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 <sys/param.h>
  34 #include <sys/callout.h>
  35 #include <sys/eventhandler.h>
  36 #include <sys/mbuf.h>
  37 #include <sys/systm.h>
  38 #include <sys/malloc.h>
  39 #include <sys/kernel.h> /* hz */
  40 #include <sys/socket.h> /* for net/if.h */
  41 #include <sys/sockio.h>
  42 #include <sys/sysctl.h>
  43 #include <machine/stdarg.h>
  44 #include <sys/lock.h>
  45 #include <sys/rwlock.h>
  46 #include <sys/taskqueue.h>
  47
  48 #include <net/if.h>
  49 #include <net/if_var.h>
  50 #include <net/if_dl.h>
  51 #include <net/ethernet.h>
  52 #include <net/if_media.h>
  53 #include <net/if_types.h>
  54
  55 #include <net/if_lagg.h>
  56 #include <net/ieee8023ad_lacp.h>
  57
  58 /*
  59  * actor system priority and port priority.
  60  * XXX should be configurable.
  61  */
  62
  63 #define LACP_SYSTEM_PRIO        0x8000
  64 #define LACP_PORT_PRIO          0x8000
  65
  66 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  67     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  68
  69 static const struct tlv_template lacp_info_tlv_template[] = {
  70         { LACP_TYPE_ACTORINFO,
  71             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  72         { LACP_TYPE_PARTNERINFO,
  73             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  74         { LACP_TYPE_COLLECTORINFO,
  75             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  76         { 0, 0 },
  77 };
  78
  79 static const struct tlv_template marker_info_tlv_template[] = {
  80         { MARKER_TYPE_INFO,
  81             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  82         { 0, 0 },
  83 };
  84
  85 static const struct tlv_template marker_response_tlv_template[] = {
  86         { MARKER_TYPE_RESPONSE,
  87             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  88         { 0, 0 },
  89 };
  90
  91 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  92
  93 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
  94 static void     lacp_fill_markerinfo(struct lacp_port *,
  95                     struct lacp_markerinfo *);
  96
  97 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
  98 static void     lacp_suppress_distributing(struct lacp_softc *,
  99                     struct lacp_aggregator *);
 100 static void     lacp_transit_expire(void *);
 101 static void     lacp_update_portmap(struct lacp_softc *);
 102 static void     lacp_select_active_aggregator(struct lacp_softc *);
 103 static uint16_t lacp_compose_key(struct lacp_port *);
 104 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 105                     const struct tlv_template *, boolean_t);
 106 static void     lacp_tick(void *);
 107
 108 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 109                     const struct lacp_port *);
 110 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 111                     const struct lacp_peerinfo *);
 112 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 113                     const struct lacp_port *);
 114 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 115                     const struct lacp_peerinfo *);
 116
 117 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 118                     struct lacp_port *);
 119 static void     lacp_aggregator_addref(struct lacp_softc *,
 120                     struct lacp_aggregator *);
 121 static void     lacp_aggregator_delref(struct lacp_softc *,
 122                     struct lacp_aggregator *);
 123
 124 /* receive machine */
 125
 126 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 127 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 128 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 129 static void     lacp_sm_rx_timer(struct lacp_port *);
 130 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 131 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 132                     const struct lacpdu *);
 133 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 134                     const struct lacpdu *);
 135 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 136                     const struct lacpdu *);
 137 static void     lacp_sm_rx_record_default(struct lacp_port *);
 138 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 139 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 140                     const struct lacp_peerinfo *);
 141
 142 /* mux machine */
 143
 144 static void     lacp_sm_mux(struct lacp_port *);
 145 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 146 static void     lacp_sm_mux_timer(struct lacp_port *);
 147
 148 /* periodic transmit machine */
 149
 150 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 151 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 152 static void     lacp_sm_ptx_timer(struct lacp_port *);
 153
 154 /* transmit machine */
 155
 156 static void     lacp_sm_tx(struct lacp_port *);
 157 static void     lacp_sm_assert_ntt(struct lacp_port *);
 158
 159 static void     lacp_run_timers(struct lacp_port *);
 160 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 161                     const struct lacp_peerinfo *);
 162 static int      lacp_compare_systemid(const struct lacp_systemid *,
 163                     const struct lacp_systemid *);
 164 static void     lacp_port_enable(struct lacp_port *);
 165 static void     lacp_port_disable(struct lacp_port *);
 166 static void     lacp_select(struct lacp_port *);
 167 static void     lacp_unselect(struct lacp_port *);
 168 static void     lacp_disable_collecting(struct lacp_port *);
 169 static void     lacp_enable_collecting(struct lacp_port *);
 170 static void     lacp_disable_distributing(struct lacp_port *);
 171 static void     lacp_enable_distributing(struct lacp_port *);
 172 static int      lacp_xmit_lacpdu(struct lacp_port *);
 173 static int      lacp_xmit_marker(struct lacp_port *);
 174
 175 /* Debugging */
 176
 177 static void     lacp_dump_lacpdu(const struct lacpdu *);
 178 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 179                     size_t);
 180 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 181                     const struct lacp_peerinfo *, char *, size_t);
 182 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 183                     char *, size_t);
 184 static const char *lacp_format_state(uint8_t, char *, size_t);
 185 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 186 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 187                     size_t);
 188 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 189                     size_t);
 190 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 191                     __attribute__((__format__(__printf__, 2, 3)));
 192
 193 static VNET_DEFINE(int, lacp_debug);
 194 #define V_lacp_debug    VNET(lacp_debug)
 195 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
 196 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 197     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 198
 199 static VNET_DEFINE(int, lacp_default_strict_mode) = 1;
 200 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode, CTLFLAG_RWTUN,
 201     &VNET_NAME(lacp_default_strict_mode), 0,
 202     "LACP strict protocol compliance default");
 203
 204 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 205 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 206 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 207
 208 /*
 209  * partner administration variables.
 210  * XXX should be configurable.
 211  */
 212
 213 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 214         .lip_systemid = { .lsi_prio = 0xffff },
 215         .lip_portid = { .lpi_prio = 0xffff },
 216         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 217             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 218 };
 219
 220 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 221         .lip_systemid = { .lsi_prio = 0xffff },
 222         .lip_portid = { .lpi_prio = 0xffff },
 223         .lip_state = 0,
 224 };
 225
 226 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 227         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 228         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 229         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 230 };
 231
 232 struct mbuf *
 233 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 234 {
 235         struct lacp_port *lp = LACP_PORT(lgp);
 236         uint8_t subtype;
 237
 238         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 239                 m_freem(m);
 240                 return (NULL);
 241         }
 242
 243         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 244         switch (subtype) {
 245                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 246                         lacp_pdu_input(lp, m);
 247                         return (NULL);
 248
 249                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 250                         lacp_marker_input(lp, m);
 251                         return (NULL);
 252         }
 253
 254         /* Not a subtype we are interested in */
 255         return (m);
 256 }
 257
 258 /*
 259  * lacp_pdu_input: process lacpdu
 260  */
 261 static int
 262 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 263 {
 264         struct lacp_softc *lsc = lp->lp_lsc;
 265         struct lacpdu *du;
 266         int error = 0;
 267
 268         if (m->m_pkthdr.len != sizeof(*du)) {
 269                 goto bad;
 270         }
 271
 272         if ((m->m_flags & M_MCAST) == 0) {
 273                 goto bad;
 274         }
 275
 276         if (m->m_len < sizeof(*du)) {
 277                 m = m_pullup(m, sizeof(*du));
 278                 if (m == NULL) {
 279                         return (ENOMEM);
 280                 }
 281         }
 282
 283         du = mtod(m, struct lacpdu *);
 284
 285         if (memcmp(&du->ldu_eh.ether_dhost,
 286             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 287                 goto bad;
 288         }
 289
 290         /*
 291          * ignore the version for compatibility with
 292          * the future protocol revisions.
 293          */
 294 #if 0
 295         if (du->ldu_sph.sph_version != 1) {
 296                 goto bad;
 297         }
 298 #endif
 299
 300         /*
 301          * ignore tlv types for compatibility with
 302          * the future protocol revisions.
 303          */
 304         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 305             lacp_info_tlv_template, FALSE)) {
 306                 goto bad;
 307         }
 308
 309         if (V_lacp_debug > 0) {
 310                 lacp_dprintf(lp, "lacpdu receive\n");
 311                 lacp_dump_lacpdu(du);
 312         }
 313
 314         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 315                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 316                 goto bad;
 317         }
 318
 319         LACP_LOCK(lsc);
 320         lacp_sm_rx(lp, du);
 321         LACP_UNLOCK(lsc);
 322
 323         m_freem(m);
 324         return (error);
 325
 326 bad:
 327         m_freem(m);
 328         return (EINVAL);
 329 }
 330
 331 static void
 332 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 333 {
 334         struct lagg_port *lgp = lp->lp_lagg;
 335         struct lagg_softc *sc = lgp->lp_softc;
 336
 337         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 338         memcpy(&info->lip_systemid.lsi_mac,
 339             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 340         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 341         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 342         info->lip_state = lp->lp_state;
 343 }
 344
 345 static void
 346 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 347 {
 348         struct ifnet *ifp = lp->lp_ifp;
 349
 350         /* Fill in the port index and system id (encoded as the MAC) */
 351         info->mi_rq_port = htons(ifp->if_index);
 352         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 353         info->mi_rq_xid = htonl(0);
 354 }
 355
 356 static int
 357 lacp_xmit_lacpdu(struct lacp_port *lp)
 358 {
 359         struct lagg_port *lgp = lp->lp_lagg;
 360         struct mbuf *m;
 361         struct lacpdu *du;
 362         int error;
 363
 364         LACP_LOCK_ASSERT(lp->lp_lsc);
 365
 366         m = m_gethdr(M_NOWAIT, MT_DATA);
 367         if (m == NULL) {
 368                 return (ENOMEM);
 369         }
 370         m->m_len = m->m_pkthdr.len = sizeof(*du);
 371
 372         du = mtod(m, struct lacpdu *);
 373         memset(du, 0, sizeof(*du));
 374
 375         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 376             ETHER_ADDR_LEN);
 377         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 378         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 379
 380         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 381         du->ldu_sph.sph_version = 1;
 382
 383         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 384         du->ldu_actor = lp->lp_actor;
 385
 386         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 387             sizeof(du->ldu_partner));
 388         du->ldu_partner = lp->lp_partner;
 389
 390         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 391             sizeof(du->ldu_collector));
 392         du->ldu_collector.lci_maxdelay = 0;
 393
 394         if (V_lacp_debug > 0) {
 395                 lacp_dprintf(lp, "lacpdu transmit\n");
 396                 lacp_dump_lacpdu(du);
 397         }
 398
 399         m->m_flags |= M_MCAST;
 400
 401         /*
 402          * XXX should use higher priority queue.
 403          * otherwise network congestion can break aggregation.
 404          */
 405
 406         error = lagg_enqueue(lp->lp_ifp, m);
 407         return (error);
 408 }
 409
 410 static int
 411 lacp_xmit_marker(struct lacp_port *lp)
 412 {
 413         struct lagg_port *lgp = lp->lp_lagg;
 414         struct mbuf *m;
 415         struct markerdu *mdu;
 416         int error;
 417
 418         LACP_LOCK_ASSERT(lp->lp_lsc);
 419
 420         m = m_gethdr(M_NOWAIT, MT_DATA);
 421         if (m == NULL) {
 422                 return (ENOMEM);
 423         }
 424         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 425
 426         mdu = mtod(m, struct markerdu *);
 427         memset(mdu, 0, sizeof(*mdu));
 428
 429         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 430             ETHER_ADDR_LEN);
 431         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 432         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 433
 434         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 435         mdu->mdu_sph.sph_version = 1;
 436
 437         /* Bump the transaction id and copy over the marker info */
 438         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 439         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 440         mdu->mdu_info = lp->lp_marker;
 441
 442         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 443             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 444             ntohl(mdu->mdu_info.mi_rq_xid)));
 445
 446         m->m_flags |= M_MCAST;
 447         error = lagg_enqueue(lp->lp_ifp, m);
 448         return (error);
 449 }
 450
 451 void
 452 lacp_linkstate(struct lagg_port *lgp)
 453 {
 454         struct lacp_port *lp = LACP_PORT(lgp);
 455         struct lacp_softc *lsc = lp->lp_lsc;
 456         struct ifnet *ifp = lgp->lp_ifp;
 457         struct ifmediareq ifmr;
 458         int error = 0;
 459         u_int media;
 460         uint8_t old_state;
 461         uint16_t old_key;
 462
 463         bzero((char *)&ifmr, sizeof(ifmr));
 464         error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 465         if (error != 0)
 466                 return;
 467
 468         LACP_LOCK(lsc);
 469         media = ifmr.ifm_active;
 470         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 471             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 472             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 473         old_state = lp->lp_state;
 474         old_key = lp->lp_key;
 475
 476         lp->lp_media = media;
 477         /*
 478          * If the port is not an active full duplex Ethernet link then it can
 479          * not be aggregated.
 480          */
 481         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 482             ifp->if_link_state != LINK_STATE_UP) {
 483                 lacp_port_disable(lp);
 484         } else {
 485                 lacp_port_enable(lp);
 486         }
 487         lp->lp_key = lacp_compose_key(lp);
 488
 489         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 490                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 491                 lp->lp_selected = LACP_UNSELECTED;
 492         }
 493         LACP_UNLOCK(lsc);
 494 }
 495
 496 static void
 497 lacp_tick(void *arg)
 498 {
 499         struct lacp_softc *lsc = arg;
 500         struct lacp_port *lp;
 501
 502         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 503                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 504                         continue;
 505
 506                 CURVNET_SET(lp->lp_ifp->if_vnet);
 507                 lacp_run_timers(lp);
 508
 509                 lacp_select(lp);
 510                 lacp_sm_mux(lp);
 511                 lacp_sm_tx(lp);
 512                 lacp_sm_ptx_tx_schedule(lp);
 513                 CURVNET_RESTORE();
 514         }
 515         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 516 }
 517
 518 int
 519 lacp_port_create(struct lagg_port *lgp)
 520 {
 521         struct lagg_softc *sc = lgp->lp_softc;
 522         struct lacp_softc *lsc = LACP_SOFTC(sc);
 523         struct lacp_port *lp;
 524         struct ifnet *ifp = lgp->lp_ifp;
 525         struct sockaddr_dl sdl;
 526         struct ifmultiaddr *rifma = NULL;
 527         int error;
 528
 529         boolean_t active = TRUE; /* XXX should be configurable */
 530         boolean_t fast = FALSE; /* Configurable via ioctl */
 531
 532         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 533         sdl.sdl_alen = ETHER_ADDR_LEN;
 534
 535         bcopy(&ethermulticastaddr_slowprotocols,
 536             LLADDR(&sdl), ETHER_ADDR_LEN);
 537         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 538         if (error) {
 539                 printf("%s: ADDMULTI failed on %s\n", __func__,
 540                     lgp->lp_ifp->if_xname);
 541                 return (error);
 542         }
 543
 544         lp = malloc(sizeof(struct lacp_port),
 545             M_DEVBUF, M_NOWAIT|M_ZERO);
 546         if (lp == NULL)
 547                 return (ENOMEM);
 548
 549         LACP_LOCK(lsc);
 550         lgp->lp_psc = lp;
 551         lp->lp_ifp = ifp;
 552         lp->lp_lagg = lgp;
 553         lp->lp_lsc = lsc;
 554         lp->lp_ifma = rifma;
 555
 556         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 557
 558         lacp_fill_actorinfo(lp, &lp->lp_actor);
 559         lacp_fill_markerinfo(lp, &lp->lp_marker);
 560         lp->lp_state =
 561             (active ? LACP_STATE_ACTIVITY : 0) |
 562             (fast ? LACP_STATE_TIMEOUT : 0);
 563         lp->lp_aggregator = NULL;
 564         lacp_sm_rx_set_expired(lp);
 565         LACP_UNLOCK(lsc);
 566         lacp_linkstate(lgp);
 567
 568         return (0);
 569 }
 570
 571 void
 572 lacp_port_destroy(struct lagg_port *lgp)
 573 {
 574         struct lacp_port *lp = LACP_PORT(lgp);
 575         struct lacp_softc *lsc = lp->lp_lsc;
 576         int i;
 577
 578         LACP_LOCK(lsc);
 579         for (i = 0; i < LACP_NTIMER; i++) {
 580                 LACP_TIMER_DISARM(lp, i);
 581         }
 582
 583         lacp_disable_collecting(lp);
 584         lacp_disable_distributing(lp);
 585         lacp_unselect(lp);
 586
 587         LIST_REMOVE(lp, lp_next);
 588         LACP_UNLOCK(lsc);
 589
 590         /* The address may have already been removed by if_purgemaddrs() */
 591         if (!lgp->lp_detaching)
 592                 if_delmulti_ifma(lp->lp_ifma);
 593
 594         free(lp, M_DEVBUF);
 595 }
 596
 597 void
 598 lacp_req(struct lagg_softc *sc, void *data)
 599 {
 600         struct lacp_opreq *req = (struct lacp_opreq *)data;
 601         struct lacp_softc *lsc = LACP_SOFTC(sc);
 602         struct lacp_aggregator *la;
 603
 604         bzero(req, sizeof(struct lacp_opreq));
 605
 606         /*
 607          * If the LACP softc is NULL, return with the opreq structure full of
 608          * zeros.  It is normal for the softc to be NULL while the lagg is
 609          * being destroyed.
 610          */
 611         if (NULL == lsc)
 612                 return;
 613
 614         la = lsc->lsc_active_aggregator;
 615         LACP_LOCK(lsc);
 616         if (la != NULL) {
 617                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 618                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 619                     ETHER_ADDR_LEN);
 620                 req->actor_key = ntohs(la->la_actor.lip_key);
 621                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 622                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 623                 req->actor_state = la->la_actor.lip_state;
 624
 625                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 626                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 627                     ETHER_ADDR_LEN);
 628                 req->partner_key = ntohs(la->la_partner.lip_key);
 629                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 630                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 631                 req->partner_state = la->la_partner.lip_state;
 632         }
 633         LACP_UNLOCK(lsc);
 634 }
 635
 636 void
 637 lacp_portreq(struct lagg_port *lgp, void *data)
 638 {
 639         struct lacp_opreq *req = (struct lacp_opreq *)data;
 640         struct lacp_port *lp = LACP_PORT(lgp);
 641         struct lacp_softc *lsc = lp->lp_lsc;
 642
 643         LACP_LOCK(lsc);
 644         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 645         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 646             ETHER_ADDR_LEN);
 647         req->actor_key = ntohs(lp->lp_actor.lip_key);
 648         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 649         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 650         req->actor_state = lp->lp_actor.lip_state;
 651
 652         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 653         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 654             ETHER_ADDR_LEN);
 655         req->partner_key = ntohs(lp->lp_partner.lip_key);
 656         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 657         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 658         req->partner_state = lp->lp_partner.lip_state;
 659         LACP_UNLOCK(lsc);
 660 }
 661
 662 static void
 663 lacp_disable_collecting(struct lacp_port *lp)
 664 {
 665         LACP_DPRINTF((lp, "collecting disabled\n"));
 666         lp->lp_state &= ~LACP_STATE_COLLECTING;
 667 }
 668
 669 static void
 670 lacp_enable_collecting(struct lacp_port *lp)
 671 {
 672         LACP_DPRINTF((lp, "collecting enabled\n"));
 673         lp->lp_state |= LACP_STATE_COLLECTING;
 674 }
 675
 676 static void
 677 lacp_disable_distributing(struct lacp_port *lp)
 678 {
 679         struct lacp_aggregator *la = lp->lp_aggregator;
 680         struct lacp_softc *lsc = lp->lp_lsc;
 681         struct lagg_softc *sc = lsc->lsc_softc;
 682         char buf[LACP_LAGIDSTR_MAX+1];
 683
 684         LACP_LOCK_ASSERT(lsc);
 685
 686         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 687                 return;
 688         }
 689
 690         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 691         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 692         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 693
 694         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 695             "nports %d -> %d\n",
 696             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 697             la->la_nports, la->la_nports - 1));
 698
 699         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 700         la->la_nports--;
 701         sc->sc_active = la->la_nports;
 702
 703         if (lsc->lsc_active_aggregator == la) {
 704                 lacp_suppress_distributing(lsc, la);
 705                 lacp_select_active_aggregator(lsc);
 706                 /* regenerate the port map, the active aggregator has changed */
 707                 lacp_update_portmap(lsc);
 708         }
 709
 710         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 711 }
 712
 713 static void
 714 lacp_enable_distributing(struct lacp_port *lp)
 715 {
 716         struct lacp_aggregator *la = lp->lp_aggregator;
 717         struct lacp_softc *lsc = lp->lp_lsc;
 718         struct lagg_softc *sc = lsc->lsc_softc;
 719         char buf[LACP_LAGIDSTR_MAX+1];
 720
 721         LACP_LOCK_ASSERT(lsc);
 722
 723         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 724                 return;
 725         }
 726
 727         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 728             "nports %d -> %d\n",
 729             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 730             la->la_nports, la->la_nports + 1));
 731
 732         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 733         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 734         la->la_nports++;
 735         sc->sc_active = la->la_nports;
 736
 737         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 738
 739         if (lsc->lsc_active_aggregator == la) {
 740                 lacp_suppress_distributing(lsc, la);
 741                 lacp_update_portmap(lsc);
 742         } else
 743                 /* try to become the active aggregator */
 744                 lacp_select_active_aggregator(lsc);
 745 }
 746
 747 static void
 748 lacp_transit_expire(void *vp)
 749 {
 750         struct lacp_softc *lsc = vp;
 751
 752         LACP_LOCK_ASSERT(lsc);
 753
 754         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 755         LACP_TRACE(NULL);
 756         CURVNET_RESTORE();
 757
 758         lsc->lsc_suppress_distributing = FALSE;
 759 }
 760
 761 void
 762 lacp_attach(struct lagg_softc *sc)
 763 {
 764         struct lacp_softc *lsc;
 765
 766         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 767
 768         sc->sc_psc = lsc;
 769         lsc->lsc_softc = sc;
 770
 771         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 772         lsc->lsc_active_aggregator = NULL;
 773         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 774         LACP_LOCK_INIT(lsc);
 775         TAILQ_INIT(&lsc->lsc_aggregators);
 776         LIST_INIT(&lsc->lsc_ports);
 777
 778         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 779         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 780
 781         /* if the lagg is already up then do the same */
 782         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 783                 lacp_init(sc);
 784 }
 785
 786 void
 787 lacp_detach(void *psc)
 788 {
 789         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 790
 791         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 792             ("aggregators still active"));
 793         KASSERT(lsc->lsc_active_aggregator == NULL,
 794             ("aggregator still attached"));
 795
 796         callout_drain(&lsc->lsc_transit_callout);
 797         callout_drain(&lsc->lsc_callout);
 798
 799         LACP_LOCK_DESTROY(lsc);
 800         free(lsc, M_DEVBUF);
 801 }
 802
 803 void
 804 lacp_init(struct lagg_softc *sc)
 805 {
 806         struct lacp_softc *lsc = LACP_SOFTC(sc);
 807
 808         LACP_LOCK(lsc);
 809         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 810         LACP_UNLOCK(lsc);
 811 }
 812
 813 void
 814 lacp_stop(struct lagg_softc *sc)
 815 {
 816         struct lacp_softc *lsc = LACP_SOFTC(sc);
 817
 818         LACP_LOCK(lsc);
 819         callout_stop(&lsc->lsc_transit_callout);
 820         callout_stop(&lsc->lsc_callout);
 821         LACP_UNLOCK(lsc);
 822 }
 823
 824 struct lagg_port *
 825 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 826 {
 827         struct lacp_softc *lsc = LACP_SOFTC(sc);
 828         struct lacp_portmap *pm;
 829         struct lacp_port *lp;
 830         uint32_t hash;
 831
 832         if (__predict_false(lsc->lsc_suppress_distributing)) {
 833                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 834                 return (NULL);
 835         }
 836
 837         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 838         if (pm->pm_count == 0) {
 839                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 840                 return (NULL);
 841         }
 842
 843         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 844             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 845                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 846         else
 847                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 848         hash %= pm->pm_count;
 849         lp = pm->pm_map[hash];
 850
 851         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 852             ("aggregated port is not distributing"));
 853
 854         return (lp->lp_lagg);
 855 }
 856 /*
 857  * lacp_suppress_distributing: drop transmit packets for a while
 858  * to preserve packet ordering.
 859  */
 860
 861 static void
 862 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 863 {
 864         struct lacp_port *lp;
 865
 866         if (lsc->lsc_active_aggregator != la) {
 867                 return;
 868         }
 869
 870         LACP_TRACE(NULL);
 871
 872         lsc->lsc_suppress_distributing = TRUE;
 873
 874         /* send a marker frame down each port to verify the queues are empty */
 875         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 876                 lp->lp_flags |= LACP_PORT_MARK;
 877                 lacp_xmit_marker(lp);
 878         }
 879
 880         /* set a timeout for the marker frames */
 881         callout_reset(&lsc->lsc_transit_callout,
 882             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 883 }
 884
 885 static int
 886 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 887     const struct lacp_peerinfo *b)
 888 {
 889         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 890 }
 891
 892 static int
 893 lacp_compare_systemid(const struct lacp_systemid *a,
 894     const struct lacp_systemid *b)
 895 {
 896         return (memcmp(a, b, sizeof(*a)));
 897 }
 898
 899 #if 0   /* unused */
 900 static int
 901 lacp_compare_portid(const struct lacp_portid *a,
 902     const struct lacp_portid *b)
 903 {
 904         return (memcmp(a, b, sizeof(*a)));
 905 }
 906 #endif
 907
 908 static uint64_t
 909 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 910 {
 911         struct lacp_port *lp;
 912         uint64_t speed;
 913
 914         lp = TAILQ_FIRST(&la->la_ports);
 915         if (lp == NULL) {
 916                 return (0);
 917         }
 918
 919         speed = ifmedia_baudrate(lp->lp_media);
 920         speed *= la->la_nports;
 921         if (speed == 0) {
 922                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 923                     lp->lp_media, la->la_nports));
 924         }
 925
 926         return (speed);
 927 }
 928
 929 /*
 930  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 931  * packets from lagg(4) interface.
 932  */
 933
 934 static void
 935 lacp_select_active_aggregator(struct lacp_softc *lsc)
 936 {
 937         struct lacp_aggregator *la;
 938         struct lacp_aggregator *best_la = NULL;
 939         uint64_t best_speed = 0;
 940         char buf[LACP_LAGIDSTR_MAX+1];
 941
 942         LACP_TRACE(NULL);
 943
 944         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 945                 uint64_t speed;
 946
 947                 if (la->la_nports == 0) {
 948                         continue;
 949                 }
 950
 951                 speed = lacp_aggregator_bandwidth(la);
 952                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 953                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 954                     speed, la->la_nports));
 955
 956                 /*
 957                  * This aggregator is chosen if the partner has a better
 958                  * system priority or, the total aggregated speed is higher
 959                  * or, it is already the chosen aggregator
 960                  */
 961                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 962                     LACP_SYS_PRI(best_la->la_partner)) ||
 963                     speed > best_speed ||
 964                     (speed == best_speed &&
 965                     la == lsc->lsc_active_aggregator)) {
 966                         best_la = la;
 967                         best_speed = speed;
 968                 }
 969         }
 970
 971         KASSERT(best_la == NULL || best_la->la_nports > 0,
 972             ("invalid aggregator refcnt"));
 973         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
 974             ("invalid aggregator list"));
 975
 976         if (lsc->lsc_active_aggregator != best_la) {
 977                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
 978                 LACP_DPRINTF((NULL, "old %s\n",
 979                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
 980                     buf, sizeof(buf))));
 981         } else {
 982                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
 983         }
 984         LACP_DPRINTF((NULL, "new %s\n",
 985             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
 986
 987         if (lsc->lsc_active_aggregator != best_la) {
 988                 lsc->lsc_active_aggregator = best_la;
 989                 lacp_update_portmap(lsc);
 990                 if (best_la) {
 991                         lacp_suppress_distributing(lsc, best_la);
 992                 }
 993         }
 994 }
 995
 996 /*
 997  * Updated the inactive portmap array with the new list of ports and
 998  * make it live.
 999  */
1000 static void
1001 lacp_update_portmap(struct lacp_softc *lsc)
1002 {
1003         struct lagg_softc *sc = lsc->lsc_softc;
1004         struct lacp_aggregator *la;
1005         struct lacp_portmap *p;
1006         struct lacp_port *lp;
1007         uint64_t speed;
1008         u_int newmap;
1009         int i;
1010
1011         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1012         p = &lsc->lsc_pmap[newmap];
1013         la = lsc->lsc_active_aggregator;
1014         speed = 0;
1015         bzero(p, sizeof(struct lacp_portmap));
1016
1017         if (la != NULL && la->la_nports > 0) {
1018                 p->pm_count = la->la_nports;
1019                 i = 0;
1020                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1021                         p->pm_map[i++] = lp;
1022                 KASSERT(i == p->pm_count, ("Invalid port count"));
1023                 speed = lacp_aggregator_bandwidth(la);
1024         }
1025         sc->sc_ifp->if_baudrate = speed;
1026
1027         /* switch the active portmap over */
1028         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1029         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1030                     lsc->lsc_activemap,
1031                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1032 }
1033
1034 static uint16_t
1035 lacp_compose_key(struct lacp_port *lp)
1036 {
1037         struct lagg_port *lgp = lp->lp_lagg;
1038         struct lagg_softc *sc = lgp->lp_softc;
1039         u_int media = lp->lp_media;
1040         uint16_t key;
1041
1042         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1043
1044                 /*
1045                  * non-aggregatable links should have unique keys.
1046                  *
1047                  * XXX this isn't really unique as if_index is 16 bit.
1048                  */
1049
1050                 /* bit 0..14:   (some bits of) if_index of this port */
1051                 key = lp->lp_ifp->if_index;
1052                 /* bit 15:      1 */
1053                 key |= 0x8000;
1054         } else {
1055                 u_int subtype = IFM_SUBTYPE(media);
1056
1057                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1058                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1059
1060                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1061                 switch (subtype) {
1062                 case IFM_10_T:
1063                 case IFM_10_2:
1064                 case IFM_10_5:
1065                 case IFM_10_STP:
1066                 case IFM_10_FL:
1067                         key = IFM_10_T;
1068                         break;
1069                 case IFM_100_TX:
1070                 case IFM_100_FX:
1071                 case IFM_100_T4:
1072                 case IFM_100_VG:
1073                 case IFM_100_T2:
1074                 case IFM_100_T:
1075                         key = IFM_100_TX;
1076                         break;
1077                 case IFM_1000_SX:
1078                 case IFM_1000_LX:
1079                 case IFM_1000_CX:
1080                 case IFM_1000_T:
1081                 case IFM_1000_KX:
1082                 case IFM_1000_SGMII:
1083                 case IFM_1000_CX_SGMII:
1084                         key = IFM_1000_SX;
1085                         break;
1086                 case IFM_10G_LR:
1087                 case IFM_10G_SR:
1088                 case IFM_10G_CX4:
1089                 case IFM_10G_TWINAX:
1090                 case IFM_10G_TWINAX_LONG:
1091                 case IFM_10G_LRM:
1092                 case IFM_10G_T:
1093                 case IFM_10G_KX4:
1094                 case IFM_10G_KR:
1095                 case IFM_10G_CR1:
1096                 case IFM_10G_ER:
1097                 case IFM_10G_SFI:
1098                         key = IFM_10G_LR;
1099                         break;
1100                 case IFM_20G_KR2:
1101                         key = IFM_20G_KR2;
1102                         break;
1103                 case IFM_2500_KX:
1104                 case IFM_2500_T:
1105                         key = IFM_2500_KX;
1106                         break;
1107                 case IFM_5000_T:
1108                         key = IFM_5000_T;
1109                         break;
1110                 case IFM_50G_PCIE:
1111                 case IFM_50G_CR2:
1112                 case IFM_50G_KR2:
1113                         key = IFM_50G_PCIE;
1114                         break;
1115                 case IFM_56G_R4:
1116                         key = IFM_56G_R4;
1117                         break;
1118                 case IFM_25G_PCIE:
1119                 case IFM_25G_CR:
1120                 case IFM_25G_KR:
1121                 case IFM_25G_SR:
1122                         key = IFM_25G_PCIE;
1123                         break;
1124                 case IFM_40G_CR4:
1125                 case IFM_40G_SR4:
1126                 case IFM_40G_LR4:
1127                 case IFM_40G_XLPPI:
1128                 case IFM_40G_KR4:
1129                         key = IFM_40G_CR4;
1130                         break;
1131                 case IFM_100G_CR4:
1132                 case IFM_100G_SR4:
1133                 case IFM_100G_KR4:
1134                 case IFM_100G_LR4:
1135                         key = IFM_100G_CR4;
1136                         break;
1137                 default:
1138                         key = subtype;
1139                         break;
1140                 }
1141                 /* bit 5..14:   (some bits of) if_index of lagg device */
1142                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1143                 /* bit 15:      0 */
1144         }
1145         return (htons(key));
1146 }
1147
1148 static void
1149 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1150 {
1151         char buf[LACP_LAGIDSTR_MAX+1];
1152
1153         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1154             __func__,
1155             lacp_format_lagid(&la->la_actor, &la->la_partner,
1156             buf, sizeof(buf)),
1157             la->la_refcnt, la->la_refcnt + 1));
1158
1159         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1160         la->la_refcnt++;
1161         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1162 }
1163
1164 static void
1165 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1166 {
1167         char buf[LACP_LAGIDSTR_MAX+1];
1168
1169         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1170             __func__,
1171             lacp_format_lagid(&la->la_actor, &la->la_partner,
1172             buf, sizeof(buf)),
1173             la->la_refcnt, la->la_refcnt - 1));
1174
1175         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1176         la->la_refcnt--;
1177         if (la->la_refcnt > 0) {
1178                 return;
1179         }
1180
1181         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1182         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1183
1184         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1185
1186         free(la, M_DEVBUF);
1187 }
1188
1189 /*
1190  * lacp_aggregator_get: allocate an aggregator.
1191  */
1192
1193 static struct lacp_aggregator *
1194 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1195 {
1196         struct lacp_aggregator *la;
1197
1198         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1199         if (la) {
1200                 la->la_refcnt = 1;
1201                 la->la_nports = 0;
1202                 TAILQ_INIT(&la->la_ports);
1203                 la->la_pending = 0;
1204                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1205         }
1206
1207         return (la);
1208 }
1209
1210 /*
1211  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1212  */
1213
1214 static void
1215 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1216 {
1217         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1218         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1219
1220         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1221 }
1222
1223 static void
1224 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1225     const struct lacp_peerinfo *lpi_port)
1226 {
1227         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1228         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1229         lpi_aggr->lip_key = lpi_port->lip_key;
1230 }
1231
1232 /*
1233  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1234  */
1235
1236 static int
1237 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1238     const struct lacp_port *lp)
1239 {
1240         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1241             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1242                 return (0);
1243         }
1244
1245         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1246                 return (0);
1247         }
1248
1249         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1250                 return (0);
1251         }
1252
1253         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1254                 return (0);
1255         }
1256
1257         return (1);
1258 }
1259
1260 static int
1261 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1262     const struct lacp_peerinfo *b)
1263 {
1264         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1265             sizeof(a->lip_systemid))) {
1266                 return (0);
1267         }
1268
1269         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1270                 return (0);
1271         }
1272
1273         return (1);
1274 }
1275
1276 static void
1277 lacp_port_enable(struct lacp_port *lp)
1278 {
1279         lp->lp_state |= LACP_STATE_AGGREGATION;
1280 }
1281
1282 static void
1283 lacp_port_disable(struct lacp_port *lp)
1284 {
1285         lacp_set_mux(lp, LACP_MUX_DETACHED);
1286
1287         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1288         lp->lp_selected = LACP_UNSELECTED;
1289         lacp_sm_rx_record_default(lp);
1290         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1291         lp->lp_state &= ~LACP_STATE_EXPIRED;
1292 }
1293
1294 /*
1295  * lacp_select: select an aggregator.  create one if necessary.
1296  */
1297 static void
1298 lacp_select(struct lacp_port *lp)
1299 {
1300         struct lacp_softc *lsc = lp->lp_lsc;
1301         struct lacp_aggregator *la;
1302         char buf[LACP_LAGIDSTR_MAX+1];
1303
1304         if (lp->lp_aggregator) {
1305                 return;
1306         }
1307
1308         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1309             ("timer_wait_while still active"));
1310
1311         LACP_DPRINTF((lp, "port lagid=%s\n",
1312             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1313             buf, sizeof(buf))));
1314
1315         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1316                 if (lacp_aggregator_is_compatible(la, lp)) {
1317                         break;
1318                 }
1319         }
1320
1321         if (la == NULL) {
1322                 la = lacp_aggregator_get(lsc, lp);
1323                 if (la == NULL) {
1324                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1325
1326                         /*
1327                          * will retry on the next tick.
1328                          */
1329
1330                         return;
1331                 }
1332                 lacp_fill_aggregator_id(la, lp);
1333                 LACP_DPRINTF((lp, "aggregator created\n"));
1334         } else {
1335                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1336                 if (la->la_refcnt == LACP_MAX_PORTS)
1337                         return;
1338                 lacp_aggregator_addref(lsc, la);
1339         }
1340
1341         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1342             lacp_format_lagid(&la->la_actor, &la->la_partner,
1343             buf, sizeof(buf))));
1344
1345         lp->lp_aggregator = la;
1346         lp->lp_selected = LACP_SELECTED;
1347 }
1348
1349 /*
1350  * lacp_unselect: finish unselect/detach process.
1351  */
1352
1353 static void
1354 lacp_unselect(struct lacp_port *lp)
1355 {
1356         struct lacp_softc *lsc = lp->lp_lsc;
1357         struct lacp_aggregator *la = lp->lp_aggregator;
1358
1359         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1360             ("timer_wait_while still active"));
1361
1362         if (la == NULL) {
1363                 return;
1364         }
1365
1366         lp->lp_aggregator = NULL;
1367         lacp_aggregator_delref(lsc, la);
1368 }
1369
1370 /* mux machine */
1371
1372 static void
1373 lacp_sm_mux(struct lacp_port *lp)
1374 {
1375         struct lagg_port *lgp = lp->lp_lagg;
1376         struct lagg_softc *sc = lgp->lp_softc;
1377         enum lacp_mux_state new_state;
1378         boolean_t p_sync =
1379                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1380         boolean_t p_collecting =
1381             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1382         enum lacp_selected selected = lp->lp_selected;
1383         struct lacp_aggregator *la;
1384
1385         if (V_lacp_debug > 1)
1386                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1387                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1388                     lp->lp_mux_state, selected, p_sync, p_collecting);
1389
1390 re_eval:
1391         la = lp->lp_aggregator;
1392         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1393             ("MUX not detached"));
1394         new_state = lp->lp_mux_state;
1395         switch (lp->lp_mux_state) {
1396         case LACP_MUX_DETACHED:
1397                 if (selected != LACP_UNSELECTED) {
1398                         new_state = LACP_MUX_WAITING;
1399                 }
1400                 break;
1401         case LACP_MUX_WAITING:
1402                 KASSERT(la->la_pending > 0 ||
1403                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1404                     ("timer_wait_while still active"));
1405                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1406                         new_state = LACP_MUX_ATTACHED;
1407                 } else if (selected == LACP_UNSELECTED) {
1408                         new_state = LACP_MUX_DETACHED;
1409                 }
1410                 break;
1411         case LACP_MUX_ATTACHED:
1412                 if (selected == LACP_SELECTED && p_sync) {
1413                         new_state = LACP_MUX_COLLECTING;
1414                 } else if (selected != LACP_SELECTED) {
1415                         new_state = LACP_MUX_DETACHED;
1416                 }
1417                 break;
1418         case LACP_MUX_COLLECTING:
1419                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1420                         new_state = LACP_MUX_DISTRIBUTING;
1421                 } else if (selected != LACP_SELECTED || !p_sync) {
1422                         new_state = LACP_MUX_ATTACHED;
1423                 }
1424                 break;
1425         case LACP_MUX_DISTRIBUTING:
1426                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1427                         new_state = LACP_MUX_COLLECTING;
1428                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1429                         sc->sc_flapping++;
1430                 }
1431                 break;
1432         default:
1433                 panic("%s: unknown state", __func__);
1434         }
1435
1436         if (lp->lp_mux_state == new_state) {
1437                 return;
1438         }
1439
1440         lacp_set_mux(lp, new_state);
1441         goto re_eval;
1442 }
1443
1444 static void
1445 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1446 {
1447         struct lacp_aggregator *la = lp->lp_aggregator;
1448
1449         if (lp->lp_mux_state == new_state) {
1450                 return;
1451         }
1452
1453         switch (new_state) {
1454         case LACP_MUX_DETACHED:
1455                 lp->lp_state &= ~LACP_STATE_SYNC;
1456                 lacp_disable_distributing(lp);
1457                 lacp_disable_collecting(lp);
1458                 lacp_sm_assert_ntt(lp);
1459                 /* cancel timer */
1460                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1461                         KASSERT(la->la_pending > 0,
1462                             ("timer_wait_while not active"));
1463                         la->la_pending--;
1464                 }
1465                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1466                 lacp_unselect(lp);
1467                 break;
1468         case LACP_MUX_WAITING:
1469                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1470                     LACP_AGGREGATE_WAIT_TIME);
1471                 la->la_pending++;
1472                 break;
1473         case LACP_MUX_ATTACHED:
1474                 lp->lp_state |= LACP_STATE_SYNC;
1475                 lacp_disable_collecting(lp);
1476                 lacp_sm_assert_ntt(lp);
1477                 break;
1478         case LACP_MUX_COLLECTING:
1479                 lacp_enable_collecting(lp);
1480                 lacp_disable_distributing(lp);
1481                 lacp_sm_assert_ntt(lp);
1482                 break;
1483         case LACP_MUX_DISTRIBUTING:
1484                 lacp_enable_distributing(lp);
1485                 break;
1486         default:
1487                 panic("%s: unknown state", __func__);
1488         }
1489
1490         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1491
1492         lp->lp_mux_state = new_state;
1493 }
1494
1495 static void
1496 lacp_sm_mux_timer(struct lacp_port *lp)
1497 {
1498         struct lacp_aggregator *la = lp->lp_aggregator;
1499         char buf[LACP_LAGIDSTR_MAX+1];
1500
1501         KASSERT(la->la_pending > 0, ("no pending event"));
1502
1503         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1504             lacp_format_lagid(&la->la_actor, &la->la_partner,
1505             buf, sizeof(buf)),
1506             la->la_pending, la->la_pending - 1));
1507
1508         la->la_pending--;
1509 }
1510
1511 /* periodic transmit machine */
1512
1513 static void
1514 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1515 {
1516         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1517             LACP_STATE_TIMEOUT)) {
1518                 return;
1519         }
1520
1521         LACP_DPRINTF((lp, "partner timeout changed\n"));
1522
1523         /*
1524          * FAST_PERIODIC -> SLOW_PERIODIC
1525          * or
1526          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1527          *
1528          * let lacp_sm_ptx_tx_schedule to update timeout.
1529          */
1530
1531         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1532
1533         /*
1534          * if timeout has been shortened, assert NTT.
1535          */
1536
1537         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1538                 lacp_sm_assert_ntt(lp);
1539         }
1540 }
1541
1542 static void
1543 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1544 {
1545         int timeout;
1546
1547         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1548             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1549
1550                 /*
1551                  * NO_PERIODIC
1552                  */
1553
1554                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1555                 return;
1556         }
1557
1558         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1559                 return;
1560         }
1561
1562         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1563             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1564
1565         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1566 }
1567
1568 static void
1569 lacp_sm_ptx_timer(struct lacp_port *lp)
1570 {
1571         lacp_sm_assert_ntt(lp);
1572 }
1573
1574 static void
1575 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1576 {
1577         int timeout;
1578
1579         /*
1580          * check LACP_DISABLED first
1581          */
1582
1583         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1584                 return;
1585         }
1586
1587         /*
1588          * check loopback condition.
1589          */
1590
1591         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1592             &lp->lp_actor.lip_systemid)) {
1593                 return;
1594         }
1595
1596         /*
1597          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1598          */
1599
1600         lacp_sm_rx_update_selected(lp, du);
1601         lacp_sm_rx_update_ntt(lp, du);
1602         lacp_sm_rx_record_pdu(lp, du);
1603
1604         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1605             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1606         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1607
1608         lp->lp_state &= ~LACP_STATE_EXPIRED;
1609
1610         /*
1611          * kick transmit machine without waiting the next tick.
1612          */
1613
1614         lacp_sm_tx(lp);
1615 }
1616
1617 static void
1618 lacp_sm_rx_set_expired(struct lacp_port *lp)
1619 {
1620         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1621         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1622         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1623         lp->lp_state |= LACP_STATE_EXPIRED;
1624 }
1625
1626 static void
1627 lacp_sm_rx_timer(struct lacp_port *lp)
1628 {
1629         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1630                 /* CURRENT -> EXPIRED */
1631                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1632                 lacp_sm_rx_set_expired(lp);
1633         } else {
1634                 /* EXPIRED -> DEFAULTED */
1635                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1636                 lacp_sm_rx_update_default_selected(lp);
1637                 lacp_sm_rx_record_default(lp);
1638                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1639         }
1640 }
1641
1642 static void
1643 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1644 {
1645         boolean_t active;
1646         uint8_t oldpstate;
1647         char buf[LACP_STATESTR_MAX+1];
1648
1649         LACP_TRACE(lp);
1650
1651         oldpstate = lp->lp_partner.lip_state;
1652
1653         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1654             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1655             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1656
1657         lp->lp_partner = du->ldu_actor;
1658         if (active &&
1659             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1660             LACP_STATE_AGGREGATION) &&
1661             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1662             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1663                 /* XXX nothing? */
1664         } else {
1665                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1666         }
1667
1668         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1669
1670         if (oldpstate != lp->lp_partner.lip_state) {
1671                 LACP_DPRINTF((lp, "old pstate %s\n",
1672                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1673                 LACP_DPRINTF((lp, "new pstate %s\n",
1674                     lacp_format_state(lp->lp_partner.lip_state, buf,
1675                     sizeof(buf))));
1676         }
1677
1678         /* XXX Hack, still need to implement 5.4.9 para 2,3,4 */
1679         if (lp->lp_lsc->lsc_strict_mode)
1680                 lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1681
1682         lacp_sm_ptx_update_timeout(lp, oldpstate);
1683 }
1684
1685 static void
1686 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1687 {
1688
1689         LACP_TRACE(lp);
1690
1691         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1692             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1693             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1694                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1695                 lacp_sm_assert_ntt(lp);
1696         }
1697 }
1698
1699 static void
1700 lacp_sm_rx_record_default(struct lacp_port *lp)
1701 {
1702         uint8_t oldpstate;
1703
1704         LACP_TRACE(lp);
1705
1706         oldpstate = lp->lp_partner.lip_state;
1707         if (lp->lp_lsc->lsc_strict_mode)
1708                 lp->lp_partner = lacp_partner_admin_strict;
1709         else
1710                 lp->lp_partner = lacp_partner_admin_optimistic;
1711         lp->lp_state |= LACP_STATE_DEFAULTED;
1712         lacp_sm_ptx_update_timeout(lp, oldpstate);
1713 }
1714
1715 static void
1716 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1717     const struct lacp_peerinfo *info)
1718 {
1719
1720         LACP_TRACE(lp);
1721
1722         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1723             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1724             LACP_STATE_AGGREGATION)) {
1725                 lp->lp_selected = LACP_UNSELECTED;
1726                 /* mux machine will clean up lp->lp_aggregator */
1727         }
1728 }
1729
1730 static void
1731 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1732 {
1733
1734         LACP_TRACE(lp);
1735
1736         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1737 }
1738
1739 static void
1740 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1741 {
1742
1743         LACP_TRACE(lp);
1744
1745         if (lp->lp_lsc->lsc_strict_mode)
1746                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1747                     &lacp_partner_admin_strict);
1748         else
1749                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1750                     &lacp_partner_admin_optimistic);
1751 }
1752
1753 /* transmit machine */
1754
1755 static void
1756 lacp_sm_tx(struct lacp_port *lp)
1757 {
1758         int error = 0;
1759
1760         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1761 #if 1
1762             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1763             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1764 #endif
1765             ) {
1766                 lp->lp_flags &= ~LACP_PORT_NTT;
1767         }
1768
1769         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1770                 return;
1771         }
1772
1773         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1774         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1775                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1776                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1777                 return;
1778         }
1779
1780         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1781                 error = lacp_xmit_lacpdu(lp);
1782         } else {
1783                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1784         }
1785
1786         if (error == 0) {
1787                 lp->lp_flags &= ~LACP_PORT_NTT;
1788         } else {
1789                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1790                     error));
1791         }
1792 }
1793
1794 static void
1795 lacp_sm_assert_ntt(struct lacp_port *lp)
1796 {
1797
1798         lp->lp_flags |= LACP_PORT_NTT;
1799 }
1800
1801 static void
1802 lacp_run_timers(struct lacp_port *lp)
1803 {
1804         int i;
1805
1806         for (i = 0; i < LACP_NTIMER; i++) {
1807                 KASSERT(lp->lp_timer[i] >= 0,
1808                     ("invalid timer value %d", lp->lp_timer[i]));
1809                 if (lp->lp_timer[i] == 0) {
1810                         continue;
1811                 } else if (--lp->lp_timer[i] <= 0) {
1812                         if (lacp_timer_funcs[i]) {
1813                                 (*lacp_timer_funcs[i])(lp);
1814                         }
1815                 }
1816         }
1817 }
1818
1819 int
1820 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1821 {
1822         struct lacp_softc *lsc = lp->lp_lsc;
1823         struct lagg_port *lgp = lp->lp_lagg;
1824         struct lacp_port *lp2;
1825         struct markerdu *mdu;
1826         int error = 0;
1827         int pending = 0;
1828
1829         if (m->m_pkthdr.len != sizeof(*mdu)) {
1830                 goto bad;
1831         }
1832
1833         if ((m->m_flags & M_MCAST) == 0) {
1834                 goto bad;
1835         }
1836
1837         if (m->m_len < sizeof(*mdu)) {
1838                 m = m_pullup(m, sizeof(*mdu));
1839                 if (m == NULL) {
1840                         return (ENOMEM);
1841                 }
1842         }
1843
1844         mdu = mtod(m, struct markerdu *);
1845
1846         if (memcmp(&mdu->mdu_eh.ether_dhost,
1847             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1848                 goto bad;
1849         }
1850
1851         if (mdu->mdu_sph.sph_version != 1) {
1852                 goto bad;
1853         }
1854
1855         switch (mdu->mdu_tlv.tlv_type) {
1856         case MARKER_TYPE_INFO:
1857                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1858                     marker_info_tlv_template, TRUE)) {
1859                         goto bad;
1860                 }
1861                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1862                 memcpy(&mdu->mdu_eh.ether_dhost,
1863                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1864                 memcpy(&mdu->mdu_eh.ether_shost,
1865                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1866                 error = lagg_enqueue(lp->lp_ifp, m);
1867                 break;
1868
1869         case MARKER_TYPE_RESPONSE:
1870                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1871                     marker_response_tlv_template, TRUE)) {
1872                         goto bad;
1873                 }
1874                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1875                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1876                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1877
1878                 /* Verify that it is the last marker we sent out */
1879                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1880                     sizeof(struct lacp_markerinfo)))
1881                         goto bad;
1882
1883                 LACP_LOCK(lsc);
1884                 lp->lp_flags &= ~LACP_PORT_MARK;
1885
1886                 if (lsc->lsc_suppress_distributing) {
1887                         /* Check if any ports are waiting for a response */
1888                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1889                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1890                                         pending = 1;
1891                                         break;
1892                                 }
1893                         }
1894
1895                         if (pending == 0) {
1896                                 /* All interface queues are clear */
1897                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1898                                 lsc->lsc_suppress_distributing = FALSE;
1899                         }
1900                 }
1901                 LACP_UNLOCK(lsc);
1902                 m_freem(m);
1903                 break;
1904
1905         default:
1906                 goto bad;
1907         }
1908
1909         return (error);
1910
1911 bad:
1912         LACP_DPRINTF((lp, "bad marker frame\n"));
1913         m_freem(m);
1914         return (EINVAL);
1915 }
1916
1917 static int
1918 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1919     const struct tlv_template *tmpl, boolean_t check_type)
1920 {
1921         while (/* CONSTCOND */ 1) {
1922                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1923                         return (EINVAL);
1924                 }
1925                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1926                     tlv->tlv_length != tmpl->tmpl_length) {
1927                         return (EINVAL);
1928                 }
1929                 if (tmpl->tmpl_type == 0) {
1930                         break;
1931                 }
1932                 tlv = (const struct tlvhdr *)
1933                     ((const char *)tlv + tlv->tlv_length);
1934                 tmpl++;
1935         }
1936
1937         return (0);
1938 }
1939
1940 /* Debugging */
1941 const char *
1942 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1943 {
1944         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1945             (int)mac[0],
1946             (int)mac[1],
1947             (int)mac[2],
1948             (int)mac[3],
1949             (int)mac[4],
1950             (int)mac[5]);
1951
1952         return (buf);
1953 }
1954
1955 const char *
1956 lacp_format_systemid(const struct lacp_systemid *sysid,
1957     char *buf, size_t buflen)
1958 {
1959         char macbuf[LACP_MACSTR_MAX+1];
1960
1961         snprintf(buf, buflen, "%04X,%s",
1962             ntohs(sysid->lsi_prio),
1963             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1964
1965         return (buf);
1966 }
1967
1968 const char *
1969 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
1970 {
1971         snprintf(buf, buflen, "%04X,%04X",
1972             ntohs(portid->lpi_prio),
1973             ntohs(portid->lpi_portno));
1974
1975         return (buf);
1976 }
1977
1978 const char *
1979 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
1980 {
1981         char sysid[LACP_SYSTEMIDSTR_MAX+1];
1982         char portid[LACP_PORTIDSTR_MAX+1];
1983
1984         snprintf(buf, buflen, "(%s,%04X,%s)",
1985             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
1986             ntohs(peer->lip_key),
1987             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
1988
1989         return (buf);
1990 }
1991
1992 const char *
1993 lacp_format_lagid(const struct lacp_peerinfo *a,
1994     const struct lacp_peerinfo *b, char *buf, size_t buflen)
1995 {
1996         char astr[LACP_PARTNERSTR_MAX+1];
1997         char bstr[LACP_PARTNERSTR_MAX+1];
1998
1999 #if 0
2000         /*
2001          * there's a convention to display small numbered peer
2002          * in the left.
2003          */
2004
2005         if (lacp_compare_peerinfo(a, b) > 0) {
2006                 const struct lacp_peerinfo *t;
2007
2008                 t = a;
2009                 a = b;
2010                 b = t;
2011         }
2012 #endif
2013
2014         snprintf(buf, buflen, "[%s,%s]",
2015             lacp_format_partner(a, astr, sizeof(astr)),
2016             lacp_format_partner(b, bstr, sizeof(bstr)));
2017
2018         return (buf);
2019 }
2020
2021 const char *
2022 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2023     char *buf, size_t buflen)
2024 {
2025         if (la == NULL) {
2026                 return ("(none)");
2027         }
2028
2029         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2030 }
2031
2032 const char *
2033 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2034 {
2035         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2036         return (buf);
2037 }
2038
2039 static void
2040 lacp_dump_lacpdu(const struct lacpdu *du)
2041 {
2042         char buf[LACP_PARTNERSTR_MAX+1];
2043         char buf2[LACP_STATESTR_MAX+1];
2044
2045         printf("actor=%s\n",
2046             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2047         printf("actor.state=%s\n",
2048             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2049         printf("partner=%s\n",
2050             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2051         printf("partner.state=%s\n",
2052             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2053
2054         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2055 }
2056
2057 static void
2058 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2059 {
2060         va_list va;
2061
2062         if (lp) {
2063                 printf("%s: ", lp->lp_ifp->if_xname);
2064         }
2065
2066         va_start(va, fmt);
2067         vprintf(fmt, va);
2068         va_end(va);
2069 }