64_t speed;
1007 	u_int newmap;
1008 	int i;
1009 
1010 	newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1011 	p = &lsc->lsc_pmap[newmap];
1012 	la = lsc->lsc_active_aggregator;
1013 	speed = 0;
1014 	bzero(p, sizeof(struct lacp_portmap));
1015 
1016 	if (la != NULL && la->la_nports > 0) {
1017 		p->pm_count = la->la_nports;
1018 		i = 0;
1019 		TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1020 			p->pm_map[i++] = lp;
1021 		KASSERT(i == p->pm_count, ("Invalid port count"));
1022 		speed = lacp_aggregator_bandwidth(la);
1023 	}
1024 	sc->sc_ifp->if_baudrate = speed;
1025 
1026 	/* switch the active portmap over */
1027 	atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1028 	LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1029 		    lsc->lsc_activemap,
1030 		    lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1031 }
1032 
1033 static uint16_t
1034 lacp_compose_key(struct lacp_port *lp)
1035 {
1036 	struct lagg_port *lgp = lp->lp_lagg;
1037 	struct lagg_softc *sc = lgp->lp_softc;
1038 	u_int media = lp->lp_media;
1039 	uint16_t key;
1040 
1041 	if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1042 
1043 		/*
1044 		 * non-aggregatable links should have unique keys.
1045 		 *
1046 		 * XXX this isn't really unique as if_index is 16 bit.
1047 		 */
1048 
1049 		/* bit 0..14:	(some bits of) if_index of this port */
1050 		key = lp->lp_ifp->if_index;
1051 		/* bit 15:	1 */
1052 		key |= 0x8000;
1053 	} else {
1054 		u_int subtype = IFM_SUBTYPE(media);
1055 
1056 		KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1057 		KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1058 
1059 		/* bit 0..4:	IFM_SUBTYPE modulo speed */
1060 		switch (subtype) {
1061 		case IFM_10_T:
1062 		case IFM_10_2:
1063 		case IFM_10_5:
1064 		case IFM_10_STP:
1065 		case IFM_10_FL:
1066 			key = IFM_10_T;
1067 			break;
1068 		case IFM_100_TX:
1069 		case IFM_100_FX:
1070 		case IFM_100_T4:
1071 		case IFM_100_VG:
1072 		case IFM_100_T2:
1073 			key = IFM_100_TX;
1074 			break;
1075 		case IFM_1000_SX:
1076 		case IFM_1000_LX:
1077 		case IFM_1000_CX:
1078 		case IFM_1000_T:
1079 			key = IFM_1000_SX;
1080 			break;
1081 		case IFM_10G_LR:
1082 		case IFM_10G_SR:
1083 		case IFM_10G_CX4:
1084 		case IFM_10G_TWINAX:
1085 		case IFM_10G_TWINAX_LONG:
1086 		case IFM_10G_LRM:
1087 		case IFM_10G_T:
1088 			key = IFM_10G_LR;
1089 			break;
1090 		case IFM_40G_CR4:
1091 		case IFM_40G_SR4:
1092 		case IFM_40G_LR4:
1093 			key = IFM_40G_CR4;
1094 			break;
1095 		default:
1096 			key = subtype;
1097 		}
1098 		/* bit 5..14:	(some bits of) if_index of lagg device */
1099 		key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1100 		/* bit 15:	0 */
1101 	}
1102 	return (htons(key));
1103 }
1104 
1105 static void
1106 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1107 {
1108 	char buf[LACP_LAGIDSTR_MAX+1];
1109 
1110 	LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1111 	    __func__,
1112 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1113 	    buf, sizeof(buf)),
1114 	    la->la_refcnt, la->la_refcnt + 1));
1115 
1116 	KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1117 	la->la_refcnt++;
1118 	KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1119 }
1120 
1121 static void
1122 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1123 {
1124 	char buf[LACP_LAGIDSTR_MAX+1];
1125 
1126 	LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1127 	    __func__,
1128 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1129 	    buf, sizeof(buf)),
1130 	    la->la_refcnt, la->la_refcnt - 1));
1131 
1132 	KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1133 	la->la_refcnt--;
1134 	if (la->la_refcnt > 0) {
1135 		return;
1136 	}
1137 
1138 	KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1139 	KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1140 
1141 	TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1142 
1143 	free(la, M_DEVBUF);
1144 }
1145 
1146 /*
1147  * lacp_aggregator_get: allocate an aggregator.
1148  */
1149 
1150 static struct lacp_aggregator *
1151 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1152 {
1153 	struct lacp_aggregator *la;
1154 
1155 	la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1156 	if (la) {
1157 		la->la_refcnt = 1;
1158 		la->la_nports = 0;
1159 		TAILQ_INIT(&la->la_ports);
1160 		la->la_pending = 0;
1161 		TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1162 	}
1163 
1164 	return (la);
1165 }
1166 
1167 /*
1168  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1169  */
1170 
1171 static void
1172 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1173 {
1174 	lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1175 	lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1176 
1177 	la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1178 }
1179 
1180 static void
1181 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1182     const struct lacp_peerinfo *lpi_port)
1183 {
1184 	memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1185 	lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1186 	lpi_aggr->lip_key = lpi_port->lip_key;
1187 }
1188 
1189 /*
1190  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1191  */
1192 
1193 static int
1194 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1195     const struct lacp_port *lp)
1196 {
1197 	if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1198 	    !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1199 		return (0);
1200 	}
1201 
1202 	if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1203 		return (0);
1204 	}
1205 
1206 	if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1207 		return (0);
1208 	}
1209 
1210 	if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1211 		return (0);
1212 	}
1213 
1214 	return (1);
1215 }
1216 
1217 static int
1218 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1219     const struct lacp_peerinfo *b)
1220 {
1221 	if (memcmp(&a->lip_systemid, &b->lip_systemid,
1222 	    sizeof(a->lip_systemid))) {
1223 		return (0);
1224 	}
1225 
1226 	if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1227 		return (0);
1228 	}
1229 
1230 	return (1);
1231 }
1232 
1233 static void
1234 lacp_port_enable(struct lacp_port *lp)
1235 {
1236 	lp->lp_state |= LACP_STATE_AGGREGATION;
1237 }
1238 
1239 static void
1240 lacp_port_disable(struct lacp_port *lp)
1241 {
1242 	lacp_set_mux(lp, LACP_MUX_DETACHED);
1243 
1244 	lp->lp_state &= ~LACP_STATE_AGGREGATION;
1245 	lp->lp_selected = LACP_UNSELECTED;
1246 	lacp_sm_rx_record_default(lp);
1247 	lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1248 	lp->lp_state &= ~LACP_STATE_EXPIRED;
1249 }
1250 
1251 /*
1252  * lacp_select: select an aggregator.  create one if necessary.
1253  */
1254 static void
1255 lacp_select(struct lacp_port *lp)
1256 {
1257 	struct lacp_softc *lsc = lp->lp_lsc;
1258 	struct lacp_aggregator *la;
1259 	char buf[LACP_LAGIDSTR_MAX+1];
1260 
1261 	if (lp->lp_aggregator) {
1262 		return;
1263 	}
1264 
1265 	KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1266 	    ("timer_wait_while still active"));
1267 
1268 	LACP_DPRINTF((lp, "port lagid=%s\n",
1269 	    lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1270 	    buf, sizeof(buf))));
1271 
1272 	TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1273 		if (lacp_aggregator_is_compatible(la, lp)) {
1274 			break;
1275 		}
1276 	}
1277 
1278 	if (la == NULL) {
1279 		la = lacp_aggregator_get(lsc, lp);
1280 		if (la == NULL) {
1281 			LACP_DPRINTF((lp, "aggregator creation failed\n"));
1282 
1283 			/*
1284 			 * will retry on the next tick.
1285 			 */
1286 
1287 			return;
1288 		}
1289 		lacp_fill_aggregator_id(la, lp);
1290 		LACP_DPRINTF((lp, "aggregator created\n"));
1291 	} else {
1292 		LACP_DPRINTF((lp, "compatible aggregator found\n"));
1293 		if (la->la_refcnt == LACP_MAX_PORTS)
1294 			return;
1295 		lacp_aggregator_addref(lsc, la);
1296 	}
1297 
1298 	LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1299 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1300 	    buf, sizeof(buf))));
1301 
1302 	lp->lp_aggregator = la;
1303 	lp->lp_selected = LACP_SELECTED;
1304 }
1305 
1306 /*
1307  * lacp_unselect: finish unselect/detach process.
1308  */
1309 
1310 static void
1311 lacp_unselect(struct lacp_port *lp)
1312 {
1313 	struct lacp_softc *lsc = lp->lp_lsc;
1314 	struct lacp_aggregator *la = lp->lp_aggregator;
1315 
1316 	KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1317 	    ("timer_wait_while still active"));
1318 
1319 	if (la == NULL) {
1320 		return;
1321 	}
1322 
1323 	lp->lp_aggregator = NULL;
1324 	lacp_aggregator_delref(lsc, la);
1325 }
1326 
1327 /* mux machine */
1328 
1329 static void
1330 lacp_sm_mux(struct lacp_port *lp)
1331 {
1332 	struct lagg_port *lgp = lp->lp_lagg;
1333 	struct lagg_softc *sc = lgp->lp_softc;
1334 	enum lacp_mux_state new_state;
1335 	boolean_t p_sync =
1336 		    (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1337 	boolean_t p_collecting =
1338 	    (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1339 	enum lacp_selected selected = lp->lp_selected;
1340 	struct lacp_aggregator *la;
1341 
1342 	if (V_lacp_debug > 1)
1343 		lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1344 		    "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1345 		    lp->lp_mux_state, selected, p_sync, p_collecting);
1346 
1347 re_eval:
1348 	la = lp->lp_aggregator;
1349 	KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1350 	    ("MUX not detached"));
1351 	new_state = lp->lp_mux_state;
1352 	switch (lp->lp_mux_state) {
1353 	case LACP_MUX_DETACHED:
1354 		if (selected != LACP_UNSELECTED) {
1355 			new_state = LACP_MUX_WAITING;
1356 		}
1357 		break;
1358 	case LACP_MUX_WAITING:
1359 		KASSERT(la->la_pending > 0 ||
1360 		    !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1361 		    ("timer_wait_while still active"));
1362 		if (selected == LACP_SELECTED && la->la_pending == 0) {
1363 			new_state = LACP_MUX_ATTACHED;
1364 		} else if (selected == LACP_UNSELECTED) {
1365 			new_state = LACP_MUX_DETACHED;
1366 		}
1367 		break;
1368 	case LACP_MUX_ATTACHED:
1369 		if (selected == LACP_SELECTED && p_sync) {
1370 			new_state = LACP_MUX_COLLECTING;
1371 		} else if (selected != LACP_SELECTED) {
1372 			new_state = LACP_MUX_DETACHED;
1373 		}
1374 		break;
1375 	case LACP_MUX_COLLECTING:
1376 		if (selected == LACP_SELECTED && p_sync && p_collecting) {
1377 			new_state = LACP_MUX_DISTRIBUTING;
1378 		} else if (selected != LACP_SELECTED || !p_sync) {
1379 			new_state = LACP_MUX_ATTACHED;
1380 		}
1381 		break;
1382 	case LACP_MUX_DISTRIBUTING:
1383 		if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1384 			new_state = LACP_MUX_COLLECTING;
1385 			lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1386 			sc->sc_flapping++;
1387 		}
1388 		break;
1389 	default:
1390 		panic("%s: unknown state", __func__);
1391 	}
1392 
1393 	if (lp->lp_mux_state == new_state) {
1394 		return;
1395 	}
1396 
1397 	lacp_set_mux(lp, new_state);
1398 	goto re_eval;
1399 }
1400 
1401 static void
1402 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1403 {
1404 	struct lacp_aggregator *la = lp->lp_aggregator;
1405 
1406 	if (lp->lp_mux_state == new_state) {
1407 		return;
1408 	}
1409 
1410 	switch (new_state) {
1411 	case LACP_MUX_DETACHED:
1412 		lp->lp_state &= ~LACP_STATE_SYNC;
1413 		lacp_disable_distributing(lp);
1414 		lacp_disable_collecting(lp);
1415 		lacp_sm_assert_ntt(lp);
1416 		/* cancel timer */
1417 		if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1418 			KASSERT(la->la_pending > 0,
1419 			    ("timer_wait_while not active"));
1420 			la->la_pending--;
1421 		}
1422 		LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1423 		lacp_unselect(lp);
1424 		break;
1425 	case LACP_MUX_WAITING:
1426 		LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1427 		    LACP_AGGREGATE_WAIT_TIME);
1428 		la->la_pending++;
1429 		break;
1430 	case LACP_MUX_ATTACHED:
1431 		lp->lp_state |= LACP_STATE_SYNC;
1432 		lacp_disable_collecting(lp);
1433 		lacp_sm_assert_ntt(lp);
1434 		break;
1435 	case LACP_MUX_COLLECTING:
1436 		lacp_enable_collecting(lp);
1437 		lacp_disable_distributing(lp);
1438 		lacp_sm_assert_ntt(lp);
1439 		break;
1440 	case LACP_MUX_DISTRIBUTING:
1441 		lacp_enable_distributing(lp);
1442 		break;
1443 	default:
1444 		panic("%s: unknown state", __func__);
1445 	}
1446 
1447 	LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1448 
1449 	lp->lp_mux_state = new_state;
1450 }
1451 
1452 static void
1453 lacp_sm_mux_timer(struct lacp_port *lp)
1454 {
1455 	struct lacp_aggregator *la = lp->lp_aggregator;
1456 	char buf[LACP_LAGIDSTR_MAX+1];
1457 
1458 	KASSERT(la->la_pending > 0, ("no pending event"));
1459 
1460 	LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1461 	    lacp_format_lagid(&la->la_actor, &la->la_partner,
1462 	    buf, sizeof(buf)),
1463 	    la->la_pending, la->la_pending - 1));
1464 
1465 	la->la_pending--;
1466 }
1467 
1468 /* periodic transmit machine */
1469 
1470 static void
1471 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1472 {
1473 	if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1474 	    LACP_STATE_TIMEOUT)) {
1475 		return;
1476 	}
1477 
1478 	LACP_DPRINTF((lp, "partner timeout changed\n"));
1479 
1480 	/*
1481 	 * FAST_PERIODIC -> SLOW_PERIODIC
1482 	 * or
1483 	 * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1484 	 *
1485 	 * let lacp_sm_ptx_tx_schedule to update timeout.
1486 	 */
1487 
1488 	LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1489 
1490 	/*
1491 	 * if timeout has been shortened, assert NTT.
1492 	 */
1493 
1494 	if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1495 		lacp_sm_assert_ntt(lp);
1496 	}
1497 }
1498 
1499 static void
1500 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1501 {
1502 	int timeout;
1503 
1504 	if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1505 	    !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1506 
1507 		/*
1508 		 * NO_PERIODIC
1509 		 */
1510 
1511 		LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1512 		return;
1513 	}
1514 
1515 	if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1516 		return;
1517 	}
1518 
1519 	timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1520 	    LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1521 
1522 	LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1523 }
1524 
1525 static void
1526 lacp_sm_ptx_timer(struct lacp_port *lp)
1527 {
1528 	lacp_sm_assert_ntt(lp);
1529 }
1530 
1531 static void
1532 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1533 {
1534 	int timeout;
1535 
1536 	/*
1537 	 * check LACP_DISABLED first
1538 	 */
1539 
1540 	if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1541 		return;
1542 	}
1543 
1544 	/*
1545 	 * check loopback condition.
1546 	 */
1547 
1548 	if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1549 	    &lp->lp_actor.lip_systemid)) {
1550 		return;
1551 	}
1552 
1553 	/*
1554 	 * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1555 	 */
1556 
1557 	lacp_sm_rx_update_selected(lp, du);
1558 	lacp_sm_rx_update_ntt(lp, du);
1559 	lacp_sm_rx_record_pdu(lp, du);
1560 
1561 	timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1562 	    LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1563 	LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1564 
1565 	lp->lp_state &= ~LACP_STATE_EXPIRED;
1566 
1567 	/*
1568 	 * kick transmit machine without waiting the next tick.
1569 	 */
1570 
1571 	lacp_sm_tx(lp);
1572 }
1573 
1574 static void
1575 lacp_sm_rx_set_expired(struct lacp_port *lp)
1576 {
1577 	lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1578 	lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1579 	LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1580 	lp->lp_state |= LACP_STATE_EXPIRED;
1581 }
1582 
1583 static void
1584 lacp_sm_rx_timer(struct lacp_port *lp)
1585 {
1586 	if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1587 		/* CURRENT -> EXPIRED */
1588 		LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1589 		lacp_sm_rx_set_expired(lp);
1590 	} else {
1591 		/* EXPIRED -> DEFAULTED */
1592 		LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1593 		lacp_sm_rx_update_default_selected(lp);
1594 		lacp_sm_rx_record_default(lp);
1595 		lp->lp_state &= ~LACP_STATE_EXPIRED;
1596 	}
1597 }
1598 
1599 static void
1600 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1601 {
1602 	boolean_t active;
1603 	uint8_t oldpstate;
1604 	char buf[LACP_STATESTR_MAX+1];
1605 
1606 	LACP_TRACE(lp);
1607 
1608 	oldpstate = lp->lp_partner.lip_state;
1609 
1610 	active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1611 	    || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1612 	    (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1613 
1614 	lp->lp_partner = du->ldu_actor;
1615 	if (active &&
1616 	    ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1617 	    LACP_STATE_AGGREGATION) &&
1618 	    !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1619 	    || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1620 		/* XXX nothing? */
1621 	} else {
1622 		lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1623 	}
1624 
1625 	lp->lp_state &= ~LACP_STATE_DEFAULTED;
1626 
1627 	if (oldpstate != lp->lp_partner.lip_state) {
1628 		LACP_DPRINTF((lp, "old pstate %s\n",
1629 		    lacp_format_state(oldpstate, buf, sizeof(buf))));
1630 		LACP_DPRINTF((lp, "new pstate %s\n",
1631 		    lacp_format_state(lp->lp_partner.lip_state, buf,
1632 		    sizeof(buf))));
1633 	}
1634 
1635 	/* XXX Hack, still need to implement 5.4.9 para 2,3,4 */
1636 	if (lp->lp_lsc->lsc_strict_mode)
1637 		lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1638 
1639 	lacp_sm_ptx_update_timeout(lp, oldpstate);
1640 }
1641 
1642 static void
1643 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1644 {
1645 
1646 	LACP_TRACE(lp);
1647 
1648 	if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1649 	    !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1650 	    LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1651 		LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1652 		lacp_sm_assert_ntt(lp);
1653 	}
1654 }
1655 
1656 static void
1657 lacp_sm_rx_record_default(struct lacp_port *lp)
1658 {
1659 	uint8_t oldpstate;
1660 
1661 	LACP_TRACE(lp);
1662 
1663 	oldpstate = lp->lp_partner.lip_state;
1664 	if (lp->lp_lsc->lsc_strict_mode)
1665 		lp->lp_partner = lacp_partner_admin_strict;
1666 	else
1667 		lp->lp_partner = lacp_partner_admin_optimistic;;
1668 	lp->lp_state |= LACP_STATE_DEFAULTED;
1669 	lacp_sm_ptx_update_timeout(lp, oldpstate);
1670 }
1671 
1672 static void
1673 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1674     const struct lacp_peerinfo *info)
1675 {
1676 
1677 	LACP_TRACE(lp);
1678 
1679 	if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1680 	    !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1681 	    LACP_STATE_AGGREGATION)) {
1682 		lp->lp_selected = LACP_UNSELECTED;
1683 		/* mux machine will clean up lp->lp_aggregator */
1684 	}
1685 }
1686 
1687 static void
1688 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1689 {
1690 
1691 	LACP_TRACE(lp);
1692 
1693 	lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1694 }
1695 
1696 static void
1697 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1698 {
1699 
1700 	LACP_TRACE(lp);
1701 
1702 	if (lp->lp_lsc->lsc_strict_mode)
1703 		lacp_sm_rx_update_selected_from_peerinfo(lp,
1704 		    &lacp_partner_admin_strict);
1705 	else
1706 		lacp_sm_rx_update_selected_from_peerinfo(lp,
1707 		    &lacp_partner_admin_optimistic);
1708 }
1709 
1710 /* transmit machine */
1711 
1712 static void
1713 lacp_sm_tx(struct lacp_port *lp)
1714 {
1715 	int error = 0;
1716 
1717 	if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1718 #if 1
1719 	    || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1720 	    && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1721 #endif
1722 	    ) {
1723 		lp->lp_flags &= ~LACP_PORT_NTT;
1724 	}
1725 
1726 	if (!(lp->lp_flags & LACP_PORT_NTT)) {
1727 		return;
1728 	}
1729 
1730 	/* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1731 	if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1732 		    (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1733 		LACP_DPRINTF((lp, "rate limited pdu\n"));
1734 		return;
1735 	}
1736 
1737 	if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1738 		error = lacp_xmit_lacpdu(lp);
1739 	} else {
1740 		LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1741 	}
1742 
1743 	if (error == 0) {
1744 		lp->lp_flags &= ~LACP_PORT_NTT;
1745 	} else {
1746 		LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1747 		    error));
1748 	}
1749 }
1750 
1751 static void
1752 lacp_sm_assert_ntt(struct lacp_port *lp)
1753 {
1754 
1755 	lp->lp_flags |= LACP_PORT_NTT;
1756 }
1757 
1758 static void
1759 lacp_run_timers(struct lacp_port *lp)
1760 {
1761 	int i;
1762 
1763 	for (i = 0; i < LACP_NTIMER; i++) {
1764 		KASSERT(lp->lp_timer[i] >= 0,
1765 		    ("invalid timer value %d", lp->lp_timer[i]));
1766 		if (lp->lp_timer[i] == 0) {
1767 			continue;
1768 		} else if (--lp->lp_timer[i] <= 0) {
1769 			if (lacp_timer_funcs[i]) {
1770 				(*lacp_timer_funcs[i])(lp);
1771 			}
1772 		}
1773 	}
1774 }
1775 
1776 int
1777 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1778 {
1779 	struct lacp_softc *lsc = lp->lp_lsc;
1780 	struct lagg_port *lgp = lp->lp_lagg;
1781 	struct lacp_port *lp2;
1782 	struct markerdu *mdu;
1783 	int error = 0;
1784 	int pending = 0;
1785 
1786 	if (m->m_pkthdr.len != sizeof(*mdu)) {
1787 		goto bad;
1788 	}
1789 
1790 	if ((m->m_flags & M_MCAST) == 0) {
1791 		goto bad;
1792 	}
1793 
1794 	if (m->m_len < sizeof(*mdu)) {
1795 		m = m_pullup(m, sizeof(*mdu));
1796 		if (m == NULL) {
1797 			return (ENOMEM);
1798 		}
1799 	}
1800 
1801 	mdu = mtod(m, struct markerdu *);
1802 
1803 	if (memcmp(&mdu->mdu_eh.ether_dhost,
1804 	    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1805 		goto bad;
1806 	}
1807 
1808 	if (mdu->mdu_sph.sph_version != 1) {
1809 		goto bad;
1810 	}
1811 
1812 	switch (mdu->mdu_tlv.tlv_type) {
1813 	case MARKER_TYPE_INFO:
1814 		if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1815 		    marker_info_tlv_template, TRUE)) {
1816 			goto bad;
1817 		}
1818 		mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1819 		memcpy(&mdu->mdu_eh.ether_dhost,
1820 		    &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1821 		memcpy(&mdu->mdu_eh.ether_shost,
1822 		    lgp->lp_lladdr, ETHER_ADDR_LEN);
1823 		error = lagg_enqueue(lp->lp_ifp, m);
1824 		break;
1825 
1826 	case MARKER_TYPE_RESPONSE:
1827 		if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1828 		    marker_response_tlv_template, TRUE)) {
1829 			goto bad;
1830 		}
1831 		LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1832 		    ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1833 		    ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1834 
1835 		/* Verify that it is the last marker we sent out */
1836 		if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1837 		    sizeof(struct lacp_markerinfo)))
1838 			goto bad;
1839 
1840 		LACP_LOCK(lsc);
1841 		lp->lp_flags &= ~LACP_PORT_MARK;
1842 
1843 		if (lsc->lsc_suppress_distributing) {
1844 			/* Check if any ports are waiting for a response */
1845 			LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1846 				if (lp2->lp_flags & LACP_PORT_MARK) {
1847 					pending = 1;
1848 					break;
1849 				}
1850 			}
1851 
1852 			if (pending == 0) {
1853 				/* All interface queues are clear */
1854 				LACP_DPRINTF((NULL, "queue flush complete\n"));
1855 				lsc->lsc_suppress_distributing = FALSE;
1856 			}
1857 		}
1858 		LACP_UNLOCK(lsc);
1859 		m_freem(m);
1860 		break;
1861 
1862 	default:
1863 		goto bad;
1864 	}
1865 
1866 	return (error);
1867 
1868 bad:
1869 	LACP_DPRINTF((lp, "bad marker frame\n"));
1870 	m_freem(m);
1871 	return (EINVAL);
1872 }
1873 
1874 static int
1875 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1876     const struct tlv_template *tmpl, boolean_t check_type)
1877 {
1878 	while (/* CONSTCOND */ 1) {
1879 		if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1880 			return (EINVAL);
1881 		}
1882 		if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1883 		    tlv->tlv_length != tmpl->tmpl_length) {
1884 			return (EINVAL);
1885 		}
1886 		if (tmpl->tmpl_type == 0) {
1887 			break;
1888 		}
1889 		tlv = (const struct tlvhdr *)
1890 		    ((const char *)tlv + tlv->tlv_length);
1891 		tmpl++;
1892 	}
1893 
1894 	return (0);
1895 }
1896 
1897 /* Debugging */
1898 const char *
1899 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1900 {
1901 	snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1902 	    (int)mac[0],
1903 	    (int)mac[1],
1904 	    (int)mac[2],
1905 	    (int)mac[3],
1906 	    (int)mac[4],
1907 	    (int)mac[5]);
1908 
1909 	return (buf);
1910 }
1911 
1912 const char *
1913 lacp_format_systemid(const struct lacp_systemid *sysid,
1914     char *buf, size_t buflen)
1915 {
1916 	char macbuf[LACP_MACSTR_MAX+1];
1917 
1918 	snprintf(buf, buflen, "%04X,%s",
1919 	    ntohs(sysid->lsi_prio),
1920 	    lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1921 
1922 	return (buf);
1923 }
1924 
1925 const char *
1926 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
1927 {
1928 	snprintf(buf, buflen, "%04X,%04X",
1929 	    ntohs(portid->lpi_prio),
1930 	    ntohs(portid->lpi_portno));
1931 
1932 	return (buf);
1933 }
1934 
1935 const char *
1936 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
1937 {
1938 	char sysid[LACP_SYSTEMIDSTR_MAX+1];
1939 	char portid[LACP_PORTIDSTR_MAX+1];
1940 
1941 	snprintf(buf, buflen, "(%s,%04X,%s)",
1942 	    lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
1943 	    ntohs(peer->lip_key),
1944 	    lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
1945 
1946 	return (buf);
1947 }
1948 
1949 const char *
1950 lacp_format_lagid(const struct lacp_peerinfo *a,
1951     const struct lacp_peerinfo *b, char *buf, size_t buflen)
1952 {
1953 	char astr[LACP_PARTNERSTR_MAX+1];
1954 	char bstr[LACP_PARTNERSTR_MAX+1];
1955 
1956 #if 0
1957 	/*
1958 	 * there's a convention to display small numbered peer
1959 	 * in the left.
1960 	 */
1961 
1962 	if (lacp_compare_peerinfo(a, b) > 0) {
1963 		const struct lacp_peerinfo *t;
1964 
1965 		t = a;
1966 		a = b;
1967 		b = t;
1968 	}
1969 #endif
1970 
1971 	snprintf(buf, buflen, "[%s,%s]",
1972 	    lacp_format_partner(a, astr, sizeof(astr)),
1973 	    lacp_format_partner(b, bstr, sizeof(bstr)));
1974 
1975 	return (buf);
1976 }
1977 
1978 const char *
1979 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
1980     char *buf, size_t buflen)
1981 {
1982 	if (la == NULL) {
1983 		return ("(none)");
1984 	}
1985 
1986 	return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
1987 }
1988 
1989 const char *
1990 lacp_format_state(uint8_t state, char *buf, size_t buflen)
1991 {
1992 	snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
1993 	return (buf);
1994 }
1995 
1996 static void
1997 lacp_dump_lacpdu(const struct lacpdu *du)
1998 {
1999 	char buf[LACP_PARTNERSTR_MAX+1];
2000 	char buf2[LACP_STATESTR_MAX+1];
2001 
2002 	printf("actor=%s\n",
2003 	    lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2004 	printf("actor.state=%s\n",
2005 	    lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2006 	printf("partner=%s\n",
2007 	    lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2008 	printf("partner.state=%s\n",
2009 	    lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2010 
2011 	printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2012 }
2013 
2014 static void
2015 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2016 {
2017 	va_list va;
2018 
2019 	if (lp) {
2020 		printf("%s: ", lp->lp_ifp->if_xname);
2021 	}
2022 
2023 	va_start(va, fmt);
2024 	vprintf(fmt, va);
2025 	va_end(va);
2026 }
2027