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,
 201     CTLFLAG_RWTUN | CTLFLAG_VNET, &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, SIOCGIFXMEDIA, (caddr_t)&ifmr);
 465         if (error != 0) {
 466                 bzero((char *)&ifmr, sizeof(ifmr));
 467                 error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 468         }
 469         if (error != 0)
 470                 return;
 471
 472         LACP_LOCK(lsc);
 473         media = ifmr.ifm_active;
 474         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 475             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 476             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 477         old_state = lp->lp_state;
 478         old_key = lp->lp_key;
 479
 480         lp->lp_media = media;
 481         /*
 482          * If the port is not an active full duplex Ethernet link then it can
 483          * not be aggregated.
 484          */
 485         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 486             ifp->if_link_state != LINK_STATE_UP) {
 487                 lacp_port_disable(lp);
 488         } else {
 489                 lacp_port_enable(lp);
 490         }
 491         lp->lp_key = lacp_compose_key(lp);
 492
 493         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 494                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 495                 lp->lp_selected = LACP_UNSELECTED;
 496         }
 497         LACP_UNLOCK(lsc);
 498 }
 499
 500 static void
 501 lacp_tick(void *arg)
 502 {
 503         struct lacp_softc *lsc = arg;
 504         struct lacp_port *lp;
 505
 506         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 507                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 508                         continue;
 509
 510                 CURVNET_SET(lp->lp_ifp->if_vnet);
 511                 lacp_run_timers(lp);
 512
 513                 lacp_select(lp);
 514                 lacp_sm_mux(lp);
 515                 lacp_sm_tx(lp);
 516                 lacp_sm_ptx_tx_schedule(lp);
 517                 CURVNET_RESTORE();
 518         }
 519         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 520 }
 521
 522 int
 523 lacp_port_create(struct lagg_port *lgp)
 524 {
 525         struct lagg_softc *sc = lgp->lp_softc;
 526         struct lacp_softc *lsc = LACP_SOFTC(sc);
 527         struct lacp_port *lp;
 528         struct ifnet *ifp = lgp->lp_ifp;
 529         struct sockaddr_dl sdl;
 530         struct ifmultiaddr *rifma = NULL;
 531         int error;
 532
 533         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 534         sdl.sdl_alen = ETHER_ADDR_LEN;
 535
 536         bcopy(&ethermulticastaddr_slowprotocols,
 537             LLADDR(&sdl), ETHER_ADDR_LEN);
 538         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 539         if (error) {
 540                 printf("%s: ADDMULTI failed on %s\n", __func__,
 541                     lgp->lp_ifp->if_xname);
 542                 return (error);
 543         }
 544
 545         lp = malloc(sizeof(struct lacp_port),
 546             M_DEVBUF, M_NOWAIT|M_ZERO);
 547         if (lp == NULL)
 548                 return (ENOMEM);
 549
 550         LACP_LOCK(lsc);
 551         lgp->lp_psc = lp;
 552         lp->lp_ifp = ifp;
 553         lp->lp_lagg = lgp;
 554         lp->lp_lsc = lsc;
 555         lp->lp_ifma = rifma;
 556
 557         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 558
 559         lacp_fill_actorinfo(lp, &lp->lp_actor);
 560         lacp_fill_markerinfo(lp, &lp->lp_marker);
 561         lp->lp_state = LACP_STATE_ACTIVITY;
 562         lp->lp_aggregator = NULL;
 563         lacp_sm_rx_set_expired(lp);
 564         LACP_UNLOCK(lsc);
 565         lacp_linkstate(lgp);
 566
 567         return (0);
 568 }
 569
 570 void
 571 lacp_port_destroy(struct lagg_port *lgp)
 572 {
 573         struct lacp_port *lp = LACP_PORT(lgp);
 574         struct lacp_softc *lsc = lp->lp_lsc;
 575         int i;
 576
 577         LACP_LOCK(lsc);
 578         for (i = 0; i < LACP_NTIMER; i++) {
 579                 LACP_TIMER_DISARM(lp, i);
 580         }
 581
 582         lacp_disable_collecting(lp);
 583         lacp_disable_distributing(lp);
 584         lacp_unselect(lp);
 585
 586         LIST_REMOVE(lp, lp_next);
 587         LACP_UNLOCK(lsc);
 588
 589         /* The address may have already been removed by if_purgemaddrs() */
 590         if (!lgp->lp_detaching)
 591                 if_delmulti_ifma(lp->lp_ifma);
 592
 593         free(lp, M_DEVBUF);
 594 }
 595
 596 void
 597 lacp_req(struct lagg_softc *sc, void *data)
 598 {
 599         struct lacp_opreq *req = (struct lacp_opreq *)data;
 600         struct lacp_softc *lsc = LACP_SOFTC(sc);
 601         struct lacp_aggregator *la;
 602
 603         bzero(req, sizeof(struct lacp_opreq));
 604
 605         /*
 606          * If the LACP softc is NULL, return with the opreq structure full of
 607          * zeros.  It is normal for the softc to be NULL while the lagg is
 608          * being destroyed.
 609          */
 610         if (NULL == lsc)
 611                 return;
 612
 613         la = lsc->lsc_active_aggregator;
 614         LACP_LOCK(lsc);
 615         if (la != NULL) {
 616                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 617                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 618                     ETHER_ADDR_LEN);
 619                 req->actor_key = ntohs(la->la_actor.lip_key);
 620                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 621                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 622                 req->actor_state = la->la_actor.lip_state;
 623
 624                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 625                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 626                     ETHER_ADDR_LEN);
 627                 req->partner_key = ntohs(la->la_partner.lip_key);
 628                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 629                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 630                 req->partner_state = la->la_partner.lip_state;
 631         }
 632         LACP_UNLOCK(lsc);
 633 }
 634
 635 void
 636 lacp_portreq(struct lagg_port *lgp, void *data)
 637 {
 638         struct lacp_opreq *req = (struct lacp_opreq *)data;
 639         struct lacp_port *lp = LACP_PORT(lgp);
 640         struct lacp_softc *lsc = lp->lp_lsc;
 641
 642         LACP_LOCK(lsc);
 643         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 644         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 645             ETHER_ADDR_LEN);
 646         req->actor_key = ntohs(lp->lp_actor.lip_key);
 647         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 648         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 649         req->actor_state = lp->lp_actor.lip_state;
 650
 651         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 652         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 653             ETHER_ADDR_LEN);
 654         req->partner_key = ntohs(lp->lp_partner.lip_key);
 655         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 656         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 657         req->partner_state = lp->lp_partner.lip_state;
 658         LACP_UNLOCK(lsc);
 659 }
 660
 661 static void
 662 lacp_disable_collecting(struct lacp_port *lp)
 663 {
 664         LACP_DPRINTF((lp, "collecting disabled\n"));
 665         lp->lp_state &= ~LACP_STATE_COLLECTING;
 666 }
 667
 668 static void
 669 lacp_enable_collecting(struct lacp_port *lp)
 670 {
 671         LACP_DPRINTF((lp, "collecting enabled\n"));
 672         lp->lp_state |= LACP_STATE_COLLECTING;
 673 }
 674
 675 static void
 676 lacp_disable_distributing(struct lacp_port *lp)
 677 {
 678         struct lacp_aggregator *la = lp->lp_aggregator;
 679         struct lacp_softc *lsc = lp->lp_lsc;
 680         struct lagg_softc *sc = lsc->lsc_softc;
 681         char buf[LACP_LAGIDSTR_MAX+1];
 682
 683         LACP_LOCK_ASSERT(lsc);
 684
 685         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 686                 return;
 687         }
 688
 689         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 690         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 691         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 692
 693         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 694             "nports %d -> %d\n",
 695             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 696             la->la_nports, la->la_nports - 1));
 697
 698         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 699         la->la_nports--;
 700         sc->sc_active = la->la_nports;
 701
 702         if (lsc->lsc_active_aggregator == la) {
 703                 lacp_suppress_distributing(lsc, la);
 704                 lacp_select_active_aggregator(lsc);
 705                 /* regenerate the port map, the active aggregator has changed */
 706                 lacp_update_portmap(lsc);
 707         }
 708
 709         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 710 }
 711
 712 static void
 713 lacp_enable_distributing(struct lacp_port *lp)
 714 {
 715         struct lacp_aggregator *la = lp->lp_aggregator;
 716         struct lacp_softc *lsc = lp->lp_lsc;
 717         struct lagg_softc *sc = lsc->lsc_softc;
 718         char buf[LACP_LAGIDSTR_MAX+1];
 719
 720         LACP_LOCK_ASSERT(lsc);
 721
 722         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 723                 return;
 724         }
 725
 726         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 727             "nports %d -> %d\n",
 728             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 729             la->la_nports, la->la_nports + 1));
 730
 731         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 732         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 733         la->la_nports++;
 734         sc->sc_active = la->la_nports;
 735
 736         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 737
 738         if (lsc->lsc_active_aggregator == la) {
 739                 lacp_suppress_distributing(lsc, la);
 740                 lacp_update_portmap(lsc);
 741         } else
 742                 /* try to become the active aggregator */
 743                 lacp_select_active_aggregator(lsc);
 744 }
 745
 746 static void
 747 lacp_transit_expire(void *vp)
 748 {
 749         struct lacp_softc *lsc = vp;
 750
 751         LACP_LOCK_ASSERT(lsc);
 752
 753         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 754         LACP_TRACE(NULL);
 755         CURVNET_RESTORE();
 756
 757         lsc->lsc_suppress_distributing = FALSE;
 758 }
 759
 760 void
 761 lacp_attach(struct lagg_softc *sc)
 762 {
 763         struct lacp_softc *lsc;
 764
 765         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 766
 767         sc->sc_psc = lsc;
 768         lsc->lsc_softc = sc;
 769
 770         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 771         lsc->lsc_active_aggregator = NULL;
 772         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 773         LACP_LOCK_INIT(lsc);
 774         TAILQ_INIT(&lsc->lsc_aggregators);
 775         LIST_INIT(&lsc->lsc_ports);
 776
 777         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 778         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 779
 780         /* if the lagg is already up then do the same */
 781         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 782                 lacp_init(sc);
 783 }
 784
 785 void
 786 lacp_detach(void *psc)
 787 {
 788         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 789
 790         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 791             ("aggregators still active"));
 792         KASSERT(lsc->lsc_active_aggregator == NULL,
 793             ("aggregator still attached"));
 794
 795         callout_drain(&lsc->lsc_transit_callout);
 796         callout_drain(&lsc->lsc_callout);
 797
 798         LACP_LOCK_DESTROY(lsc);
 799         free(lsc, M_DEVBUF);
 800 }
 801
 802 void
 803 lacp_init(struct lagg_softc *sc)
 804 {
 805         struct lacp_softc *lsc = LACP_SOFTC(sc);
 806
 807         LACP_LOCK(lsc);
 808         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 809         LACP_UNLOCK(lsc);
 810 }
 811
 812 void
 813 lacp_stop(struct lagg_softc *sc)
 814 {
 815         struct lacp_softc *lsc = LACP_SOFTC(sc);
 816
 817         LACP_LOCK(lsc);
 818         callout_stop(&lsc->lsc_transit_callout);
 819         callout_stop(&lsc->lsc_callout);
 820         LACP_UNLOCK(lsc);
 821 }
 822
 823 struct lagg_port *
 824 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 825 {
 826         struct lacp_softc *lsc = LACP_SOFTC(sc);
 827         struct lacp_portmap *pm;
 828         struct lacp_port *lp;
 829         uint32_t hash;
 830
 831         if (__predict_false(lsc->lsc_suppress_distributing)) {
 832                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 833                 return (NULL);
 834         }
 835
 836         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 837         if (pm->pm_count == 0) {
 838                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 839                 return (NULL);
 840         }
 841
 842         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 843             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 844                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 845         else
 846                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 847         hash %= pm->pm_count;
 848         lp = pm->pm_map[hash];
 849
 850         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 851             ("aggregated port is not distributing"));
 852
 853         return (lp->lp_lagg);
 854 }
 855 /*
 856  * lacp_suppress_distributing: drop transmit packets for a while
 857  * to preserve packet ordering.
 858  */
 859
 860 static void
 861 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 862 {
 863         struct lacp_port *lp;
 864
 865         if (lsc->lsc_active_aggregator != la) {
 866                 return;
 867         }
 868
 869         LACP_TRACE(NULL);
 870
 871         lsc->lsc_suppress_distributing = TRUE;
 872
 873         /* send a marker frame down each port to verify the queues are empty */
 874         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 875                 lp->lp_flags |= LACP_PORT_MARK;
 876                 lacp_xmit_marker(lp);
 877         }
 878
 879         /* set a timeout for the marker frames */
 880         callout_reset(&lsc->lsc_transit_callout,
 881             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 882 }
 883
 884 static int
 885 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 886     const struct lacp_peerinfo *b)
 887 {
 888         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 889 }
 890
 891 static int
 892 lacp_compare_systemid(const struct lacp_systemid *a,
 893     const struct lacp_systemid *b)
 894 {
 895         return (memcmp(a, b, sizeof(*a)));
 896 }
 897
 898 #if 0   /* unused */
 899 static int
 900 lacp_compare_portid(const struct lacp_portid *a,
 901     const struct lacp_portid *b)
 902 {
 903         return (memcmp(a, b, sizeof(*a)));
 904 }
 905 #endif
 906
 907 static uint64_t
 908 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 909 {
 910         struct lacp_port *lp;
 911         uint64_t speed;
 912
 913         lp = TAILQ_FIRST(&la->la_ports);
 914         if (lp == NULL) {
 915                 return (0);
 916         }
 917
 918         speed = ifmedia_baudrate(lp->lp_media);
 919         speed *= la->la_nports;
 920         if (speed == 0) {
 921                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 922                     lp->lp_media, la->la_nports));
 923         }
 924
 925         return (speed);
 926 }
 927
 928 /*
 929  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 930  * packets from lagg(4) interface.
 931  */
 932
 933 static void
 934 lacp_select_active_aggregator(struct lacp_softc *lsc)
 935 {
 936         struct lacp_aggregator *la;
 937         struct lacp_aggregator *best_la = NULL;
 938         uint64_t best_speed = 0;
 939         char buf[LACP_LAGIDSTR_MAX+1];
 940
 941         LACP_TRACE(NULL);
 942
 943         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 944                 uint64_t speed;
 945
 946                 if (la->la_nports == 0) {
 947                         continue;
 948                 }
 949
 950                 speed = lacp_aggregator_bandwidth(la);
 951                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 952                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 953                     speed, la->la_nports));
 954
 955                 /*
 956                  * This aggregator is chosen if the partner has a better
 957                  * system priority or, the total aggregated speed is higher
 958                  * or, it is already the chosen aggregator
 959                  */
 960                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 961                     LACP_SYS_PRI(best_la->la_partner)) ||
 962                     speed > best_speed ||
 963                     (speed == best_speed &&
 964                     la == lsc->lsc_active_aggregator)) {
 965                         best_la = la;
 966                         best_speed = speed;
 967                 }
 968         }
 969
 970         KASSERT(best_la == NULL || best_la->la_nports > 0,
 971             ("invalid aggregator refcnt"));
 972         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
 973             ("invalid aggregator list"));
 974
 975         if (lsc->lsc_active_aggregator != best_la) {
 976                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
 977                 LACP_DPRINTF((NULL, "old %s\n",
 978                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
 979                     buf, sizeof(buf))));
 980         } else {
 981                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
 982         }
 983         LACP_DPRINTF((NULL, "new %s\n",
 984             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
 985
 986         if (lsc->lsc_active_aggregator != best_la) {
 987                 lsc->lsc_active_aggregator = best_la;
 988                 lacp_update_portmap(lsc);
 989                 if (best_la) {
 990                         lacp_suppress_distributing(lsc, best_la);
 991                 }
 992         }
 993 }
 994
 995 /*
 996  * Updated the inactive portmap array with the new list of ports and
 997  * make it live.
 998  */
 999 static void
1000 lacp_update_portmap(struct lacp_softc *lsc)
1001 {
1002         struct lagg_softc *sc = lsc->lsc_softc;
1003         struct lacp_aggregator *la;
1004         struct lacp_portmap *p;
1005         struct lacp_port *lp;
1006         uint64_t speed;
1007         u_int newmap;
1008         int i;
1009
1010         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1011         p = &lsc->lsc_pmap[newmap];
1012         la = lsc->lsc_active_aggregator;
1013         speed = 0;
1014         bzero(p, sizeof(struct lacp_portmap));
1015
1016         if (la != NULL && la->la_nports > 0) {
1017                 p->pm_count = la->la_nports;
1018                 i = 0;
1019                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1020                         p->pm_map[i++] = lp;
1021                 KASSERT(i == p->pm_count, ("Invalid port count"));
1022                 speed = lacp_aggregator_bandwidth(la);
1023         }
1024         sc->sc_ifp->if_baudrate = speed;
1025
1026         /* switch the active portmap over */
1027         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1028         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1029                     lsc->lsc_activemap,
1030                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1031 }
1032
1033 static uint16_t
1034 lacp_compose_key(struct lacp_port *lp)
1035 {
1036         struct lagg_port *lgp = lp->lp_lagg;
1037         struct lagg_softc *sc = lgp->lp_softc;
1038         u_int media = lp->lp_media;
1039         uint16_t key;
1040
1041         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1042
1043                 /*
1044                  * non-aggregatable links should have unique keys.
1045                  *
1046                  * XXX this isn't really unique as if_index is 16 bit.
1047                  */
1048
1049                 /* bit 0..14:   (some bits of) if_index of this port */
1050                 key = lp->lp_ifp->if_index;
1051                 /* bit 15:      1 */
1052                 key |= 0x8000;
1053         } else {
1054                 u_int subtype = IFM_SUBTYPE(media);
1055
1056                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1057                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1058
1059                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1060                 switch (subtype) {
1061                 case IFM_10_T:
1062                 case IFM_10_2:
1063                 case IFM_10_5:
1064                 case IFM_10_STP:
1065                 case IFM_10_FL:
1066                         key = IFM_10_T;
1067                         break;
1068                 case IFM_100_TX:
1069                 case IFM_100_FX:
1070                 case IFM_100_T4:
1071                 case IFM_100_VG:
1072                 case IFM_100_T2:
1073                 case IFM_100_T:
1074                 case IFM_100_SGMII:
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                 case IFM_10G_AOC:
1099                         key = IFM_10G_LR;
1100                         break;
1101                 case IFM_20G_KR2:
1102                         key = IFM_20G_KR2;
1103                         break;
1104                 case IFM_2500_KX:
1105                 case IFM_2500_T:
1106                 case IFM_2500_X:
1107                         key = IFM_2500_KX;
1108                         break;
1109                 case IFM_5000_T:
1110                 case IFM_5000_KR:
1111                 case IFM_5000_KR_S:
1112                 case IFM_5000_KR1:
1113                         key = IFM_5000_T;
1114                         break;
1115                 case IFM_50G_PCIE:
1116                 case IFM_50G_CR2:
1117                 case IFM_50G_KR2:
1118                 case IFM_50G_SR2:
1119                 case IFM_50G_LR2:
1120                 case IFM_50G_LAUI2_AC:
1121                 case IFM_50G_LAUI2:
1122                 case IFM_50G_AUI2_AC:
1123                 case IFM_50G_AUI2:
1124                 case IFM_50G_CP:
1125                 case IFM_50G_SR:
1126                 case IFM_50G_LR:
1127                 case IFM_50G_FR:
1128                 case IFM_50G_KR_PAM4:
1129                 case IFM_50G_AUI1_AC:
1130                 case IFM_50G_AUI1:
1131                         key = IFM_50G_PCIE;
1132                         break;
1133                 case IFM_56G_R4:
1134                         key = IFM_56G_R4;
1135                         break;
1136                 case IFM_25G_PCIE:
1137                 case IFM_25G_CR:
1138                 case IFM_25G_KR:
1139                 case IFM_25G_SR:
1140                 case IFM_25G_LR:
1141                 case IFM_25G_ACC:
1142                 case IFM_25G_AOC:
1143                 case IFM_25G_T:
1144                 case IFM_25G_CR_S:
1145                 case IFM_25G_CR1:
1146                 case IFM_25G_KR_S:
1147                 case IFM_25G_AUI:
1148                 case IFM_25G_KR1:
1149                         key = IFM_25G_PCIE;
1150                         break;
1151                 case IFM_40G_CR4:
1152                 case IFM_40G_SR4:
1153                 case IFM_40G_LR4:
1154                 case IFM_40G_XLPPI:
1155                 case IFM_40G_KR4:
1156                 case IFM_40G_XLAUI:
1157                 case IFM_40G_XLAUI_AC:
1158                 case IFM_40G_ER4:
1159                         key = IFM_40G_CR4;
1160                         break;
1161                 case IFM_100G_CR4:
1162                 case IFM_100G_SR4:
1163                 case IFM_100G_KR4:
1164                 case IFM_100G_LR4:
1165                 case IFM_100G_CAUI4_AC:
1166                 case IFM_100G_CAUI4:
1167                 case IFM_100G_AUI4_AC:
1168                 case IFM_100G_AUI4:
1169                 case IFM_100G_CR_PAM4:
1170                 case IFM_100G_KR_PAM4:
1171                 case IFM_100G_CP2:
1172                 case IFM_100G_SR2:
1173                 case IFM_100G_DR:
1174                 case IFM_100G_KR2_PAM4:
1175                 case IFM_100G_CAUI2_AC:
1176                 case IFM_100G_CAUI2:
1177                 case IFM_100G_AUI2_AC:
1178                 case IFM_100G_AUI2:
1179                         key = IFM_100G_CR4;
1180                         break;
1181                 case IFM_200G_CR4_PAM4:
1182                 case IFM_200G_SR4:
1183                 case IFM_200G_FR4:
1184                 case IFM_200G_LR4:
1185                 case IFM_200G_DR4:
1186                 case IFM_200G_KR4_PAM4:
1187                 case IFM_200G_AUI4_AC:
1188                 case IFM_200G_AUI4:
1189                 case IFM_200G_AUI8_AC:
1190                 case IFM_200G_AUI8:
1191                         key = IFM_200G_CR4_PAM4;
1192                         break;
1193                 case IFM_400G_FR8:
1194                 case IFM_400G_LR8:
1195                 case IFM_400G_DR4:
1196                 case IFM_400G_AUI8_AC:
1197                 case IFM_400G_AUI8:
1198                         key = IFM_400G_FR8;
1199                         break;
1200                 default:
1201                         key = subtype;
1202                         break;
1203                 }
1204                 /* bit 5..14:   (some bits of) if_index of lagg device */
1205                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1206                 /* bit 15:      0 */
1207         }
1208         return (htons(key));
1209 }
1210
1211 static void
1212 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1213 {
1214         char buf[LACP_LAGIDSTR_MAX+1];
1215
1216         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1217             __func__,
1218             lacp_format_lagid(&la->la_actor, &la->la_partner,
1219             buf, sizeof(buf)),
1220             la->la_refcnt, la->la_refcnt + 1));
1221
1222         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1223         la->la_refcnt++;
1224         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1225 }
1226
1227 static void
1228 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1229 {
1230         char buf[LACP_LAGIDSTR_MAX+1];
1231
1232         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1233             __func__,
1234             lacp_format_lagid(&la->la_actor, &la->la_partner,
1235             buf, sizeof(buf)),
1236             la->la_refcnt, la->la_refcnt - 1));
1237
1238         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1239         la->la_refcnt--;
1240         if (la->la_refcnt > 0) {
1241                 return;
1242         }
1243
1244         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1245         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1246
1247         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1248
1249         free(la, M_DEVBUF);
1250 }
1251
1252 /*
1253  * lacp_aggregator_get: allocate an aggregator.
1254  */
1255
1256 static struct lacp_aggregator *
1257 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1258 {
1259         struct lacp_aggregator *la;
1260
1261         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1262         if (la) {
1263                 la->la_refcnt = 1;
1264                 la->la_nports = 0;
1265                 TAILQ_INIT(&la->la_ports);
1266                 la->la_pending = 0;
1267                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1268         }
1269
1270         return (la);
1271 }
1272
1273 /*
1274  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1275  */
1276
1277 static void
1278 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1279 {
1280         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1281         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1282
1283         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1284 }
1285
1286 static void
1287 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1288     const struct lacp_peerinfo *lpi_port)
1289 {
1290         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1291         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1292         lpi_aggr->lip_key = lpi_port->lip_key;
1293 }
1294
1295 /*
1296  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1297  */
1298
1299 static int
1300 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1301     const struct lacp_port *lp)
1302 {
1303         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1304             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1305                 return (0);
1306         }
1307
1308         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1309                 return (0);
1310         }
1311
1312         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1313                 return (0);
1314         }
1315
1316         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1317                 return (0);
1318         }
1319
1320         return (1);
1321 }
1322
1323 static int
1324 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1325     const struct lacp_peerinfo *b)
1326 {
1327         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1328             sizeof(a->lip_systemid))) {
1329                 return (0);
1330         }
1331
1332         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1333                 return (0);
1334         }
1335
1336         return (1);
1337 }
1338
1339 static void
1340 lacp_port_enable(struct lacp_port *lp)
1341 {
1342         lp->lp_state |= LACP_STATE_AGGREGATION;
1343 }
1344
1345 static void
1346 lacp_port_disable(struct lacp_port *lp)
1347 {
1348         lacp_set_mux(lp, LACP_MUX_DETACHED);
1349
1350         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1351         lp->lp_selected = LACP_UNSELECTED;
1352         lacp_sm_rx_record_default(lp);
1353         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1354         lp->lp_state &= ~LACP_STATE_EXPIRED;
1355 }
1356
1357 /*
1358  * lacp_select: select an aggregator.  create one if necessary.
1359  */
1360 static void
1361 lacp_select(struct lacp_port *lp)
1362 {
1363         struct lacp_softc *lsc = lp->lp_lsc;
1364         struct lacp_aggregator *la;
1365         char buf[LACP_LAGIDSTR_MAX+1];
1366
1367         if (lp->lp_aggregator) {
1368                 return;
1369         }
1370
1371         /* If we haven't heard from our peer, skip this step. */
1372         if (lp->lp_state & LACP_STATE_DEFAULTED)
1373                 return;
1374
1375         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1376             ("timer_wait_while still active"));
1377
1378         LACP_DPRINTF((lp, "port lagid=%s\n",
1379             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1380             buf, sizeof(buf))));
1381
1382         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1383                 if (lacp_aggregator_is_compatible(la, lp)) {
1384                         break;
1385                 }
1386         }
1387
1388         if (la == NULL) {
1389                 la = lacp_aggregator_get(lsc, lp);
1390                 if (la == NULL) {
1391                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1392
1393                         /*
1394                          * will retry on the next tick.
1395                          */
1396
1397                         return;
1398                 }
1399                 lacp_fill_aggregator_id(la, lp);
1400                 LACP_DPRINTF((lp, "aggregator created\n"));
1401         } else {
1402                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1403                 if (la->la_refcnt == LACP_MAX_PORTS)
1404                         return;
1405                 lacp_aggregator_addref(lsc, la);
1406         }
1407
1408         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1409             lacp_format_lagid(&la->la_actor, &la->la_partner,
1410             buf, sizeof(buf))));
1411
1412         lp->lp_aggregator = la;
1413         lp->lp_selected = LACP_SELECTED;
1414 }
1415
1416 /*
1417  * lacp_unselect: finish unselect/detach process.
1418  */
1419
1420 static void
1421 lacp_unselect(struct lacp_port *lp)
1422 {
1423         struct lacp_softc *lsc = lp->lp_lsc;
1424         struct lacp_aggregator *la = lp->lp_aggregator;
1425
1426         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1427             ("timer_wait_while still active"));
1428
1429         if (la == NULL) {
1430                 return;
1431         }
1432
1433         lp->lp_aggregator = NULL;
1434         lacp_aggregator_delref(lsc, la);
1435 }
1436
1437 /* mux machine */
1438
1439 static void
1440 lacp_sm_mux(struct lacp_port *lp)
1441 {
1442         struct lagg_port *lgp = lp->lp_lagg;
1443         struct lagg_softc *sc = lgp->lp_softc;
1444         enum lacp_mux_state new_state;
1445         boolean_t p_sync =
1446                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1447         boolean_t p_collecting =
1448             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1449         enum lacp_selected selected = lp->lp_selected;
1450         struct lacp_aggregator *la;
1451
1452         if (V_lacp_debug > 1)
1453                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1454                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1455                     lp->lp_mux_state, selected, p_sync, p_collecting);
1456
1457 re_eval:
1458         la = lp->lp_aggregator;
1459         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1460             ("MUX not detached"));
1461         new_state = lp->lp_mux_state;
1462         switch (lp->lp_mux_state) {
1463         case LACP_MUX_DETACHED:
1464                 if (selected != LACP_UNSELECTED) {
1465                         new_state = LACP_MUX_WAITING;
1466                 }
1467                 break;
1468         case LACP_MUX_WAITING:
1469                 KASSERT(la->la_pending > 0 ||
1470                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1471                     ("timer_wait_while still active"));
1472                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1473                         new_state = LACP_MUX_ATTACHED;
1474                 } else if (selected == LACP_UNSELECTED) {
1475                         new_state = LACP_MUX_DETACHED;
1476                 }
1477                 break;
1478         case LACP_MUX_ATTACHED:
1479                 if (selected == LACP_SELECTED && p_sync) {
1480                         new_state = LACP_MUX_COLLECTING;
1481                 } else if (selected != LACP_SELECTED) {
1482                         new_state = LACP_MUX_DETACHED;
1483                 }
1484                 break;
1485         case LACP_MUX_COLLECTING:
1486                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1487                         new_state = LACP_MUX_DISTRIBUTING;
1488                 } else if (selected != LACP_SELECTED || !p_sync) {
1489                         new_state = LACP_MUX_ATTACHED;
1490                 }
1491                 break;
1492         case LACP_MUX_DISTRIBUTING:
1493                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1494                         new_state = LACP_MUX_COLLECTING;
1495                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1496                         sc->sc_flapping++;
1497                 }
1498                 break;
1499         default:
1500                 panic("%s: unknown state", __func__);
1501         }
1502
1503         if (lp->lp_mux_state == new_state) {
1504                 return;
1505         }
1506
1507         lacp_set_mux(lp, new_state);
1508         goto re_eval;
1509 }
1510
1511 static void
1512 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1513 {
1514         struct lacp_aggregator *la = lp->lp_aggregator;
1515
1516         if (lp->lp_mux_state == new_state) {
1517                 return;
1518         }
1519
1520         switch (new_state) {
1521         case LACP_MUX_DETACHED:
1522                 lp->lp_state &= ~LACP_STATE_SYNC;
1523                 lacp_disable_distributing(lp);
1524                 lacp_disable_collecting(lp);
1525                 lacp_sm_assert_ntt(lp);
1526                 /* cancel timer */
1527                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1528                         KASSERT(la->la_pending > 0,
1529                             ("timer_wait_while not active"));
1530                         la->la_pending--;
1531                 }
1532                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1533                 lacp_unselect(lp);
1534                 break;
1535         case LACP_MUX_WAITING:
1536                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1537                     LACP_AGGREGATE_WAIT_TIME);
1538                 la->la_pending++;
1539                 break;
1540         case LACP_MUX_ATTACHED:
1541                 lp->lp_state |= LACP_STATE_SYNC;
1542                 lacp_disable_collecting(lp);
1543                 lacp_sm_assert_ntt(lp);
1544                 break;
1545         case LACP_MUX_COLLECTING:
1546                 lacp_enable_collecting(lp);
1547                 lacp_disable_distributing(lp);
1548                 lacp_sm_assert_ntt(lp);
1549                 break;
1550         case LACP_MUX_DISTRIBUTING:
1551                 lacp_enable_distributing(lp);
1552                 break;
1553         default:
1554                 panic("%s: unknown state", __func__);
1555         }
1556
1557         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1558
1559         lp->lp_mux_state = new_state;
1560 }
1561
1562 static void
1563 lacp_sm_mux_timer(struct lacp_port *lp)
1564 {
1565         struct lacp_aggregator *la = lp->lp_aggregator;
1566         char buf[LACP_LAGIDSTR_MAX+1];
1567
1568         KASSERT(la->la_pending > 0, ("no pending event"));
1569
1570         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1571             lacp_format_lagid(&la->la_actor, &la->la_partner,
1572             buf, sizeof(buf)),
1573             la->la_pending, la->la_pending - 1));
1574
1575         la->la_pending--;
1576 }
1577
1578 /* periodic transmit machine */
1579
1580 static void
1581 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1582 {
1583         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1584             LACP_STATE_TIMEOUT)) {
1585                 return;
1586         }
1587
1588         LACP_DPRINTF((lp, "partner timeout changed\n"));
1589
1590         /*
1591          * FAST_PERIODIC -> SLOW_PERIODIC
1592          * or
1593          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1594          *
1595          * let lacp_sm_ptx_tx_schedule to update timeout.
1596          */
1597
1598         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1599
1600         /*
1601          * if timeout has been shortened, assert NTT.
1602          */
1603
1604         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1605                 lacp_sm_assert_ntt(lp);
1606         }
1607 }
1608
1609 static void
1610 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1611 {
1612         int timeout;
1613
1614         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1615             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1616
1617                 /*
1618                  * NO_PERIODIC
1619                  */
1620
1621                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1622                 return;
1623         }
1624
1625         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1626                 return;
1627         }
1628
1629         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1630             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1631
1632         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1633 }
1634
1635 static void
1636 lacp_sm_ptx_timer(struct lacp_port *lp)
1637 {
1638         lacp_sm_assert_ntt(lp);
1639 }
1640
1641 static void
1642 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1643 {
1644         int timeout;
1645
1646         /*
1647          * check LACP_DISABLED first
1648          */
1649
1650         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1651                 return;
1652         }
1653
1654         /*
1655          * check loopback condition.
1656          */
1657
1658         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1659             &lp->lp_actor.lip_systemid)) {
1660                 return;
1661         }
1662
1663         /*
1664          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1665          */
1666
1667         lacp_sm_rx_update_selected(lp, du);
1668         lacp_sm_rx_update_ntt(lp, du);
1669         lacp_sm_rx_record_pdu(lp, du);
1670
1671         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1672             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1673         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1674
1675         lp->lp_state &= ~LACP_STATE_EXPIRED;
1676
1677         /*
1678          * kick transmit machine without waiting the next tick.
1679          */
1680
1681         lacp_sm_tx(lp);
1682 }
1683
1684 static void
1685 lacp_sm_rx_set_expired(struct lacp_port *lp)
1686 {
1687         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1688         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1689         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1690         lp->lp_state |= LACP_STATE_EXPIRED;
1691 }
1692
1693 static void
1694 lacp_sm_rx_timer(struct lacp_port *lp)
1695 {
1696         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1697                 /* CURRENT -> EXPIRED */
1698                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1699                 lacp_sm_rx_set_expired(lp);
1700         } else {
1701                 /* EXPIRED -> DEFAULTED */
1702                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1703                 lacp_sm_rx_update_default_selected(lp);
1704                 lacp_sm_rx_record_default(lp);
1705                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1706         }
1707 }
1708
1709 static void
1710 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1711 {
1712         boolean_t active;
1713         uint8_t oldpstate;
1714         char buf[LACP_STATESTR_MAX+1];
1715
1716         LACP_TRACE(lp);
1717
1718         oldpstate = lp->lp_partner.lip_state;
1719
1720         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1721             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1722             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1723
1724         lp->lp_partner = du->ldu_actor;
1725         if (active &&
1726             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1727             LACP_STATE_AGGREGATION) &&
1728             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1729             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1730                 /*
1731                  * XXX Maintain legacy behavior of leaving the
1732                  * LACP_STATE_SYNC bit unchanged from the partner's
1733                  * advertisement if lsc_strict_mode is false.
1734                  * TODO: We should re-examine the concept of the "strict mode"
1735                  * to ensure it makes sense to maintain a non-strict mode.
1736                  */
1737                 if (lp->lp_lsc->lsc_strict_mode)
1738                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1739         } else {
1740                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1741         }
1742
1743         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1744
1745         if (oldpstate != lp->lp_partner.lip_state) {
1746                 LACP_DPRINTF((lp, "old pstate %s\n",
1747                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1748                 LACP_DPRINTF((lp, "new pstate %s\n",
1749                     lacp_format_state(lp->lp_partner.lip_state, buf,
1750                     sizeof(buf))));
1751         }
1752
1753         lacp_sm_ptx_update_timeout(lp, oldpstate);
1754 }
1755
1756 static void
1757 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1758 {
1759
1760         LACP_TRACE(lp);
1761
1762         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1763             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1764             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1765                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1766                 lacp_sm_assert_ntt(lp);
1767         }
1768 }
1769
1770 static void
1771 lacp_sm_rx_record_default(struct lacp_port *lp)
1772 {
1773         uint8_t oldpstate;
1774
1775         LACP_TRACE(lp);
1776
1777         oldpstate = lp->lp_partner.lip_state;
1778         if (lp->lp_lsc->lsc_strict_mode)
1779                 lp->lp_partner = lacp_partner_admin_strict;
1780         else
1781                 lp->lp_partner = lacp_partner_admin_optimistic;
1782         lp->lp_state |= LACP_STATE_DEFAULTED;
1783         lacp_sm_ptx_update_timeout(lp, oldpstate);
1784 }
1785
1786 static void
1787 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1788     const struct lacp_peerinfo *info)
1789 {
1790
1791         LACP_TRACE(lp);
1792
1793         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1794             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1795             LACP_STATE_AGGREGATION)) {
1796                 lp->lp_selected = LACP_UNSELECTED;
1797                 /* mux machine will clean up lp->lp_aggregator */
1798         }
1799 }
1800
1801 static void
1802 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1803 {
1804
1805         LACP_TRACE(lp);
1806
1807         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1808 }
1809
1810 static void
1811 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1812 {
1813
1814         LACP_TRACE(lp);
1815
1816         if (lp->lp_lsc->lsc_strict_mode)
1817                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1818                     &lacp_partner_admin_strict);
1819         else
1820                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1821                     &lacp_partner_admin_optimistic);
1822 }
1823
1824 /* transmit machine */
1825
1826 static void
1827 lacp_sm_tx(struct lacp_port *lp)
1828 {
1829         int error = 0;
1830
1831         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1832 #if 1
1833             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1834             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1835 #endif
1836             ) {
1837                 lp->lp_flags &= ~LACP_PORT_NTT;
1838         }
1839
1840         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1841                 return;
1842         }
1843
1844         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1845         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1846                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1847                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1848                 return;
1849         }
1850
1851         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1852                 error = lacp_xmit_lacpdu(lp);
1853         } else {
1854                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1855         }
1856
1857         if (error == 0) {
1858                 lp->lp_flags &= ~LACP_PORT_NTT;
1859         } else {
1860                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1861                     error));
1862         }
1863 }
1864
1865 static void
1866 lacp_sm_assert_ntt(struct lacp_port *lp)
1867 {
1868
1869         lp->lp_flags |= LACP_PORT_NTT;
1870 }
1871
1872 static void
1873 lacp_run_timers(struct lacp_port *lp)
1874 {
1875         int i;
1876
1877         for (i = 0; i < LACP_NTIMER; i++) {
1878                 KASSERT(lp->lp_timer[i] >= 0,
1879                     ("invalid timer value %d", lp->lp_timer[i]));
1880                 if (lp->lp_timer[i] == 0) {
1881                         continue;
1882                 } else if (--lp->lp_timer[i] <= 0) {
1883                         if (lacp_timer_funcs[i]) {
1884                                 (*lacp_timer_funcs[i])(lp);
1885                         }
1886                 }
1887         }
1888 }
1889
1890 int
1891 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1892 {
1893         struct lacp_softc *lsc = lp->lp_lsc;
1894         struct lagg_port *lgp = lp->lp_lagg;
1895         struct lacp_port *lp2;
1896         struct markerdu *mdu;
1897         int error = 0;
1898         int pending = 0;
1899
1900         if (m->m_pkthdr.len != sizeof(*mdu)) {
1901                 goto bad;
1902         }
1903
1904         if ((m->m_flags & M_MCAST) == 0) {
1905                 goto bad;
1906         }
1907
1908         if (m->m_len < sizeof(*mdu)) {
1909                 m = m_pullup(m, sizeof(*mdu));
1910                 if (m == NULL) {
1911                         return (ENOMEM);
1912                 }
1913         }
1914
1915         mdu = mtod(m, struct markerdu *);
1916
1917         if (memcmp(&mdu->mdu_eh.ether_dhost,
1918             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1919                 goto bad;
1920         }
1921
1922         if (mdu->mdu_sph.sph_version != 1) {
1923                 goto bad;
1924         }
1925
1926         switch (mdu->mdu_tlv.tlv_type) {
1927         case MARKER_TYPE_INFO:
1928                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1929                     marker_info_tlv_template, TRUE)) {
1930                         goto bad;
1931                 }
1932                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1933                 memcpy(&mdu->mdu_eh.ether_dhost,
1934                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1935                 memcpy(&mdu->mdu_eh.ether_shost,
1936                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1937                 error = lagg_enqueue(lp->lp_ifp, m);
1938                 break;
1939
1940         case MARKER_TYPE_RESPONSE:
1941                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1942                     marker_response_tlv_template, TRUE)) {
1943                         goto bad;
1944                 }
1945                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1946                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1947                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1948
1949                 /* Verify that it is the last marker we sent out */
1950                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1951                     sizeof(struct lacp_markerinfo)))
1952                         goto bad;
1953
1954                 LACP_LOCK(lsc);
1955                 lp->lp_flags &= ~LACP_PORT_MARK;
1956
1957                 if (lsc->lsc_suppress_distributing) {
1958                         /* Check if any ports are waiting for a response */
1959                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1960                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1961                                         pending = 1;
1962                                         break;
1963                                 }
1964                         }
1965
1966                         if (pending == 0) {
1967                                 /* All interface queues are clear */
1968                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1969                                 lsc->lsc_suppress_distributing = FALSE;
1970                         }
1971                 }
1972                 LACP_UNLOCK(lsc);
1973                 m_freem(m);
1974                 break;
1975
1976         default:
1977                 goto bad;
1978         }
1979
1980         return (error);
1981
1982 bad:
1983         LACP_DPRINTF((lp, "bad marker frame\n"));
1984         m_freem(m);
1985         return (EINVAL);
1986 }
1987
1988 static int
1989 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1990     const struct tlv_template *tmpl, boolean_t check_type)
1991 {
1992         while (/* CONSTCOND */ 1) {
1993                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1994                         return (EINVAL);
1995                 }
1996                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1997                     tlv->tlv_length != tmpl->tmpl_length) {
1998                         return (EINVAL);
1999                 }
2000                 if (tmpl->tmpl_type == 0) {
2001                         break;
2002                 }
2003                 tlv = (const struct tlvhdr *)
2004                     ((const char *)tlv + tlv->tlv_length);
2005                 tmpl++;
2006         }
2007
2008         return (0);
2009 }
2010
2011 /* Debugging */
2012 const char *
2013 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
2014 {
2015         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
2016             (int)mac[0],
2017             (int)mac[1],
2018             (int)mac[2],
2019             (int)mac[3],
2020             (int)mac[4],
2021             (int)mac[5]);
2022
2023         return (buf);
2024 }
2025
2026 const char *
2027 lacp_format_systemid(const struct lacp_systemid *sysid,
2028     char *buf, size_t buflen)
2029 {
2030         char macbuf[LACP_MACSTR_MAX+1];
2031
2032         snprintf(buf, buflen, "%04X,%s",
2033             ntohs(sysid->lsi_prio),
2034             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2035
2036         return (buf);
2037 }
2038
2039 const char *
2040 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2041 {
2042         snprintf(buf, buflen, "%04X,%04X",
2043             ntohs(portid->lpi_prio),
2044             ntohs(portid->lpi_portno));
2045
2046         return (buf);
2047 }
2048
2049 const char *
2050 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2051 {
2052         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2053         char portid[LACP_PORTIDSTR_MAX+1];
2054
2055         snprintf(buf, buflen, "(%s,%04X,%s)",
2056             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2057             ntohs(peer->lip_key),
2058             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2059
2060         return (buf);
2061 }
2062
2063 const char *
2064 lacp_format_lagid(const struct lacp_peerinfo *a,
2065     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2066 {
2067         char astr[LACP_PARTNERSTR_MAX+1];
2068         char bstr[LACP_PARTNERSTR_MAX+1];
2069
2070 #if 0
2071         /*
2072          * there's a convention to display small numbered peer
2073          * in the left.
2074          */
2075
2076         if (lacp_compare_peerinfo(a, b) > 0) {
2077                 const struct lacp_peerinfo *t;
2078
2079                 t = a;
2080                 a = b;
2081                 b = t;
2082         }
2083 #endif
2084
2085         snprintf(buf, buflen, "[%s,%s]",
2086             lacp_format_partner(a, astr, sizeof(astr)),
2087             lacp_format_partner(b, bstr, sizeof(bstr)));
2088
2089         return (buf);
2090 }
2091
2092 const char *
2093 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2094     char *buf, size_t buflen)
2095 {
2096         if (la == NULL) {
2097                 return ("(none)");
2098         }
2099
2100         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2101 }
2102
2103 const char *
2104 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2105 {
2106         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2107         return (buf);
2108 }
2109
2110 static void
2111 lacp_dump_lacpdu(const struct lacpdu *du)
2112 {
2113         char buf[LACP_PARTNERSTR_MAX+1];
2114         char buf2[LACP_STATESTR_MAX+1];
2115
2116         printf("actor=%s\n",
2117             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2118         printf("actor.state=%s\n",
2119             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2120         printf("partner=%s\n",
2121             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2122         printf("partner.state=%s\n",
2123             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2124
2125         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2126 }
2127
2128 static void
2129 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2130 {
2131         va_list va;
2132
2133         if (lp) {
2134                 printf("%s: ", lp->lp_ifp->if_xname);
2135         }
2136
2137         va_start(va, fmt);
2138         vprintf(fmt, va);
2139         va_end(va);
2140 }