sys/net/ieee8023ad_lacp.c

   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 "opt_ratelimit.h"
  34
  35 #include <sys/param.h>
  36 #include <sys/callout.h>
  37 #include <sys/eventhandler.h>
  38 #include <sys/mbuf.h>
  39 #include <sys/systm.h>
  40 #include <sys/malloc.h>
  41 #include <sys/kernel.h> /* hz */
  42 #include <sys/socket.h> /* for net/if.h */
  43 #include <sys/sockio.h>
  44 #include <sys/sysctl.h>
  45 #include <machine/stdarg.h>
  46 #include <sys/lock.h>
  47 #include <sys/rwlock.h>
  48 #include <sys/taskqueue.h>
  49
  50 #include <net/if.h>
  51 #include <net/if_var.h>
  52 #include <net/if_dl.h>
  53 #include <net/ethernet.h>
  54 #include <net/if_media.h>
  55 #include <net/if_types.h>
  56
  57 #include <net/if_lagg.h>
  58 #include <net/ieee8023ad_lacp.h>
  59
  60 /*
  61  * actor system priority and port priority.
  62  * XXX should be configurable.
  63  */
  64
  65 #define LACP_SYSTEM_PRIO        0x8000
  66 #define LACP_PORT_PRIO          0x8000
  67
  68 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  69     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  70
  71 static const struct tlv_template lacp_info_tlv_template[] = {
  72         { LACP_TYPE_ACTORINFO,
  73             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  74         { LACP_TYPE_PARTNERINFO,
  75             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  76         { LACP_TYPE_COLLECTORINFO,
  77             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  78         { 0, 0 },
  79 };
  80
  81 static const struct tlv_template marker_info_tlv_template[] = {
  82         { MARKER_TYPE_INFO,
  83             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  84         { 0, 0 },
  85 };
  86
  87 static const struct tlv_template marker_response_tlv_template[] = {
  88         { MARKER_TYPE_RESPONSE,
  89             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  90         { 0, 0 },
  91 };
  92
  93 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  94
  95 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
  96 static void     lacp_fill_markerinfo(struct lacp_port *,
  97                     struct lacp_markerinfo *);
  98
  99 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
 100 static void     lacp_suppress_distributing(struct lacp_softc *,
 101                     struct lacp_aggregator *);
 102 static void     lacp_transit_expire(void *);
 103 static void     lacp_update_portmap(struct lacp_softc *);
 104 static void     lacp_select_active_aggregator(struct lacp_softc *);
 105 static uint16_t lacp_compose_key(struct lacp_port *);
 106 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 107                     const struct tlv_template *, boolean_t);
 108 static void     lacp_tick(void *);
 109
 110 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 111                     const struct lacp_port *);
 112 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 113                     const struct lacp_peerinfo *);
 114 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 115                     const struct lacp_port *);
 116 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 117                     const struct lacp_peerinfo *);
 118
 119 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 120                     struct lacp_port *);
 121 static void     lacp_aggregator_addref(struct lacp_softc *,
 122                     struct lacp_aggregator *);
 123 static void     lacp_aggregator_delref(struct lacp_softc *,
 124                     struct lacp_aggregator *);
 125
 126 /* receive machine */
 127
 128 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 129 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 130 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 131 static void     lacp_sm_rx_timer(struct lacp_port *);
 132 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 133 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 134                     const struct lacpdu *);
 135 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 136                     const struct lacpdu *);
 137 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 138                     const struct lacpdu *);
 139 static void     lacp_sm_rx_record_default(struct lacp_port *);
 140 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 141 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 142                     const struct lacp_peerinfo *);
 143
 144 /* mux machine */
 145
 146 static void     lacp_sm_mux(struct lacp_port *);
 147 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 148 static void     lacp_sm_mux_timer(struct lacp_port *);
 149
 150 /* periodic transmit machine */
 151
 152 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 153 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 154 static void     lacp_sm_ptx_timer(struct lacp_port *);
 155
 156 /* transmit machine */
 157
 158 static void     lacp_sm_tx(struct lacp_port *);
 159 static void     lacp_sm_assert_ntt(struct lacp_port *);
 160
 161 static void     lacp_run_timers(struct lacp_port *);
 162 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 163                     const struct lacp_peerinfo *);
 164 static int      lacp_compare_systemid(const struct lacp_systemid *,
 165                     const struct lacp_systemid *);
 166 static void     lacp_port_enable(struct lacp_port *);
 167 static void     lacp_port_disable(struct lacp_port *);
 168 static void     lacp_select(struct lacp_port *);
 169 static void     lacp_unselect(struct lacp_port *);
 170 static void     lacp_disable_collecting(struct lacp_port *);
 171 static void     lacp_enable_collecting(struct lacp_port *);
 172 static void     lacp_disable_distributing(struct lacp_port *);
 173 static void     lacp_enable_distributing(struct lacp_port *);
 174 static int      lacp_xmit_lacpdu(struct lacp_port *);
 175 static int      lacp_xmit_marker(struct lacp_port *);
 176
 177 /* Debugging */
 178
 179 static void     lacp_dump_lacpdu(const struct lacpdu *);
 180 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 181                     size_t);
 182 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 183                     const struct lacp_peerinfo *, char *, size_t);
 184 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 185                     char *, size_t);
 186 static const char *lacp_format_state(uint8_t, char *, size_t);
 187 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 188 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 189                     size_t);
 190 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 191                     size_t);
 192 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 193                     __attribute__((__format__(__printf__, 2, 3)));
 194
 195 static VNET_DEFINE(int, lacp_debug);
 196 #define V_lacp_debug    VNET(lacp_debug)
 197 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
 198 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 199     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 200
 201 static VNET_DEFINE(int, lacp_default_strict_mode) = 1;
 202 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode, CTLFLAG_RWTUN,
 203     &VNET_NAME(lacp_default_strict_mode), 0,
 204     "LACP strict protocol compliance default");
 205
 206 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 207 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 208 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 209
 210 /*
 211  * partner administration variables.
 212  * XXX should be configurable.
 213  */
 214
 215 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 216         .lip_systemid = { .lsi_prio = 0xffff },
 217         .lip_portid = { .lpi_prio = 0xffff },
 218         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 219             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 220 };
 221
 222 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 223         .lip_systemid = { .lsi_prio = 0xffff },
 224         .lip_portid = { .lpi_prio = 0xffff },
 225         .lip_state = 0,
 226 };
 227
 228 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 229         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 230         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 231         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 232 };
 233
 234 struct mbuf *
 235 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 236 {
 237         struct lacp_port *lp = LACP_PORT(lgp);
 238         uint8_t subtype;
 239
 240         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 241                 m_freem(m);
 242                 return (NULL);
 243         }
 244
 245         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 246         switch (subtype) {
 247                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 248                         lacp_pdu_input(lp, m);
 249                         return (NULL);
 250
 251                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 252                         lacp_marker_input(lp, m);
 253                         return (NULL);
 254         }
 255
 256         /* Not a subtype we are interested in */
 257         return (m);
 258 }
 259
 260 /*
 261  * lacp_pdu_input: process lacpdu
 262  */
 263 static int
 264 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 265 {
 266         struct lacp_softc *lsc = lp->lp_lsc;
 267         struct lacpdu *du;
 268         int error = 0;
 269
 270         if (m->m_pkthdr.len != sizeof(*du)) {
 271                 goto bad;
 272         }
 273
 274         if ((m->m_flags & M_MCAST) == 0) {
 275                 goto bad;
 276         }
 277
 278         if (m->m_len < sizeof(*du)) {
 279                 m = m_pullup(m, sizeof(*du));
 280                 if (m == NULL) {
 281                         return (ENOMEM);
 282                 }
 283         }
 284
 285         du = mtod(m, struct lacpdu *);
 286
 287         if (memcmp(&du->ldu_eh.ether_dhost,
 288             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 289                 goto bad;
 290         }
 291
 292         /*
 293          * ignore the version for compatibility with
 294          * the future protocol revisions.
 295          */
 296 #if 0
 297         if (du->ldu_sph.sph_version != 1) {
 298                 goto bad;
 299         }
 300 #endif
 301
 302         /*
 303          * ignore tlv types for compatibility with
 304          * the future protocol revisions.
 305          */
 306         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 307             lacp_info_tlv_template, FALSE)) {
 308                 goto bad;
 309         }
 310
 311         if (V_lacp_debug > 0) {
 312                 lacp_dprintf(lp, "lacpdu receive\n");
 313                 lacp_dump_lacpdu(du);
 314         }
 315
 316         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 317                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 318                 goto bad;
 319         }
 320
 321         LACP_LOCK(lsc);
 322         lacp_sm_rx(lp, du);
 323         LACP_UNLOCK(lsc);
 324
 325         m_freem(m);
 326         return (error);
 327
 328 bad:
 329         m_freem(m);
 330         return (EINVAL);
 331 }
 332
 333 static void
 334 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 335 {
 336         struct lagg_port *lgp = lp->lp_lagg;
 337         struct lagg_softc *sc = lgp->lp_softc;
 338
 339         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 340         memcpy(&info->lip_systemid.lsi_mac,
 341             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 342         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 343         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 344         info->lip_state = lp->lp_state;
 345 }
 346
 347 static void
 348 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 349 {
 350         struct ifnet *ifp = lp->lp_ifp;
 351
 352         /* Fill in the port index and system id (encoded as the MAC) */
 353         info->mi_rq_port = htons(ifp->if_index);
 354         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 355         info->mi_rq_xid = htonl(0);
 356 }
 357
 358 static int
 359 lacp_xmit_lacpdu(struct lacp_port *lp)
 360 {
 361         struct lagg_port *lgp = lp->lp_lagg;
 362         struct mbuf *m;
 363         struct lacpdu *du;
 364         int error;
 365
 366         LACP_LOCK_ASSERT(lp->lp_lsc);
 367
 368         m = m_gethdr(M_NOWAIT, MT_DATA);
 369         if (m == NULL) {
 370                 return (ENOMEM);
 371         }
 372         m->m_len = m->m_pkthdr.len = sizeof(*du);
 373
 374         du = mtod(m, struct lacpdu *);
 375         memset(du, 0, sizeof(*du));
 376
 377         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 378             ETHER_ADDR_LEN);
 379         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 380         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 381
 382         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 383         du->ldu_sph.sph_version = 1;
 384
 385         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 386         du->ldu_actor = lp->lp_actor;
 387
 388         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 389             sizeof(du->ldu_partner));
 390         du->ldu_partner = lp->lp_partner;
 391
 392         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 393             sizeof(du->ldu_collector));
 394         du->ldu_collector.lci_maxdelay = 0;
 395
 396         if (V_lacp_debug > 0) {
 397                 lacp_dprintf(lp, "lacpdu transmit\n");
 398                 lacp_dump_lacpdu(du);
 399         }
 400
 401         m->m_flags |= M_MCAST;
 402
 403         /*
 404          * XXX should use higher priority queue.
 405          * otherwise network congestion can break aggregation.
 406          */
 407
 408         error = lagg_enqueue(lp->lp_ifp, m);
 409         return (error);
 410 }
 411
 412 static int
 413 lacp_xmit_marker(struct lacp_port *lp)
 414 {
 415         struct lagg_port *lgp = lp->lp_lagg;
 416         struct mbuf *m;
 417         struct markerdu *mdu;
 418         int error;
 419
 420         LACP_LOCK_ASSERT(lp->lp_lsc);
 421
 422         m = m_gethdr(M_NOWAIT, MT_DATA);
 423         if (m == NULL) {
 424                 return (ENOMEM);
 425         }
 426         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 427
 428         mdu = mtod(m, struct markerdu *);
 429         memset(mdu, 0, sizeof(*mdu));
 430
 431         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 432             ETHER_ADDR_LEN);
 433         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 434         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 435
 436         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 437         mdu->mdu_sph.sph_version = 1;
 438
 439         /* Bump the transaction id and copy over the marker info */
 440         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 441         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 442         mdu->mdu_info = lp->lp_marker;
 443
 444         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 445             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 446             ntohl(mdu->mdu_info.mi_rq_xid)));
 447
 448         m->m_flags |= M_MCAST;
 449         error = lagg_enqueue(lp->lp_ifp, m);
 450         return (error);
 451 }
 452
 453 void
 454 lacp_linkstate(struct lagg_port *lgp)
 455 {
 456         struct lacp_port *lp = LACP_PORT(lgp);
 457         struct lacp_softc *lsc = lp->lp_lsc;
 458         struct ifnet *ifp = lgp->lp_ifp;
 459         struct ifmediareq ifmr;
 460         int error = 0;
 461         u_int media;
 462         uint8_t old_state;
 463         uint16_t old_key;
 464
 465         bzero((char *)&ifmr, sizeof(ifmr));
 466         error = (*ifp->if_ioctl)(ifp, SIOCGIFXMEDIA, (caddr_t)&ifmr);
 467         if (error != 0) {
 468                 bzero((char *)&ifmr, sizeof(ifmr));
 469                 error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 470         }
 471         if (error != 0)
 472                 return;
 473
 474         LACP_LOCK(lsc);
 475         media = ifmr.ifm_active;
 476         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 477             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 478             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 479         old_state = lp->lp_state;
 480         old_key = lp->lp_key;
 481
 482         lp->lp_media = media;
 483         /*
 484          * If the port is not an active full duplex Ethernet link then it can
 485          * not be aggregated.
 486          */
 487         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 488             ifp->if_link_state != LINK_STATE_UP) {
 489                 lacp_port_disable(lp);
 490         } else {
 491                 lacp_port_enable(lp);
 492         }
 493         lp->lp_key = lacp_compose_key(lp);
 494
 495         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 496                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 497                 lp->lp_selected = LACP_UNSELECTED;
 498         }
 499         LACP_UNLOCK(lsc);
 500 }
 501
 502 static void
 503 lacp_tick(void *arg)
 504 {
 505         struct lacp_softc *lsc = arg;
 506         struct lacp_port *lp;
 507
 508         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 509                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 510                         continue;
 511
 512                 CURVNET_SET(lp->lp_ifp->if_vnet);
 513                 lacp_run_timers(lp);
 514
 515                 lacp_select(lp);
 516                 lacp_sm_mux(lp);
 517                 lacp_sm_tx(lp);
 518                 lacp_sm_ptx_tx_schedule(lp);
 519                 CURVNET_RESTORE();
 520         }
 521         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 522 }
 523
 524 int
 525 lacp_port_create(struct lagg_port *lgp)
 526 {
 527         struct lagg_softc *sc = lgp->lp_softc;
 528         struct lacp_softc *lsc = LACP_SOFTC(sc);
 529         struct lacp_port *lp;
 530         struct ifnet *ifp = lgp->lp_ifp;
 531         struct sockaddr_dl sdl;
 532         struct ifmultiaddr *rifma = NULL;
 533         int error;
 534
 535         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 536         sdl.sdl_alen = ETHER_ADDR_LEN;
 537
 538         bcopy(&ethermulticastaddr_slowprotocols,
 539             LLADDR(&sdl), ETHER_ADDR_LEN);
 540         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 541         if (error) {
 542                 printf("%s: ADDMULTI failed on %s\n", __func__,
 543                     lgp->lp_ifp->if_xname);
 544                 return (error);
 545         }
 546
 547         lp = malloc(sizeof(struct lacp_port),
 548             M_DEVBUF, M_NOWAIT|M_ZERO);
 549         if (lp == NULL)
 550                 return (ENOMEM);
 551
 552         LACP_LOCK(lsc);
 553         lgp->lp_psc = lp;
 554         lp->lp_ifp = ifp;
 555         lp->lp_lagg = lgp;
 556         lp->lp_lsc = lsc;
 557         lp->lp_ifma = rifma;
 558
 559         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 560
 561         lacp_fill_actorinfo(lp, &lp->lp_actor);
 562         lacp_fill_markerinfo(lp, &lp->lp_marker);
 563         lp->lp_state = LACP_STATE_ACTIVITY;
 564         lp->lp_aggregator = NULL;
 565         lacp_sm_rx_set_expired(lp);
 566         LACP_UNLOCK(lsc);
 567         lacp_linkstate(lgp);
 568
 569         return (0);
 570 }
 571
 572 void
 573 lacp_port_destroy(struct lagg_port *lgp)
 574 {
 575         struct lacp_port *lp = LACP_PORT(lgp);
 576         struct lacp_softc *lsc = lp->lp_lsc;
 577         int i;
 578
 579         LACP_LOCK(lsc);
 580         for (i = 0; i < LACP_NTIMER; i++) {
 581                 LACP_TIMER_DISARM(lp, i);
 582         }
 583
 584         lacp_disable_collecting(lp);
 585         lacp_disable_distributing(lp);
 586         lacp_unselect(lp);
 587
 588         LIST_REMOVE(lp, lp_next);
 589         LACP_UNLOCK(lsc);
 590
 591         /* The address may have already been removed by if_purgemaddrs() */
 592         if (!lgp->lp_detaching)
 593                 if_delmulti_ifma(lp->lp_ifma);
 594
 595         free(lp, M_DEVBUF);
 596 }
 597
 598 void
 599 lacp_req(struct lagg_softc *sc, void *data)
 600 {
 601         struct lacp_opreq *req = (struct lacp_opreq *)data;
 602         struct lacp_softc *lsc = LACP_SOFTC(sc);
 603         struct lacp_aggregator *la;
 604
 605         bzero(req, sizeof(struct lacp_opreq));
 606
 607         /*
 608          * If the LACP softc is NULL, return with the opreq structure full of
 609          * zeros.  It is normal for the softc to be NULL while the lagg is
 610          * being destroyed.
 611          */
 612         if (NULL == lsc)
 613                 return;
 614
 615         la = lsc->lsc_active_aggregator;
 616         LACP_LOCK(lsc);
 617         if (la != NULL) {
 618                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 619                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 620                     ETHER_ADDR_LEN);
 621                 req->actor_key = ntohs(la->la_actor.lip_key);
 622                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 623                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 624                 req->actor_state = la->la_actor.lip_state;
 625
 626                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 627                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 628                     ETHER_ADDR_LEN);
 629                 req->partner_key = ntohs(la->la_partner.lip_key);
 630                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 631                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 632                 req->partner_state = la->la_partner.lip_state;
 633         }
 634         LACP_UNLOCK(lsc);
 635 }
 636
 637 void
 638 lacp_portreq(struct lagg_port *lgp, void *data)
 639 {
 640         struct lacp_opreq *req = (struct lacp_opreq *)data;
 641         struct lacp_port *lp = LACP_PORT(lgp);
 642         struct lacp_softc *lsc = lp->lp_lsc;
 643
 644         LACP_LOCK(lsc);
 645         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 646         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 647             ETHER_ADDR_LEN);
 648         req->actor_key = ntohs(lp->lp_actor.lip_key);
 649         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 650         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 651         req->actor_state = lp->lp_actor.lip_state;
 652
 653         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 654         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 655             ETHER_ADDR_LEN);
 656         req->partner_key = ntohs(lp->lp_partner.lip_key);
 657         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 658         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 659         req->partner_state = lp->lp_partner.lip_state;
 660         LACP_UNLOCK(lsc);
 661 }
 662
 663 static void
 664 lacp_disable_collecting(struct lacp_port *lp)
 665 {
 666         LACP_DPRINTF((lp, "collecting disabled\n"));
 667         lp->lp_state &= ~LACP_STATE_COLLECTING;
 668 }
 669
 670 static void
 671 lacp_enable_collecting(struct lacp_port *lp)
 672 {
 673         LACP_DPRINTF((lp, "collecting enabled\n"));
 674         lp->lp_state |= LACP_STATE_COLLECTING;
 675 }
 676
 677 static void
 678 lacp_disable_distributing(struct lacp_port *lp)
 679 {
 680         struct lacp_aggregator *la = lp->lp_aggregator;
 681         struct lacp_softc *lsc = lp->lp_lsc;
 682         struct lagg_softc *sc = lsc->lsc_softc;
 683         char buf[LACP_LAGIDSTR_MAX+1];
 684
 685         LACP_LOCK_ASSERT(lsc);
 686
 687         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 688                 return;
 689         }
 690
 691         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 692         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 693         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 694
 695         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 696             "nports %d -> %d\n",
 697             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 698             la->la_nports, la->la_nports - 1));
 699
 700         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 701         la->la_nports--;
 702         sc->sc_active = la->la_nports;
 703
 704         if (lsc->lsc_active_aggregator == la) {
 705                 lacp_suppress_distributing(lsc, la);
 706                 lacp_select_active_aggregator(lsc);
 707                 /* regenerate the port map, the active aggregator has changed */
 708                 lacp_update_portmap(lsc);
 709         }
 710
 711         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 712 }
 713
 714 static void
 715 lacp_enable_distributing(struct lacp_port *lp)
 716 {
 717         struct lacp_aggregator *la = lp->lp_aggregator;
 718         struct lacp_softc *lsc = lp->lp_lsc;
 719         struct lagg_softc *sc = lsc->lsc_softc;
 720         char buf[LACP_LAGIDSTR_MAX+1];
 721
 722         LACP_LOCK_ASSERT(lsc);
 723
 724         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 725                 return;
 726         }
 727
 728         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 729             "nports %d -> %d\n",
 730             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 731             la->la_nports, la->la_nports + 1));
 732
 733         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 734         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 735         la->la_nports++;
 736         sc->sc_active = la->la_nports;
 737
 738         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 739
 740         if (lsc->lsc_active_aggregator == la) {
 741                 lacp_suppress_distributing(lsc, la);
 742                 lacp_update_portmap(lsc);
 743         } else
 744                 /* try to become the active aggregator */
 745                 lacp_select_active_aggregator(lsc);
 746 }
 747
 748 static void
 749 lacp_transit_expire(void *vp)
 750 {
 751         struct lacp_softc *lsc = vp;
 752
 753         LACP_LOCK_ASSERT(lsc);
 754
 755         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 756         LACP_TRACE(NULL);
 757         CURVNET_RESTORE();
 758
 759         lsc->lsc_suppress_distributing = FALSE;
 760 }
 761
 762 void
 763 lacp_attach(struct lagg_softc *sc)
 764 {
 765         struct lacp_softc *lsc;
 766
 767         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 768
 769         sc->sc_psc = lsc;
 770         lsc->lsc_softc = sc;
 771
 772         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 773         lsc->lsc_active_aggregator = NULL;
 774         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 775         LACP_LOCK_INIT(lsc);
 776         TAILQ_INIT(&lsc->lsc_aggregators);
 777         LIST_INIT(&lsc->lsc_ports);
 778
 779         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 780         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 781
 782         /* if the lagg is already up then do the same */
 783         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 784                 lacp_init(sc);
 785 }
 786
 787 void
 788 lacp_detach(void *psc)
 789 {
 790         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 791
 792         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 793             ("aggregators still active"));
 794         KASSERT(lsc->lsc_active_aggregator == NULL,
 795             ("aggregator still attached"));
 796
 797         callout_drain(&lsc->lsc_transit_callout);
 798         callout_drain(&lsc->lsc_callout);
 799
 800         LACP_LOCK_DESTROY(lsc);
 801         free(lsc, M_DEVBUF);
 802 }
 803
 804 void
 805 lacp_init(struct lagg_softc *sc)
 806 {
 807         struct lacp_softc *lsc = LACP_SOFTC(sc);
 808
 809         LACP_LOCK(lsc);
 810         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 811         LACP_UNLOCK(lsc);
 812 }
 813
 814 void
 815 lacp_stop(struct lagg_softc *sc)
 816 {
 817         struct lacp_softc *lsc = LACP_SOFTC(sc);
 818
 819         LACP_LOCK(lsc);
 820         callout_stop(&lsc->lsc_transit_callout);
 821         callout_stop(&lsc->lsc_callout);
 822         LACP_UNLOCK(lsc);
 823 }
 824
 825 struct lagg_port *
 826 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 827 {
 828         struct lacp_softc *lsc = LACP_SOFTC(sc);
 829         struct lacp_portmap *pm;
 830         struct lacp_port *lp;
 831         uint32_t hash;
 832
 833         if (__predict_false(lsc->lsc_suppress_distributing)) {
 834                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 835                 return (NULL);
 836         }
 837
 838         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 839         if (pm->pm_count == 0) {
 840                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 841                 return (NULL);
 842         }
 843
 844         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 845             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 846                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 847         else
 848                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 849         hash %= pm->pm_count;
 850         lp = pm->pm_map[hash];
 851
 852         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 853             ("aggregated port is not distributing"));
 854
 855         return (lp->lp_lagg);
 856 }
 857
 858 #ifdef RATELIMIT
 859 struct lagg_port *
 860 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t flowid)
 861 {
 862         struct lacp_softc *lsc = LACP_SOFTC(sc);
 863         struct lacp_portmap *pm;
 864         struct lacp_port *lp;
 865         uint32_t hash;
 866
 867         if (__predict_false(lsc->lsc_suppress_distributing)) {
 868                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 869                 return (NULL);
 870         }
 871
 872         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 873         if (pm->pm_count == 0) {
 874                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 875                 return (NULL);
 876         }
 877
 878         hash = flowid >> sc->flowid_shift;
 879         hash %= pm->pm_count;
 880         lp = pm->pm_map[hash];
 881
 882         return (lp->lp_lagg);
 883 }
 884 #endif
 885
 886 /*
 887  * lacp_suppress_distributing: drop transmit packets for a while
 888  * to preserve packet ordering.
 889  */
 890
 891 static void
 892 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 893 {
 894         struct lacp_port *lp;
 895
 896         if (lsc->lsc_active_aggregator != la) {
 897                 return;
 898         }
 899
 900         LACP_TRACE(NULL);
 901
 902         lsc->lsc_suppress_distributing = TRUE;
 903
 904         /* send a marker frame down each port to verify the queues are empty */
 905         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 906                 lp->lp_flags |= LACP_PORT_MARK;
 907                 lacp_xmit_marker(lp);
 908         }
 909
 910         /* set a timeout for the marker frames */
 911         callout_reset(&lsc->lsc_transit_callout,
 912             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 913 }
 914
 915 static int
 916 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 917     const struct lacp_peerinfo *b)
 918 {
 919         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 920 }
 921
 922 static int
 923 lacp_compare_systemid(const struct lacp_systemid *a,
 924     const struct lacp_systemid *b)
 925 {
 926         return (memcmp(a, b, sizeof(*a)));
 927 }
 928
 929 #if 0   /* unused */
 930 static int
 931 lacp_compare_portid(const struct lacp_portid *a,
 932     const struct lacp_portid *b)
 933 {
 934         return (memcmp(a, b, sizeof(*a)));
 935 }
 936 #endif
 937
 938 static uint64_t
 939 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 940 {
 941         struct lacp_port *lp;
 942         uint64_t speed;
 943
 944         lp = TAILQ_FIRST(&la->la_ports);
 945         if (lp == NULL) {
 946                 return (0);
 947         }
 948
 949         speed = ifmedia_baudrate(lp->lp_media);
 950         speed *= la->la_nports;
 951         if (speed == 0) {
 952                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 953                     lp->lp_media, la->la_nports));
 954         }
 955
 956         return (speed);
 957 }
 958
 959 /*
 960  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 961  * packets from lagg(4) interface.
 962  */
 963
 964 static void
 965 lacp_select_active_aggregator(struct lacp_softc *lsc)
 966 {
 967         struct lacp_aggregator *la;
 968         struct lacp_aggregator *best_la = NULL;
 969         uint64_t best_speed = 0;
 970         char buf[LACP_LAGIDSTR_MAX+1];
 971
 972         LACP_TRACE(NULL);
 973
 974         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 975                 uint64_t speed;
 976
 977                 if (la->la_nports == 0) {
 978                         continue;
 979                 }
 980
 981                 speed = lacp_aggregator_bandwidth(la);
 982                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 983                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 984                     speed, la->la_nports));
 985
 986                 /*
 987                  * This aggregator is chosen if the partner has a better
 988                  * system priority or, the total aggregated speed is higher
 989                  * or, it is already the chosen aggregator
 990                  */
 991                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 992                     LACP_SYS_PRI(best_la->la_partner)) ||
 993                     speed > best_speed ||
 994                     (speed == best_speed &&
 995                     la == lsc->lsc_active_aggregator)) {
 996                         best_la = la;
 997                         best_speed = speed;
 998                 }
 999         }
1000
1001         KASSERT(best_la == NULL || best_la->la_nports > 0,
1002             ("invalid aggregator refcnt"));
1003         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1004             ("invalid aggregator list"));
1005
1006         if (lsc->lsc_active_aggregator != best_la) {
1007                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1008                 LACP_DPRINTF((NULL, "old %s\n",
1009                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1010                     buf, sizeof(buf))));
1011         } else {
1012                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1013         }
1014         LACP_DPRINTF((NULL, "new %s\n",
1015             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1016
1017         if (lsc->lsc_active_aggregator != best_la) {
1018                 lsc->lsc_active_aggregator = best_la;
1019                 lacp_update_portmap(lsc);
1020                 if (best_la) {
1021                         lacp_suppress_distributing(lsc, best_la);
1022                 }
1023         }
1024 }
1025
1026 /*
1027  * Updated the inactive portmap array with the new list of ports and
1028  * make it live.
1029  */
1030 static void
1031 lacp_update_portmap(struct lacp_softc *lsc)
1032 {
1033         struct lagg_softc *sc = lsc->lsc_softc;
1034         struct lacp_aggregator *la;
1035         struct lacp_portmap *p;
1036         struct lacp_port *lp;
1037         uint64_t speed;
1038         u_int newmap;
1039         int i;
1040
1041         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1042         p = &lsc->lsc_pmap[newmap];
1043         la = lsc->lsc_active_aggregator;
1044         speed = 0;
1045         bzero(p, sizeof(struct lacp_portmap));
1046
1047         if (la != NULL && la->la_nports > 0) {
1048                 p->pm_count = la->la_nports;
1049                 i = 0;
1050                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1051                         p->pm_map[i++] = lp;
1052                 KASSERT(i == p->pm_count, ("Invalid port count"));
1053                 speed = lacp_aggregator_bandwidth(la);
1054         }
1055         sc->sc_ifp->if_baudrate = speed;
1056
1057         /* switch the active portmap over */
1058         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1059         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1060                     lsc->lsc_activemap,
1061                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1062 }
1063
1064 static uint16_t
1065 lacp_compose_key(struct lacp_port *lp)
1066 {
1067         struct lagg_port *lgp = lp->lp_lagg;
1068         struct lagg_softc *sc = lgp->lp_softc;
1069         u_int media = lp->lp_media;
1070         uint16_t key;
1071
1072         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1073
1074                 /*
1075                  * non-aggregatable links should have unique keys.
1076                  *
1077                  * XXX this isn't really unique as if_index is 16 bit.
1078                  */
1079
1080                 /* bit 0..14:   (some bits of) if_index of this port */
1081                 key = lp->lp_ifp->if_index;
1082                 /* bit 15:      1 */
1083                 key |= 0x8000;
1084         } else {
1085                 u_int subtype = IFM_SUBTYPE(media);
1086
1087                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1088                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1089
1090                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1091                 switch (subtype) {
1092                 case IFM_10_T:
1093                 case IFM_10_2:
1094                 case IFM_10_5:
1095                 case IFM_10_STP:
1096                 case IFM_10_FL:
1097                         key = IFM_10_T;
1098                         break;
1099                 case IFM_100_TX:
1100                 case IFM_100_FX:
1101                 case IFM_100_T4:
1102                 case IFM_100_VG:
1103                 case IFM_100_T2:
1104                 case IFM_100_T:
1105                         key = IFM_100_TX;
1106                         break;
1107                 case IFM_1000_SX:
1108                 case IFM_1000_LX:
1109                 case IFM_1000_CX:
1110                 case IFM_1000_T:
1111                 case IFM_1000_KX:
1112                 case IFM_1000_SGMII:
1113                 case IFM_1000_CX_SGMII:
1114                         key = IFM_1000_SX;
1115                         break;
1116                 case IFM_10G_LR:
1117                 case IFM_10G_SR:
1118                 case IFM_10G_CX4:
1119                 case IFM_10G_TWINAX:
1120                 case IFM_10G_TWINAX_LONG:
1121                 case IFM_10G_LRM:
1122                 case IFM_10G_T:
1123                 case IFM_10G_KX4:
1124                 case IFM_10G_KR:
1125                 case IFM_10G_CR1:
1126                 case IFM_10G_ER:
1127                 case IFM_10G_SFI:
1128                 case IFM_10G_AOC:
1129                         key = IFM_10G_LR;
1130                         break;
1131                 case IFM_20G_KR2:
1132                         key = IFM_20G_KR2;
1133                         break;
1134                 case IFM_2500_KX:
1135                 case IFM_2500_T:
1136                         key = IFM_2500_KX;
1137                         break;
1138                 case IFM_5000_T:
1139                         key = IFM_5000_T;
1140                         break;
1141                 case IFM_50G_PCIE:
1142                 case IFM_50G_CR2:
1143                 case IFM_50G_KR2:
1144                         key = IFM_50G_PCIE;
1145                         break;
1146                 case IFM_56G_R4:
1147                         key = IFM_56G_R4;
1148                         break;
1149                 case IFM_25G_PCIE:
1150                 case IFM_25G_CR:
1151                 case IFM_25G_KR:
1152                 case IFM_25G_SR:
1153                 case IFM_25G_LR:
1154                 case IFM_25G_ACC:
1155                 case IFM_25G_AOC:
1156                         key = IFM_25G_PCIE;
1157                         break;
1158                 case IFM_40G_CR4:
1159                 case IFM_40G_SR4:
1160                 case IFM_40G_LR4:
1161                 case IFM_40G_XLPPI:
1162                 case IFM_40G_KR4:
1163                         key = IFM_40G_CR4;
1164                         break;
1165                 case IFM_100G_CR4:
1166                 case IFM_100G_SR4:
1167                 case IFM_100G_KR4:
1168                 case IFM_100G_LR4:
1169                         key = IFM_100G_CR4;
1170                         break;
1171                 default:
1172                         key = subtype;
1173                         break;
1174                 }
1175                 /* bit 5..14:   (some bits of) if_index of lagg device */
1176                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1177                 /* bit 15:      0 */
1178         }
1179         return (htons(key));
1180 }
1181
1182 static void
1183 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1184 {
1185         char buf[LACP_LAGIDSTR_MAX+1];
1186
1187         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1188             __func__,
1189             lacp_format_lagid(&la->la_actor, &la->la_partner,
1190             buf, sizeof(buf)),
1191             la->la_refcnt, la->la_refcnt + 1));
1192
1193         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1194         la->la_refcnt++;
1195         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1196 }
1197
1198 static void
1199 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1200 {
1201         char buf[LACP_LAGIDSTR_MAX+1];
1202
1203         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1204             __func__,
1205             lacp_format_lagid(&la->la_actor, &la->la_partner,
1206             buf, sizeof(buf)),
1207             la->la_refcnt, la->la_refcnt - 1));
1208
1209         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1210         la->la_refcnt--;
1211         if (la->la_refcnt > 0) {
1212                 return;
1213         }
1214
1215         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1216         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1217
1218         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1219
1220         free(la, M_DEVBUF);
1221 }
1222
1223 /*
1224  * lacp_aggregator_get: allocate an aggregator.
1225  */
1226
1227 static struct lacp_aggregator *
1228 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1229 {
1230         struct lacp_aggregator *la;
1231
1232         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1233         if (la) {
1234                 la->la_refcnt = 1;
1235                 la->la_nports = 0;
1236                 TAILQ_INIT(&la->la_ports);
1237                 la->la_pending = 0;
1238                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1239         }
1240
1241         return (la);
1242 }
1243
1244 /*
1245  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1246  */
1247
1248 static void
1249 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1250 {
1251         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1252         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1253
1254         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1255 }
1256
1257 static void
1258 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1259     const struct lacp_peerinfo *lpi_port)
1260 {
1261         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1262         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1263         lpi_aggr->lip_key = lpi_port->lip_key;
1264 }
1265
1266 /*
1267  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1268  */
1269
1270 static int
1271 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1272     const struct lacp_port *lp)
1273 {
1274         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1275             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1276                 return (0);
1277         }
1278
1279         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1280                 return (0);
1281         }
1282
1283         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1284                 return (0);
1285         }
1286
1287         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1288                 return (0);
1289         }
1290
1291         return (1);
1292 }
1293
1294 static int
1295 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1296     const struct lacp_peerinfo *b)
1297 {
1298         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1299             sizeof(a->lip_systemid))) {
1300                 return (0);
1301         }
1302
1303         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1304                 return (0);
1305         }
1306
1307         return (1);
1308 }
1309
1310 static void
1311 lacp_port_enable(struct lacp_port *lp)
1312 {
1313         lp->lp_state |= LACP_STATE_AGGREGATION;
1314 }
1315
1316 static void
1317 lacp_port_disable(struct lacp_port *lp)
1318 {
1319         lacp_set_mux(lp, LACP_MUX_DETACHED);
1320
1321         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1322         lp->lp_selected = LACP_UNSELECTED;
1323         lacp_sm_rx_record_default(lp);
1324         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1325         lp->lp_state &= ~LACP_STATE_EXPIRED;
1326 }
1327
1328 /*
1329  * lacp_select: select an aggregator.  create one if necessary.
1330  */
1331 static void
1332 lacp_select(struct lacp_port *lp)
1333 {
1334         struct lacp_softc *lsc = lp->lp_lsc;
1335         struct lacp_aggregator *la;
1336         char buf[LACP_LAGIDSTR_MAX+1];
1337
1338         if (lp->lp_aggregator) {
1339                 return;
1340         }
1341
1342         /* If we haven't heard from our peer, skip this step. */
1343         if (lp->lp_state & LACP_STATE_DEFAULTED)
1344                 return;
1345
1346         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1347             ("timer_wait_while still active"));
1348
1349         LACP_DPRINTF((lp, "port lagid=%s\n",
1350             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1351             buf, sizeof(buf))));
1352
1353         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1354                 if (lacp_aggregator_is_compatible(la, lp)) {
1355                         break;
1356                 }
1357         }
1358
1359         if (la == NULL) {
1360                 la = lacp_aggregator_get(lsc, lp);
1361                 if (la == NULL) {
1362                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1363
1364                         /*
1365                          * will retry on the next tick.
1366                          */
1367
1368                         return;
1369                 }
1370                 lacp_fill_aggregator_id(la, lp);
1371                 LACP_DPRINTF((lp, "aggregator created\n"));
1372         } else {
1373                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1374                 if (la->la_refcnt == LACP_MAX_PORTS)
1375                         return;
1376                 lacp_aggregator_addref(lsc, la);
1377         }
1378
1379         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1380             lacp_format_lagid(&la->la_actor, &la->la_partner,
1381             buf, sizeof(buf))));
1382
1383         lp->lp_aggregator = la;
1384         lp->lp_selected = LACP_SELECTED;
1385 }
1386
1387 /*
1388  * lacp_unselect: finish unselect/detach process.
1389  */
1390
1391 static void
1392 lacp_unselect(struct lacp_port *lp)
1393 {
1394         struct lacp_softc *lsc = lp->lp_lsc;
1395         struct lacp_aggregator *la = lp->lp_aggregator;
1396
1397         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1398             ("timer_wait_while still active"));
1399
1400         if (la == NULL) {
1401                 return;
1402         }
1403
1404         lp->lp_aggregator = NULL;
1405         lacp_aggregator_delref(lsc, la);
1406 }
1407
1408 /* mux machine */
1409
1410 static void
1411 lacp_sm_mux(struct lacp_port *lp)
1412 {
1413         struct lagg_port *lgp = lp->lp_lagg;
1414         struct lagg_softc *sc = lgp->lp_softc;
1415         enum lacp_mux_state new_state;
1416         boolean_t p_sync =
1417                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1418         boolean_t p_collecting =
1419             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1420         enum lacp_selected selected = lp->lp_selected;
1421         struct lacp_aggregator *la;
1422
1423         if (V_lacp_debug > 1)
1424                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1425                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1426                     lp->lp_mux_state, selected, p_sync, p_collecting);
1427
1428 re_eval:
1429         la = lp->lp_aggregator;
1430         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1431             ("MUX not detached"));
1432         new_state = lp->lp_mux_state;
1433         switch (lp->lp_mux_state) {
1434         case LACP_MUX_DETACHED:
1435                 if (selected != LACP_UNSELECTED) {
1436                         new_state = LACP_MUX_WAITING;
1437                 }
1438                 break;
1439         case LACP_MUX_WAITING:
1440                 KASSERT(la->la_pending > 0 ||
1441                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1442                     ("timer_wait_while still active"));
1443                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1444                         new_state = LACP_MUX_ATTACHED;
1445                 } else if (selected == LACP_UNSELECTED) {
1446                         new_state = LACP_MUX_DETACHED;
1447                 }
1448                 break;
1449         case LACP_MUX_ATTACHED:
1450                 if (selected == LACP_SELECTED && p_sync) {
1451                         new_state = LACP_MUX_COLLECTING;
1452                 } else if (selected != LACP_SELECTED) {
1453                         new_state = LACP_MUX_DETACHED;
1454                 }
1455                 break;
1456         case LACP_MUX_COLLECTING:
1457                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1458                         new_state = LACP_MUX_DISTRIBUTING;
1459                 } else if (selected != LACP_SELECTED || !p_sync) {
1460                         new_state = LACP_MUX_ATTACHED;
1461                 }
1462                 break;
1463         case LACP_MUX_DISTRIBUTING:
1464                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1465                         new_state = LACP_MUX_COLLECTING;
1466                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1467                         sc->sc_flapping++;
1468                 }
1469                 break;
1470         default:
1471                 panic("%s: unknown state", __func__);
1472         }
1473
1474         if (lp->lp_mux_state == new_state) {
1475                 return;
1476         }
1477
1478         lacp_set_mux(lp, new_state);
1479         goto re_eval;
1480 }
1481
1482 static void
1483 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1484 {
1485         struct lacp_aggregator *la = lp->lp_aggregator;
1486
1487         if (lp->lp_mux_state == new_state) {
1488                 return;
1489         }
1490
1491         switch (new_state) {
1492         case LACP_MUX_DETACHED:
1493                 lp->lp_state &= ~LACP_STATE_SYNC;
1494                 lacp_disable_distributing(lp);
1495                 lacp_disable_collecting(lp);
1496                 lacp_sm_assert_ntt(lp);
1497                 /* cancel timer */
1498                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1499                         KASSERT(la->la_pending > 0,
1500                             ("timer_wait_while not active"));
1501                         la->la_pending--;
1502                 }
1503                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1504                 lacp_unselect(lp);
1505                 break;
1506         case LACP_MUX_WAITING:
1507                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1508                     LACP_AGGREGATE_WAIT_TIME);
1509                 la->la_pending++;
1510                 break;
1511         case LACP_MUX_ATTACHED:
1512                 lp->lp_state |= LACP_STATE_SYNC;
1513                 lacp_disable_collecting(lp);
1514                 lacp_sm_assert_ntt(lp);
1515                 break;
1516         case LACP_MUX_COLLECTING:
1517                 lacp_enable_collecting(lp);
1518                 lacp_disable_distributing(lp);
1519                 lacp_sm_assert_ntt(lp);
1520                 break;
1521         case LACP_MUX_DISTRIBUTING:
1522                 lacp_enable_distributing(lp);
1523                 break;
1524         default:
1525                 panic("%s: unknown state", __func__);
1526         }
1527
1528         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1529
1530         lp->lp_mux_state = new_state;
1531 }
1532
1533 static void
1534 lacp_sm_mux_timer(struct lacp_port *lp)
1535 {
1536         struct lacp_aggregator *la = lp->lp_aggregator;
1537         char buf[LACP_LAGIDSTR_MAX+1];
1538
1539         KASSERT(la->la_pending > 0, ("no pending event"));
1540
1541         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1542             lacp_format_lagid(&la->la_actor, &la->la_partner,
1543             buf, sizeof(buf)),
1544             la->la_pending, la->la_pending - 1));
1545
1546         la->la_pending--;
1547 }
1548
1549 /* periodic transmit machine */
1550
1551 static void
1552 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1553 {
1554         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1555             LACP_STATE_TIMEOUT)) {
1556                 return;
1557         }
1558
1559         LACP_DPRINTF((lp, "partner timeout changed\n"));
1560
1561         /*
1562          * FAST_PERIODIC -> SLOW_PERIODIC
1563          * or
1564          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1565          *
1566          * let lacp_sm_ptx_tx_schedule to update timeout.
1567          */
1568
1569         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1570
1571         /*
1572          * if timeout has been shortened, assert NTT.
1573          */
1574
1575         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1576                 lacp_sm_assert_ntt(lp);
1577         }
1578 }
1579
1580 static void
1581 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1582 {
1583         int timeout;
1584
1585         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1586             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1587
1588                 /*
1589                  * NO_PERIODIC
1590                  */
1591
1592                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1593                 return;
1594         }
1595
1596         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1597                 return;
1598         }
1599
1600         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1601             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1602
1603         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1604 }
1605
1606 static void
1607 lacp_sm_ptx_timer(struct lacp_port *lp)
1608 {
1609         lacp_sm_assert_ntt(lp);
1610 }
1611
1612 static void
1613 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1614 {
1615         int timeout;
1616
1617         /*
1618          * check LACP_DISABLED first
1619          */
1620
1621         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1622                 return;
1623         }
1624
1625         /*
1626          * check loopback condition.
1627          */
1628
1629         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1630             &lp->lp_actor.lip_systemid)) {
1631                 return;
1632         }
1633
1634         /*
1635          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1636          */
1637
1638         lacp_sm_rx_update_selected(lp, du);
1639         lacp_sm_rx_update_ntt(lp, du);
1640         lacp_sm_rx_record_pdu(lp, du);
1641
1642         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1643             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1644         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1645
1646         lp->lp_state &= ~LACP_STATE_EXPIRED;
1647
1648         /*
1649          * kick transmit machine without waiting the next tick.
1650          */
1651
1652         lacp_sm_tx(lp);
1653 }
1654
1655 static void
1656 lacp_sm_rx_set_expired(struct lacp_port *lp)
1657 {
1658         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1659         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1660         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1661         lp->lp_state |= LACP_STATE_EXPIRED;
1662 }
1663
1664 static void
1665 lacp_sm_rx_timer(struct lacp_port *lp)
1666 {
1667         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1668                 /* CURRENT -> EXPIRED */
1669                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1670                 lacp_sm_rx_set_expired(lp);
1671         } else {
1672                 /* EXPIRED -> DEFAULTED */
1673                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1674                 lacp_sm_rx_update_default_selected(lp);
1675                 lacp_sm_rx_record_default(lp);
1676                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1677         }
1678 }
1679
1680 static void
1681 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1682 {
1683         boolean_t active;
1684         uint8_t oldpstate;
1685         char buf[LACP_STATESTR_MAX+1];
1686
1687         LACP_TRACE(lp);
1688
1689         oldpstate = lp->lp_partner.lip_state;
1690
1691         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1692             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1693             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1694
1695         lp->lp_partner = du->ldu_actor;
1696         if (active &&
1697             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1698             LACP_STATE_AGGREGATION) &&
1699             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1700             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1701                 /*
1702                  * XXX Maintain legacy behavior of leaving the
1703                  * LACP_STATE_SYNC bit unchanged from the partner's
1704                  * advertisement if lsc_strict_mode is false.
1705                  * TODO: We should re-examine the concept of the "strict mode"
1706                  * to ensure it makes sense to maintain a non-strict mode.
1707                  */
1708                 if (lp->lp_lsc->lsc_strict_mode)
1709                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1710         } else {
1711                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1712         }
1713
1714         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1715
1716         if (oldpstate != lp->lp_partner.lip_state) {
1717                 LACP_DPRINTF((lp, "old pstate %s\n",
1718                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1719                 LACP_DPRINTF((lp, "new pstate %s\n",
1720                     lacp_format_state(lp->lp_partner.lip_state, buf,
1721                     sizeof(buf))));
1722         }
1723
1724         lacp_sm_ptx_update_timeout(lp, oldpstate);
1725 }
1726
1727 static void
1728 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1729 {
1730
1731         LACP_TRACE(lp);
1732
1733         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1734             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1735             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1736                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1737                 lacp_sm_assert_ntt(lp);
1738         }
1739 }
1740
1741 static void
1742 lacp_sm_rx_record_default(struct lacp_port *lp)
1743 {
1744         uint8_t oldpstate;
1745
1746         LACP_TRACE(lp);
1747
1748         oldpstate = lp->lp_partner.lip_state;
1749         if (lp->lp_lsc->lsc_strict_mode)
1750                 lp->lp_partner = lacp_partner_admin_strict;
1751         else
1752                 lp->lp_partner = lacp_partner_admin_optimistic;
1753         lp->lp_state |= LACP_STATE_DEFAULTED;
1754         lacp_sm_ptx_update_timeout(lp, oldpstate);
1755 }
1756
1757 static void
1758 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1759     const struct lacp_peerinfo *info)
1760 {
1761
1762         LACP_TRACE(lp);
1763
1764         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1765             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1766             LACP_STATE_AGGREGATION)) {
1767                 lp->lp_selected = LACP_UNSELECTED;
1768                 /* mux machine will clean up lp->lp_aggregator */
1769         }
1770 }
1771
1772 static void
1773 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1774 {
1775
1776         LACP_TRACE(lp);
1777
1778         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1779 }
1780
1781 static void
1782 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1783 {
1784
1785         LACP_TRACE(lp);
1786
1787         if (lp->lp_lsc->lsc_strict_mode)
1788                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1789                     &lacp_partner_admin_strict);
1790         else
1791                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1792                     &lacp_partner_admin_optimistic);
1793 }
1794
1795 /* transmit machine */
1796
1797 static void
1798 lacp_sm_tx(struct lacp_port *lp)
1799 {
1800         int error = 0;
1801
1802         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1803 #if 1
1804             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1805             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1806 #endif
1807             ) {
1808                 lp->lp_flags &= ~LACP_PORT_NTT;
1809         }
1810
1811         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1812                 return;
1813         }
1814
1815         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1816         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1817                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1818                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1819                 return;
1820         }
1821
1822         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1823                 error = lacp_xmit_lacpdu(lp);
1824         } else {
1825                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1826         }
1827
1828         if (error == 0) {
1829                 lp->lp_flags &= ~LACP_PORT_NTT;
1830         } else {
1831                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1832                     error));
1833         }
1834 }
1835
1836 static void
1837 lacp_sm_assert_ntt(struct lacp_port *lp)
1838 {
1839
1840         lp->lp_flags |= LACP_PORT_NTT;
1841 }
1842
1843 static void
1844 lacp_run_timers(struct lacp_port *lp)
1845 {
1846         int i;
1847
1848         for (i = 0; i < LACP_NTIMER; i++) {
1849                 KASSERT(lp->lp_timer[i] >= 0,
1850                     ("invalid timer value %d", lp->lp_timer[i]));
1851                 if (lp->lp_timer[i] == 0) {
1852                         continue;
1853                 } else if (--lp->lp_timer[i] <= 0) {
1854                         if (lacp_timer_funcs[i]) {
1855                                 (*lacp_timer_funcs[i])(lp);
1856                         }
1857                 }
1858         }
1859 }
1860
1861 int
1862 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1863 {
1864         struct lacp_softc *lsc = lp->lp_lsc;
1865         struct lagg_port *lgp = lp->lp_lagg;
1866         struct lacp_port *lp2;
1867         struct markerdu *mdu;
1868         int error = 0;
1869         int pending = 0;
1870
1871         if (m->m_pkthdr.len != sizeof(*mdu)) {
1872                 goto bad;
1873         }
1874
1875         if ((m->m_flags & M_MCAST) == 0) {
1876                 goto bad;
1877         }
1878
1879         if (m->m_len < sizeof(*mdu)) {
1880                 m = m_pullup(m, sizeof(*mdu));
1881                 if (m == NULL) {
1882                         return (ENOMEM);
1883                 }
1884         }
1885
1886         mdu = mtod(m, struct markerdu *);
1887
1888         if (memcmp(&mdu->mdu_eh.ether_dhost,
1889             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1890                 goto bad;
1891         }
1892
1893         if (mdu->mdu_sph.sph_version != 1) {
1894                 goto bad;
1895         }
1896
1897         switch (mdu->mdu_tlv.tlv_type) {
1898         case MARKER_TYPE_INFO:
1899                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1900                     marker_info_tlv_template, TRUE)) {
1901                         goto bad;
1902                 }
1903                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1904                 memcpy(&mdu->mdu_eh.ether_dhost,
1905                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1906                 memcpy(&mdu->mdu_eh.ether_shost,
1907                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1908                 error = lagg_enqueue(lp->lp_ifp, m);
1909                 break;
1910
1911         case MARKER_TYPE_RESPONSE:
1912                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1913                     marker_response_tlv_template, TRUE)) {
1914                         goto bad;
1915                 }
1916                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1917                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1918                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1919
1920                 /* Verify that it is the last marker we sent out */
1921                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1922                     sizeof(struct lacp_markerinfo)))
1923                         goto bad;
1924
1925                 LACP_LOCK(lsc);
1926                 lp->lp_flags &= ~LACP_PORT_MARK;
1927
1928                 if (lsc->lsc_suppress_distributing) {
1929                         /* Check if any ports are waiting for a response */
1930                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1931                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1932                                         pending = 1;
1933                                         break;
1934                                 }
1935                         }
1936
1937                         if (pending == 0) {
1938                                 /* All interface queues are clear */
1939                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1940                                 lsc->lsc_suppress_distributing = FALSE;
1941                         }
1942                 }
1943                 LACP_UNLOCK(lsc);
1944                 m_freem(m);
1945                 break;
1946
1947         default:
1948                 goto bad;
1949         }
1950
1951         return (error);
1952
1953 bad:
1954         LACP_DPRINTF((lp, "bad marker frame\n"));
1955         m_freem(m);
1956         return (EINVAL);
1957 }
1958
1959 static int
1960 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1961     const struct tlv_template *tmpl, boolean_t check_type)
1962 {
1963         while (/* CONSTCOND */ 1) {
1964                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1965                         return (EINVAL);
1966                 }
1967                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1968                     tlv->tlv_length != tmpl->tmpl_length) {
1969                         return (EINVAL);
1970                 }
1971                 if (tmpl->tmpl_type == 0) {
1972                         break;
1973                 }
1974                 tlv = (const struct tlvhdr *)
1975                     ((const char *)tlv + tlv->tlv_length);
1976                 tmpl++;
1977         }
1978
1979         return (0);
1980 }
1981
1982 /* Debugging */
1983 const char *
1984 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1985 {
1986         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1987             (int)mac[0],
1988             (int)mac[1],
1989             (int)mac[2],
1990             (int)mac[3],
1991             (int)mac[4],
1992             (int)mac[5]);
1993
1994         return (buf);
1995 }
1996
1997 const char *
1998 lacp_format_systemid(const struct lacp_systemid *sysid,
1999     char *buf, size_t buflen)
2000 {
2001         char macbuf[LACP_MACSTR_MAX+1];
2002
2003         snprintf(buf, buflen, "%04X,%s",
2004             ntohs(sysid->lsi_prio),
2005             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2006
2007         return (buf);
2008 }
2009
2010 const char *
2011 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2012 {
2013         snprintf(buf, buflen, "%04X,%04X",
2014             ntohs(portid->lpi_prio),
2015             ntohs(portid->lpi_portno));
2016
2017         return (buf);
2018 }
2019
2020 const char *
2021 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2022 {
2023         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2024         char portid[LACP_PORTIDSTR_MAX+1];
2025
2026         snprintf(buf, buflen, "(%s,%04X,%s)",
2027             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2028             ntohs(peer->lip_key),
2029             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2030
2031         return (buf);
2032 }
2033
2034 const char *
2035 lacp_format_lagid(const struct lacp_peerinfo *a,
2036     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2037 {
2038         char astr[LACP_PARTNERSTR_MAX+1];
2039         char bstr[LACP_PARTNERSTR_MAX+1];
2040
2041 #if 0
2042         /*
2043          * there's a convention to display small numbered peer
2044          * in the left.
2045          */
2046
2047         if (lacp_compare_peerinfo(a, b) > 0) {
2048                 const struct lacp_peerinfo *t;
2049
2050                 t = a;
2051                 a = b;
2052                 b = t;
2053         }
2054 #endif
2055
2056         snprintf(buf, buflen, "[%s,%s]",
2057             lacp_format_partner(a, astr, sizeof(astr)),
2058             lacp_format_partner(b, bstr, sizeof(bstr)));
2059
2060         return (buf);
2061 }
2062
2063 const char *
2064 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2065     char *buf, size_t buflen)
2066 {
2067         if (la == NULL) {
2068                 return ("(none)");
2069         }
2070
2071         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2072 }
2073
2074 const char *
2075 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2076 {
2077         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2078         return (buf);
2079 }
2080
2081 static void
2082 lacp_dump_lacpdu(const struct lacpdu *du)
2083 {
2084         char buf[LACP_PARTNERSTR_MAX+1];
2085         char buf2[LACP_STATESTR_MAX+1];
2086
2087         printf("actor=%s\n",
2088             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2089         printf("actor.state=%s\n",
2090             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2091         printf("partner=%s\n",
2092             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2093         printf("partner.state=%s\n",
2094             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2095
2096         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2097 }
2098
2099 static void
2100 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2101 {
2102         va_list va;
2103
2104         if (lp) {
2105                 printf("%s: ", lp->lp_ifp->if_xname);
2106         }
2107
2108         va_start(va, fmt);
2109         vprintf(fmt, va);
2110         va_end(va);
2111 }