sys/net/ieee8023ad_lacp.c

   1 /*      $NetBSD: ieee8023ad_lacp.c,v 1.3 2005/12/11 12:24:54 christos Exp $     */
   2
   3 /*-
   4  * Copyright (c)2005 YAMAMOTO Takashi,
   5  * Copyright (c)2008 Andrew Thompson <thompsa@FreeBSD.org>
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 #include <sys/cdefs.h>
  31 __FBSDID("$FreeBSD$");
  32
  33 #include "opt_ratelimit.h"
  34
  35 #include <sys/param.h>
  36 #include <sys/callout.h>
  37 #include <sys/eventhandler.h>
  38 #include <sys/mbuf.h>
  39 #include <sys/systm.h>
  40 #include <sys/malloc.h>
  41 #include <sys/kernel.h> /* hz */
  42 #include <sys/socket.h> /* for net/if.h */
  43 #include <sys/sockio.h>
  44 #include <sys/sysctl.h>
  45 #include <machine/stdarg.h>
  46 #include <sys/lock.h>
  47 #include <sys/rwlock.h>
  48 #include <sys/taskqueue.h>
  49
  50 #include <net/if.h>
  51 #include <net/if_var.h>
  52 #include <net/if_dl.h>
  53 #include <net/ethernet.h>
  54 #include <net/if_media.h>
  55 #include <net/if_types.h>
  56
  57 #include <net/if_lagg.h>
  58 #include <net/ieee8023ad_lacp.h>
  59
  60 /*
  61  * actor system priority and port priority.
  62  * XXX should be configurable.
  63  */
  64
  65 #define LACP_SYSTEM_PRIO        0x8000
  66 #define LACP_PORT_PRIO          0x8000
  67
  68 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  69     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  70
  71 static const struct tlv_template lacp_info_tlv_template[] = {
  72         { LACP_TYPE_ACTORINFO,
  73             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  74         { LACP_TYPE_PARTNERINFO,
  75             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  76         { LACP_TYPE_COLLECTORINFO,
  77             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  78         { 0, 0 },
  79 };
  80
  81 static const struct tlv_template marker_info_tlv_template[] = {
  82         { MARKER_TYPE_INFO,
  83             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  84         { 0, 0 },
  85 };
  86
  87 static const struct tlv_template marker_response_tlv_template[] = {
  88         { MARKER_TYPE_RESPONSE,
  89             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  90         { 0, 0 },
  91 };
  92
  93 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  94
  95 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
  96 static void     lacp_fill_markerinfo(struct lacp_port *,
  97                     struct lacp_markerinfo *);
  98
  99 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
 100 static void     lacp_suppress_distributing(struct lacp_softc *,
 101                     struct lacp_aggregator *);
 102 static void     lacp_transit_expire(void *);
 103 static void     lacp_update_portmap(struct lacp_softc *);
 104 static void     lacp_select_active_aggregator(struct lacp_softc *);
 105 static uint16_t lacp_compose_key(struct lacp_port *);
 106 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 107                     const struct tlv_template *, boolean_t);
 108 static void     lacp_tick(void *);
 109
 110 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 111                     const struct lacp_port *);
 112 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 113                     const struct lacp_peerinfo *);
 114 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 115                     const struct lacp_port *);
 116 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 117                     const struct lacp_peerinfo *);
 118
 119 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 120                     struct lacp_port *);
 121 static void     lacp_aggregator_addref(struct lacp_softc *,
 122                     struct lacp_aggregator *);
 123 static void     lacp_aggregator_delref(struct lacp_softc *,
 124                     struct lacp_aggregator *);
 125
 126 /* receive machine */
 127
 128 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 129 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 130 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 131 static void     lacp_sm_rx_timer(struct lacp_port *);
 132 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 133 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 134                     const struct lacpdu *);
 135 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 136                     const struct lacpdu *);
 137 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 138                     const struct lacpdu *);
 139 static void     lacp_sm_rx_record_default(struct lacp_port *);
 140 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 141 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 142                     const struct lacp_peerinfo *);
 143
 144 /* mux machine */
 145
 146 static void     lacp_sm_mux(struct lacp_port *);
 147 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 148 static void     lacp_sm_mux_timer(struct lacp_port *);
 149
 150 /* periodic transmit machine */
 151
 152 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 153 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 154 static void     lacp_sm_ptx_timer(struct lacp_port *);
 155
 156 /* transmit machine */
 157
 158 static void     lacp_sm_tx(struct lacp_port *);
 159 static void     lacp_sm_assert_ntt(struct lacp_port *);
 160
 161 static void     lacp_run_timers(struct lacp_port *);
 162 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 163                     const struct lacp_peerinfo *);
 164 static int      lacp_compare_systemid(const struct lacp_systemid *,
 165                     const struct lacp_systemid *);
 166 static void     lacp_port_enable(struct lacp_port *);
 167 static void     lacp_port_disable(struct lacp_port *);
 168 static void     lacp_select(struct lacp_port *);
 169 static void     lacp_unselect(struct lacp_port *);
 170 static void     lacp_disable_collecting(struct lacp_port *);
 171 static void     lacp_enable_collecting(struct lacp_port *);
 172 static void     lacp_disable_distributing(struct lacp_port *);
 173 static void     lacp_enable_distributing(struct lacp_port *);
 174 static int      lacp_xmit_lacpdu(struct lacp_port *);
 175 static int      lacp_xmit_marker(struct lacp_port *);
 176
 177 /* Debugging */
 178
 179 static void     lacp_dump_lacpdu(const struct lacpdu *);
 180 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 181                     size_t);
 182 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 183                     const struct lacp_peerinfo *, char *, size_t);
 184 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 185                     char *, size_t);
 186 static const char *lacp_format_state(uint8_t, char *, size_t);
 187 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 188 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 189                     size_t);
 190 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 191                     size_t);
 192 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 193                     __attribute__((__format__(__printf__, 2, 3)));
 194
 195 static VNET_DEFINE(int, lacp_debug);
 196 #define V_lacp_debug    VNET(lacp_debug)
 197 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
 198 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 199     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 200
 201 static VNET_DEFINE(int, lacp_default_strict_mode) = 1;
 202 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode, CTLFLAG_RWTUN,
 203     &VNET_NAME(lacp_default_strict_mode), 0,
 204     "LACP strict protocol compliance default");
 205
 206 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 207 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 208 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 209
 210 /*
 211  * partner administration variables.
 212  * XXX should be configurable.
 213  */
 214
 215 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 216         .lip_systemid = { .lsi_prio = 0xffff },
 217         .lip_portid = { .lpi_prio = 0xffff },
 218         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 219             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 220 };
 221
 222 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 223         .lip_systemid = { .lsi_prio = 0xffff },
 224         .lip_portid = { .lpi_prio = 0xffff },
 225         .lip_state = 0,
 226 };
 227
 228 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 229         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 230         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 231         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 232 };
 233
 234 struct mbuf *
 235 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 236 {
 237         struct lacp_port *lp = LACP_PORT(lgp);
 238         uint8_t subtype;
 239
 240         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 241                 m_freem(m);
 242                 return (NULL);
 243         }
 244
 245         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 246         switch (subtype) {
 247                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 248                         lacp_pdu_input(lp, m);
 249                         return (NULL);
 250
 251                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 252                         lacp_marker_input(lp, m);
 253                         return (NULL);
 254         }
 255
 256         /* Not a subtype we are interested in */
 257         return (m);
 258 }
 259
 260 /*
 261  * lacp_pdu_input: process lacpdu
 262  */
 263 static int
 264 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 265 {
 266         struct lacp_softc *lsc = lp->lp_lsc;
 267         struct lacpdu *du;
 268         int error = 0;
 269
 270         if (m->m_pkthdr.len != sizeof(*du)) {
 271                 goto bad;
 272         }
 273
 274         if ((m->m_flags & M_MCAST) == 0) {
 275                 goto bad;
 276         }
 277
 278         if (m->m_len < sizeof(*du)) {
 279                 m = m_pullup(m, sizeof(*du));
 280                 if (m == NULL) {
 281                         return (ENOMEM);
 282                 }
 283         }
 284
 285         du = mtod(m, struct lacpdu *);
 286
 287         if (memcmp(&du->ldu_eh.ether_dhost,
 288             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 289                 goto bad;
 290         }
 291
 292         /*
 293          * ignore the version for compatibility with
 294          * the future protocol revisions.
 295          */
 296 #if 0
 297         if (du->ldu_sph.sph_version != 1) {
 298                 goto bad;
 299         }
 300 #endif
 301
 302         /*
 303          * ignore tlv types for compatibility with
 304          * the future protocol revisions.
 305          */
 306         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 307             lacp_info_tlv_template, FALSE)) {
 308                 goto bad;
 309         }
 310
 311         if (V_lacp_debug > 0) {
 312                 lacp_dprintf(lp, "lacpdu receive\n");
 313                 lacp_dump_lacpdu(du);
 314         }
 315
 316         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 317                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 318                 goto bad;
 319         }
 320
 321         LACP_LOCK(lsc);
 322         lacp_sm_rx(lp, du);
 323         LACP_UNLOCK(lsc);
 324
 325         m_freem(m);
 326         return (error);
 327
 328 bad:
 329         m_freem(m);
 330         return (EINVAL);
 331 }
 332
 333 static void
 334 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 335 {
 336         struct lagg_port *lgp = lp->lp_lagg;
 337         struct lagg_softc *sc = lgp->lp_softc;
 338
 339         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 340         memcpy(&info->lip_systemid.lsi_mac,
 341             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 342         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 343         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 344         info->lip_state = lp->lp_state;
 345 }
 346
 347 static void
 348 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 349 {
 350         struct ifnet *ifp = lp->lp_ifp;
 351
 352         /* Fill in the port index and system id (encoded as the MAC) */
 353         info->mi_rq_port = htons(ifp->if_index);
 354         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 355         info->mi_rq_xid = htonl(0);
 356 }
 357
 358 static int
 359 lacp_xmit_lacpdu(struct lacp_port *lp)
 360 {
 361         struct lagg_port *lgp = lp->lp_lagg;
 362         struct mbuf *m;
 363         struct lacpdu *du;
 364         int error;
 365
 366         LACP_LOCK_ASSERT(lp->lp_lsc);
 367
 368         m = m_gethdr(M_NOWAIT, MT_DATA);
 369         if (m == NULL) {
 370                 return (ENOMEM);
 371         }
 372         m->m_len = m->m_pkthdr.len = sizeof(*du);
 373
 374         du = mtod(m, struct lacpdu *);
 375         memset(du, 0, sizeof(*du));
 376
 377         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 378             ETHER_ADDR_LEN);
 379         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 380         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 381
 382         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 383         du->ldu_sph.sph_version = 1;
 384
 385         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 386         du->ldu_actor = lp->lp_actor;
 387
 388         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 389             sizeof(du->ldu_partner));
 390         du->ldu_partner = lp->lp_partner;
 391
 392         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 393             sizeof(du->ldu_collector));
 394         du->ldu_collector.lci_maxdelay = 0;
 395
 396         if (V_lacp_debug > 0) {
 397                 lacp_dprintf(lp, "lacpdu transmit\n");
 398                 lacp_dump_lacpdu(du);
 399         }
 400
 401         m->m_flags |= M_MCAST;
 402
 403         /*
 404          * XXX should use higher priority queue.
 405          * otherwise network congestion can break aggregation.
 406          */
 407
 408         error = lagg_enqueue(lp->lp_ifp, m);
 409         return (error);
 410 }
 411
 412 static int
 413 lacp_xmit_marker(struct lacp_port *lp)
 414 {
 415         struct lagg_port *lgp = lp->lp_lagg;
 416         struct mbuf *m;
 417         struct markerdu *mdu;
 418         int error;
 419
 420         LACP_LOCK_ASSERT(lp->lp_lsc);
 421
 422         m = m_gethdr(M_NOWAIT, MT_DATA);
 423         if (m == NULL) {
 424                 return (ENOMEM);
 425         }
 426         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 427
 428         mdu = mtod(m, struct markerdu *);
 429         memset(mdu, 0, sizeof(*mdu));
 430
 431         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 432             ETHER_ADDR_LEN);
 433         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 434         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 435
 436         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 437         mdu->mdu_sph.sph_version = 1;
 438
 439         /* Bump the transaction id and copy over the marker info */
 440         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 441         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 442         mdu->mdu_info = lp->lp_marker;
 443
 444         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 445             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 446             ntohl(mdu->mdu_info.mi_rq_xid)));
 447
 448         m->m_flags |= M_MCAST;
 449         error = lagg_enqueue(lp->lp_ifp, m);
 450         return (error);
 451 }
 452
 453 void
 454 lacp_linkstate(struct lagg_port *lgp)
 455 {
 456         struct lacp_port *lp = LACP_PORT(lgp);
 457         struct lacp_softc *lsc = lp->lp_lsc;
 458         struct ifnet *ifp = lgp->lp_ifp;
 459         struct ifmediareq ifmr;
 460         int error = 0;
 461         u_int media;
 462         uint8_t old_state;
 463         uint16_t old_key;
 464
 465         bzero((char *)&ifmr, sizeof(ifmr));
 466         error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 467         if (error != 0)
 468                 return;
 469
 470         LACP_LOCK(lsc);
 471         media = ifmr.ifm_active;
 472         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 473             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 474             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 475         old_state = lp->lp_state;
 476         old_key = lp->lp_key;
 477
 478         lp->lp_media = media;
 479         /*
 480          * If the port is not an active full duplex Ethernet link then it can
 481          * not be aggregated.
 482          */
 483         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 484             ifp->if_link_state != LINK_STATE_UP) {
 485                 lacp_port_disable(lp);
 486         } else {
 487                 lacp_port_enable(lp);
 488         }
 489         lp->lp_key = lacp_compose_key(lp);
 490
 491         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 492                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 493                 lp->lp_selected = LACP_UNSELECTED;
 494         }
 495         LACP_UNLOCK(lsc);
 496 }
 497
 498 static void
 499 lacp_tick(void *arg)
 500 {
 501         struct lacp_softc *lsc = arg;
 502         struct lacp_port *lp;
 503
 504         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 505                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 506                         continue;
 507
 508                 CURVNET_SET(lp->lp_ifp->if_vnet);
 509                 lacp_run_timers(lp);
 510
 511                 lacp_select(lp);
 512                 lacp_sm_mux(lp);
 513                 lacp_sm_tx(lp);
 514                 lacp_sm_ptx_tx_schedule(lp);
 515                 CURVNET_RESTORE();
 516         }
 517         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 518 }
 519
 520 int
 521 lacp_port_create(struct lagg_port *lgp)
 522 {
 523         struct lagg_softc *sc = lgp->lp_softc;
 524         struct lacp_softc *lsc = LACP_SOFTC(sc);
 525         struct lacp_port *lp;
 526         struct ifnet *ifp = lgp->lp_ifp;
 527         struct sockaddr_dl sdl;
 528         struct ifmultiaddr *rifma = NULL;
 529         int error;
 530
 531         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 532         sdl.sdl_alen = ETHER_ADDR_LEN;
 533
 534         bcopy(&ethermulticastaddr_slowprotocols,
 535             LLADDR(&sdl), ETHER_ADDR_LEN);
 536         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 537         if (error) {
 538                 printf("%s: ADDMULTI failed on %s\n", __func__,
 539                     lgp->lp_ifp->if_xname);
 540                 return (error);
 541         }
 542
 543         lp = malloc(sizeof(struct lacp_port),
 544             M_DEVBUF, M_NOWAIT|M_ZERO);
 545         if (lp == NULL)
 546                 return (ENOMEM);
 547
 548         LACP_LOCK(lsc);
 549         lgp->lp_psc = lp;
 550         lp->lp_ifp = ifp;
 551         lp->lp_lagg = lgp;
 552         lp->lp_lsc = lsc;
 553         lp->lp_ifma = rifma;
 554
 555         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 556
 557         lacp_fill_actorinfo(lp, &lp->lp_actor);
 558         lacp_fill_markerinfo(lp, &lp->lp_marker);
 559         lp->lp_state = LACP_STATE_ACTIVITY;
 560         lp->lp_aggregator = NULL;
 561         lacp_sm_rx_set_expired(lp);
 562         LACP_UNLOCK(lsc);
 563         lacp_linkstate(lgp);
 564
 565         return (0);
 566 }
 567
 568 void
 569 lacp_port_destroy(struct lagg_port *lgp)
 570 {
 571         struct lacp_port *lp = LACP_PORT(lgp);
 572         struct lacp_softc *lsc = lp->lp_lsc;
 573         int i;
 574
 575         LACP_LOCK(lsc);
 576         for (i = 0; i < LACP_NTIMER; i++) {
 577                 LACP_TIMER_DISARM(lp, i);
 578         }
 579
 580         lacp_disable_collecting(lp);
 581         lacp_disable_distributing(lp);
 582         lacp_unselect(lp);
 583
 584         LIST_REMOVE(lp, lp_next);
 585         LACP_UNLOCK(lsc);
 586
 587         /* The address may have already been removed by if_purgemaddrs() */
 588         if (!lgp->lp_detaching)
 589                 if_delmulti_ifma(lp->lp_ifma);
 590
 591         free(lp, M_DEVBUF);
 592 }
 593
 594 void
 595 lacp_req(struct lagg_softc *sc, void *data)
 596 {
 597         struct lacp_opreq *req = (struct lacp_opreq *)data;
 598         struct lacp_softc *lsc = LACP_SOFTC(sc);
 599         struct lacp_aggregator *la;
 600
 601         bzero(req, sizeof(struct lacp_opreq));
 602
 603         /*
 604          * If the LACP softc is NULL, return with the opreq structure full of
 605          * zeros.  It is normal for the softc to be NULL while the lagg is
 606          * being destroyed.
 607          */
 608         if (NULL == lsc)
 609                 return;
 610
 611         la = lsc->lsc_active_aggregator;
 612         LACP_LOCK(lsc);
 613         if (la != NULL) {
 614                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 615                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 616                     ETHER_ADDR_LEN);
 617                 req->actor_key = ntohs(la->la_actor.lip_key);
 618                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 619                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 620                 req->actor_state = la->la_actor.lip_state;
 621
 622                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 623                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 624                     ETHER_ADDR_LEN);
 625                 req->partner_key = ntohs(la->la_partner.lip_key);
 626                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 627                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 628                 req->partner_state = la->la_partner.lip_state;
 629         }
 630         LACP_UNLOCK(lsc);
 631 }
 632
 633 void
 634 lacp_portreq(struct lagg_port *lgp, void *data)
 635 {
 636         struct lacp_opreq *req = (struct lacp_opreq *)data;
 637         struct lacp_port *lp = LACP_PORT(lgp);
 638         struct lacp_softc *lsc = lp->lp_lsc;
 639
 640         LACP_LOCK(lsc);
 641         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 642         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 643             ETHER_ADDR_LEN);
 644         req->actor_key = ntohs(lp->lp_actor.lip_key);
 645         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 646         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 647         req->actor_state = lp->lp_actor.lip_state;
 648
 649         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 650         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 651             ETHER_ADDR_LEN);
 652         req->partner_key = ntohs(lp->lp_partner.lip_key);
 653         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 654         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 655         req->partner_state = lp->lp_partner.lip_state;
 656         LACP_UNLOCK(lsc);
 657 }
 658
 659 static void
 660 lacp_disable_collecting(struct lacp_port *lp)
 661 {
 662         LACP_DPRINTF((lp, "collecting disabled\n"));
 663         lp->lp_state &= ~LACP_STATE_COLLECTING;
 664 }
 665
 666 static void
 667 lacp_enable_collecting(struct lacp_port *lp)
 668 {
 669         LACP_DPRINTF((lp, "collecting enabled\n"));
 670         lp->lp_state |= LACP_STATE_COLLECTING;
 671 }
 672
 673 static void
 674 lacp_disable_distributing(struct lacp_port *lp)
 675 {
 676         struct lacp_aggregator *la = lp->lp_aggregator;
 677         struct lacp_softc *lsc = lp->lp_lsc;
 678         struct lagg_softc *sc = lsc->lsc_softc;
 679         char buf[LACP_LAGIDSTR_MAX+1];
 680
 681         LACP_LOCK_ASSERT(lsc);
 682
 683         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 684                 return;
 685         }
 686
 687         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 688         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 689         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 690
 691         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 692             "nports %d -> %d\n",
 693             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 694             la->la_nports, la->la_nports - 1));
 695
 696         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 697         la->la_nports--;
 698         sc->sc_active = la->la_nports;
 699
 700         if (lsc->lsc_active_aggregator == la) {
 701                 lacp_suppress_distributing(lsc, la);
 702                 lacp_select_active_aggregator(lsc);
 703                 /* regenerate the port map, the active aggregator has changed */
 704                 lacp_update_portmap(lsc);
 705         }
 706
 707         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 708 }
 709
 710 static void
 711 lacp_enable_distributing(struct lacp_port *lp)
 712 {
 713         struct lacp_aggregator *la = lp->lp_aggregator;
 714         struct lacp_softc *lsc = lp->lp_lsc;
 715         struct lagg_softc *sc = lsc->lsc_softc;
 716         char buf[LACP_LAGIDSTR_MAX+1];
 717
 718         LACP_LOCK_ASSERT(lsc);
 719
 720         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 721                 return;
 722         }
 723
 724         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 725             "nports %d -> %d\n",
 726             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 727             la->la_nports, la->la_nports + 1));
 728
 729         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 730         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 731         la->la_nports++;
 732         sc->sc_active = la->la_nports;
 733
 734         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 735
 736         if (lsc->lsc_active_aggregator == la) {
 737                 lacp_suppress_distributing(lsc, la);
 738                 lacp_update_portmap(lsc);
 739         } else
 740                 /* try to become the active aggregator */
 741                 lacp_select_active_aggregator(lsc);
 742 }
 743
 744 static void
 745 lacp_transit_expire(void *vp)
 746 {
 747         struct lacp_softc *lsc = vp;
 748
 749         LACP_LOCK_ASSERT(lsc);
 750
 751         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 752         LACP_TRACE(NULL);
 753         CURVNET_RESTORE();
 754
 755         lsc->lsc_suppress_distributing = FALSE;
 756 }
 757
 758 void
 759 lacp_attach(struct lagg_softc *sc)
 760 {
 761         struct lacp_softc *lsc;
 762
 763         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 764
 765         sc->sc_psc = lsc;
 766         lsc->lsc_softc = sc;
 767
 768         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 769         lsc->lsc_active_aggregator = NULL;
 770         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 771         LACP_LOCK_INIT(lsc);
 772         TAILQ_INIT(&lsc->lsc_aggregators);
 773         LIST_INIT(&lsc->lsc_ports);
 774
 775         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 776         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 777
 778         /* if the lagg is already up then do the same */
 779         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 780                 lacp_init(sc);
 781 }
 782
 783 void
 784 lacp_detach(void *psc)
 785 {
 786         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 787
 788         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 789             ("aggregators still active"));
 790         KASSERT(lsc->lsc_active_aggregator == NULL,
 791             ("aggregator still attached"));
 792
 793         callout_drain(&lsc->lsc_transit_callout);
 794         callout_drain(&lsc->lsc_callout);
 795
 796         LACP_LOCK_DESTROY(lsc);
 797         free(lsc, M_DEVBUF);
 798 }
 799
 800 void
 801 lacp_init(struct lagg_softc *sc)
 802 {
 803         struct lacp_softc *lsc = LACP_SOFTC(sc);
 804
 805         LACP_LOCK(lsc);
 806         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 807         LACP_UNLOCK(lsc);
 808 }
 809
 810 void
 811 lacp_stop(struct lagg_softc *sc)
 812 {
 813         struct lacp_softc *lsc = LACP_SOFTC(sc);
 814
 815         LACP_LOCK(lsc);
 816         callout_stop(&lsc->lsc_transit_callout);
 817         callout_stop(&lsc->lsc_callout);
 818         LACP_UNLOCK(lsc);
 819 }
 820
 821 struct lagg_port *
 822 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 823 {
 824         struct lacp_softc *lsc = LACP_SOFTC(sc);
 825         struct lacp_portmap *pm;
 826         struct lacp_port *lp;
 827         uint32_t hash;
 828
 829         if (__predict_false(lsc->lsc_suppress_distributing)) {
 830                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 831                 return (NULL);
 832         }
 833
 834         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 835         if (pm->pm_count == 0) {
 836                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 837                 return (NULL);
 838         }
 839
 840         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 841             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 842                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 843         else
 844                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 845         hash %= pm->pm_count;
 846         lp = pm->pm_map[hash];
 847
 848         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 849             ("aggregated port is not distributing"));
 850
 851         return (lp->lp_lagg);
 852 }
 853
 854 #ifdef RATELIMIT
 855 struct lagg_port *
 856 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t flowid)
 857 {
 858         struct lacp_softc *lsc = LACP_SOFTC(sc);
 859         struct lacp_portmap *pm;
 860         struct lacp_port *lp;
 861         uint32_t hash;
 862
 863         if (__predict_false(lsc->lsc_suppress_distributing)) {
 864                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 865                 return (NULL);
 866         }
 867
 868         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 869         if (pm->pm_count == 0) {
 870                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 871                 return (NULL);
 872         }
 873
 874         hash = flowid >> sc->flowid_shift;
 875         hash %= pm->pm_count;
 876         lp = pm->pm_map[hash];
 877
 878         return (lp->lp_lagg);
 879 }
 880 #endif
 881
 882 /*
 883  * lacp_suppress_distributing: drop transmit packets for a while
 884  * to preserve packet ordering.
 885  */
 886
 887 static void
 888 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 889 {
 890         struct lacp_port *lp;
 891
 892         if (lsc->lsc_active_aggregator != la) {
 893                 return;
 894         }
 895
 896         LACP_TRACE(NULL);
 897
 898         lsc->lsc_suppress_distributing = TRUE;
 899
 900         /* send a marker frame down each port to verify the queues are empty */
 901         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 902                 lp->lp_flags |= LACP_PORT_MARK;
 903                 lacp_xmit_marker(lp);
 904         }
 905
 906         /* set a timeout for the marker frames */
 907         callout_reset(&lsc->lsc_transit_callout,
 908             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 909 }
 910
 911 static int
 912 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 913     const struct lacp_peerinfo *b)
 914 {
 915         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 916 }
 917
 918 static int
 919 lacp_compare_systemid(const struct lacp_systemid *a,
 920     const struct lacp_systemid *b)
 921 {
 922         return (memcmp(a, b, sizeof(*a)));
 923 }
 924
 925 #if 0   /* unused */
 926 static int
 927 lacp_compare_portid(const struct lacp_portid *a,
 928     const struct lacp_portid *b)
 929 {
 930         return (memcmp(a, b, sizeof(*a)));
 931 }
 932 #endif
 933
 934 static uint64_t
 935 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 936 {
 937         struct lacp_port *lp;
 938         uint64_t speed;
 939
 940         lp = TAILQ_FIRST(&la->la_ports);
 941         if (lp == NULL) {
 942                 return (0);
 943         }
 944
 945         speed = ifmedia_baudrate(lp->lp_media);
 946         speed *= la->la_nports;
 947         if (speed == 0) {
 948                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 949                     lp->lp_media, la->la_nports));
 950         }
 951
 952         return (speed);
 953 }
 954
 955 /*
 956  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 957  * packets from lagg(4) interface.
 958  */
 959
 960 static void
 961 lacp_select_active_aggregator(struct lacp_softc *lsc)
 962 {
 963         struct lacp_aggregator *la;
 964         struct lacp_aggregator *best_la = NULL;
 965         uint64_t best_speed = 0;
 966         char buf[LACP_LAGIDSTR_MAX+1];
 967
 968         LACP_TRACE(NULL);
 969
 970         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 971                 uint64_t speed;
 972
 973                 if (la->la_nports == 0) {
 974                         continue;
 975                 }
 976
 977                 speed = lacp_aggregator_bandwidth(la);
 978                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 979                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 980                     speed, la->la_nports));
 981
 982                 /*
 983                  * This aggregator is chosen if the partner has a better
 984                  * system priority or, the total aggregated speed is higher
 985                  * or, it is already the chosen aggregator
 986                  */
 987                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 988                     LACP_SYS_PRI(best_la->la_partner)) ||
 989                     speed > best_speed ||
 990                     (speed == best_speed &&
 991                     la == lsc->lsc_active_aggregator)) {
 992                         best_la = la;
 993                         best_speed = speed;
 994                 }
 995         }
 996
 997         KASSERT(best_la == NULL || best_la->la_nports > 0,
 998             ("invalid aggregator refcnt"));
 999         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1000             ("invalid aggregator list"));
1001
1002         if (lsc->lsc_active_aggregator != best_la) {
1003                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1004                 LACP_DPRINTF((NULL, "old %s\n",
1005                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1006                     buf, sizeof(buf))));
1007         } else {
1008                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1009         }
1010         LACP_DPRINTF((NULL, "new %s\n",
1011             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1012
1013         if (lsc->lsc_active_aggregator != best_la) {
1014                 lsc->lsc_active_aggregator = best_la;
1015                 lacp_update_portmap(lsc);
1016                 if (best_la) {
1017                         lacp_suppress_distributing(lsc, best_la);
1018                 }
1019         }
1020 }
1021
1022 /*
1023  * Updated the inactive portmap array with the new list of ports and
1024  * make it live.
1025  */
1026 static void
1027 lacp_update_portmap(struct lacp_softc *lsc)
1028 {
1029         struct lagg_softc *sc = lsc->lsc_softc;
1030         struct lacp_aggregator *la;
1031         struct lacp_portmap *p;
1032         struct lacp_port *lp;
1033         uint64_t speed;
1034         u_int newmap;
1035         int i;
1036
1037         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1038         p = &lsc->lsc_pmap[newmap];
1039         la = lsc->lsc_active_aggregator;
1040         speed = 0;
1041         bzero(p, sizeof(struct lacp_portmap));
1042
1043         if (la != NULL && la->la_nports > 0) {
1044                 p->pm_count = la->la_nports;
1045                 i = 0;
1046                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1047                         p->pm_map[i++] = lp;
1048                 KASSERT(i == p->pm_count, ("Invalid port count"));
1049                 speed = lacp_aggregator_bandwidth(la);
1050         }
1051         sc->sc_ifp->if_baudrate = speed;
1052
1053         /* switch the active portmap over */
1054         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1055         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1056                     lsc->lsc_activemap,
1057                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1058 }
1059
1060 static uint16_t
1061 lacp_compose_key(struct lacp_port *lp)
1062 {
1063         struct lagg_port *lgp = lp->lp_lagg;
1064         struct lagg_softc *sc = lgp->lp_softc;
1065         u_int media = lp->lp_media;
1066         uint16_t key;
1067
1068         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1069
1070                 /*
1071                  * non-aggregatable links should have unique keys.
1072                  *
1073                  * XXX this isn't really unique as if_index is 16 bit.
1074                  */
1075
1076                 /* bit 0..14:   (some bits of) if_index of this port */
1077                 key = lp->lp_ifp->if_index;
1078                 /* bit 15:      1 */
1079                 key |= 0x8000;
1080         } else {
1081                 u_int subtype = IFM_SUBTYPE(media);
1082
1083                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1084                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1085
1086                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1087                 switch (subtype) {
1088                 case IFM_10_T:
1089                 case IFM_10_2:
1090                 case IFM_10_5:
1091                 case IFM_10_STP:
1092                 case IFM_10_FL:
1093                         key = IFM_10_T;
1094                         break;
1095                 case IFM_100_TX:
1096                 case IFM_100_FX:
1097                 case IFM_100_T4:
1098                 case IFM_100_VG:
1099                 case IFM_100_T2:
1100                 case IFM_100_T:
1101                         key = IFM_100_TX;
1102                         break;
1103                 case IFM_1000_SX:
1104                 case IFM_1000_LX:
1105                 case IFM_1000_CX:
1106                 case IFM_1000_T:
1107                 case IFM_1000_KX:
1108                 case IFM_1000_SGMII:
1109                 case IFM_1000_CX_SGMII:
1110                         key = IFM_1000_SX;
1111                         break;
1112                 case IFM_10G_LR:
1113                 case IFM_10G_SR:
1114                 case IFM_10G_CX4:
1115                 case IFM_10G_TWINAX:
1116                 case IFM_10G_TWINAX_LONG:
1117                 case IFM_10G_LRM:
1118                 case IFM_10G_T:
1119                 case IFM_10G_KX4:
1120                 case IFM_10G_KR:
1121                 case IFM_10G_CR1:
1122                 case IFM_10G_ER:
1123                 case IFM_10G_SFI:
1124                         key = IFM_10G_LR;
1125                         break;
1126                 case IFM_20G_KR2:
1127                         key = IFM_20G_KR2;
1128                         break;
1129                 case IFM_2500_KX:
1130                 case IFM_2500_T:
1131                         key = IFM_2500_KX;
1132                         break;
1133                 case IFM_5000_T:
1134                         key = IFM_5000_T;
1135                         break;
1136                 case IFM_50G_PCIE:
1137                 case IFM_50G_CR2:
1138                 case IFM_50G_KR2:
1139                         key = IFM_50G_PCIE;
1140                         break;
1141                 case IFM_56G_R4:
1142                         key = IFM_56G_R4;
1143                         break;
1144                 case IFM_25G_PCIE:
1145                 case IFM_25G_CR:
1146                 case IFM_25G_KR:
1147                 case IFM_25G_SR:
1148                         key = IFM_25G_PCIE;
1149                         break;
1150                 case IFM_40G_CR4:
1151                 case IFM_40G_SR4:
1152                 case IFM_40G_LR4:
1153                 case IFM_40G_XLPPI:
1154                 case IFM_40G_KR4:
1155                         key = IFM_40G_CR4;
1156                         break;
1157                 case IFM_100G_CR4:
1158                 case IFM_100G_SR4:
1159                 case IFM_100G_KR4:
1160                 case IFM_100G_LR4:
1161                         key = IFM_100G_CR4;
1162                         break;
1163                 default:
1164                         key = subtype;
1165                         break;
1166                 }
1167                 /* bit 5..14:   (some bits of) if_index of lagg device */
1168                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1169                 /* bit 15:      0 */
1170         }
1171         return (htons(key));
1172 }
1173
1174 static void
1175 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1176 {
1177         char buf[LACP_LAGIDSTR_MAX+1];
1178
1179         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1180             __func__,
1181             lacp_format_lagid(&la->la_actor, &la->la_partner,
1182             buf, sizeof(buf)),
1183             la->la_refcnt, la->la_refcnt + 1));
1184
1185         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1186         la->la_refcnt++;
1187         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1188 }
1189
1190 static void
1191 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1192 {
1193         char buf[LACP_LAGIDSTR_MAX+1];
1194
1195         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1196             __func__,
1197             lacp_format_lagid(&la->la_actor, &la->la_partner,
1198             buf, sizeof(buf)),
1199             la->la_refcnt, la->la_refcnt - 1));
1200
1201         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1202         la->la_refcnt--;
1203         if (la->la_refcnt > 0) {
1204                 return;
1205         }
1206
1207         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1208         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1209
1210         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1211
1212         free(la, M_DEVBUF);
1213 }
1214
1215 /*
1216  * lacp_aggregator_get: allocate an aggregator.
1217  */
1218
1219 static struct lacp_aggregator *
1220 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1221 {
1222         struct lacp_aggregator *la;
1223
1224         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1225         if (la) {
1226                 la->la_refcnt = 1;
1227                 la->la_nports = 0;
1228                 TAILQ_INIT(&la->la_ports);
1229                 la->la_pending = 0;
1230                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1231         }
1232
1233         return (la);
1234 }
1235
1236 /*
1237  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1238  */
1239
1240 static void
1241 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1242 {
1243         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1244         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1245
1246         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1247 }
1248
1249 static void
1250 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1251     const struct lacp_peerinfo *lpi_port)
1252 {
1253         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1254         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1255         lpi_aggr->lip_key = lpi_port->lip_key;
1256 }
1257
1258 /*
1259  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1260  */
1261
1262 static int
1263 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1264     const struct lacp_port *lp)
1265 {
1266         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1267             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1268                 return (0);
1269         }
1270
1271         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1272                 return (0);
1273         }
1274
1275         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1276                 return (0);
1277         }
1278
1279         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1280                 return (0);
1281         }
1282
1283         return (1);
1284 }
1285
1286 static int
1287 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1288     const struct lacp_peerinfo *b)
1289 {
1290         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1291             sizeof(a->lip_systemid))) {
1292                 return (0);
1293         }
1294
1295         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1296                 return (0);
1297         }
1298
1299         return (1);
1300 }
1301
1302 static void
1303 lacp_port_enable(struct lacp_port *lp)
1304 {
1305         lp->lp_state |= LACP_STATE_AGGREGATION;
1306 }
1307
1308 static void
1309 lacp_port_disable(struct lacp_port *lp)
1310 {
1311         lacp_set_mux(lp, LACP_MUX_DETACHED);
1312
1313         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1314         lp->lp_selected = LACP_UNSELECTED;
1315         lacp_sm_rx_record_default(lp);
1316         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1317         lp->lp_state &= ~LACP_STATE_EXPIRED;
1318 }
1319
1320 /*
1321  * lacp_select: select an aggregator.  create one if necessary.
1322  */
1323 static void
1324 lacp_select(struct lacp_port *lp)
1325 {
1326         struct lacp_softc *lsc = lp->lp_lsc;
1327         struct lacp_aggregator *la;
1328         char buf[LACP_LAGIDSTR_MAX+1];
1329
1330         if (lp->lp_aggregator) {
1331                 return;
1332         }
1333
1334         /* If we haven't heard from our peer, skip this step. */
1335         if (lp->lp_state & LACP_STATE_DEFAULTED)
1336                 return;
1337
1338         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1339             ("timer_wait_while still active"));
1340
1341         LACP_DPRINTF((lp, "port lagid=%s\n",
1342             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1343             buf, sizeof(buf))));
1344
1345         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1346                 if (lacp_aggregator_is_compatible(la, lp)) {
1347                         break;
1348                 }
1349         }
1350
1351         if (la == NULL) {
1352                 la = lacp_aggregator_get(lsc, lp);
1353                 if (la == NULL) {
1354                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1355
1356                         /*
1357                          * will retry on the next tick.
1358                          */
1359
1360                         return;
1361                 }
1362                 lacp_fill_aggregator_id(la, lp);
1363                 LACP_DPRINTF((lp, "aggregator created\n"));
1364         } else {
1365                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1366                 if (la->la_refcnt == LACP_MAX_PORTS)
1367                         return;
1368                 lacp_aggregator_addref(lsc, la);
1369         }
1370
1371         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1372             lacp_format_lagid(&la->la_actor, &la->la_partner,
1373             buf, sizeof(buf))));
1374
1375         lp->lp_aggregator = la;
1376         lp->lp_selected = LACP_SELECTED;
1377 }
1378
1379 /*
1380  * lacp_unselect: finish unselect/detach process.
1381  */
1382
1383 static void
1384 lacp_unselect(struct lacp_port *lp)
1385 {
1386         struct lacp_softc *lsc = lp->lp_lsc;
1387         struct lacp_aggregator *la = lp->lp_aggregator;
1388
1389         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1390             ("timer_wait_while still active"));
1391
1392         if (la == NULL) {
1393                 return;
1394         }
1395
1396         lp->lp_aggregator = NULL;
1397         lacp_aggregator_delref(lsc, la);
1398 }
1399
1400 /* mux machine */
1401
1402 static void
1403 lacp_sm_mux(struct lacp_port *lp)
1404 {
1405         struct lagg_port *lgp = lp->lp_lagg;
1406         struct lagg_softc *sc = lgp->lp_softc;
1407         enum lacp_mux_state new_state;
1408         boolean_t p_sync =
1409                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1410         boolean_t p_collecting =
1411             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1412         enum lacp_selected selected = lp->lp_selected;
1413         struct lacp_aggregator *la;
1414
1415         if (V_lacp_debug > 1)
1416                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1417                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1418                     lp->lp_mux_state, selected, p_sync, p_collecting);
1419
1420 re_eval:
1421         la = lp->lp_aggregator;
1422         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1423             ("MUX not detached"));
1424         new_state = lp->lp_mux_state;
1425         switch (lp->lp_mux_state) {
1426         case LACP_MUX_DETACHED:
1427                 if (selected != LACP_UNSELECTED) {
1428                         new_state = LACP_MUX_WAITING;
1429                 }
1430                 break;
1431         case LACP_MUX_WAITING:
1432                 KASSERT(la->la_pending > 0 ||
1433                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1434                     ("timer_wait_while still active"));
1435                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1436                         new_state = LACP_MUX_ATTACHED;
1437                 } else if (selected == LACP_UNSELECTED) {
1438                         new_state = LACP_MUX_DETACHED;
1439                 }
1440                 break;
1441         case LACP_MUX_ATTACHED:
1442                 if (selected == LACP_SELECTED && p_sync) {
1443                         new_state = LACP_MUX_COLLECTING;
1444                 } else if (selected != LACP_SELECTED) {
1445                         new_state = LACP_MUX_DETACHED;
1446                 }
1447                 break;
1448         case LACP_MUX_COLLECTING:
1449                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1450                         new_state = LACP_MUX_DISTRIBUTING;
1451                 } else if (selected != LACP_SELECTED || !p_sync) {
1452                         new_state = LACP_MUX_ATTACHED;
1453                 }
1454                 break;
1455         case LACP_MUX_DISTRIBUTING:
1456                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1457                         new_state = LACP_MUX_COLLECTING;
1458                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1459                         sc->sc_flapping++;
1460                 }
1461                 break;
1462         default:
1463                 panic("%s: unknown state", __func__);
1464         }
1465
1466         if (lp->lp_mux_state == new_state) {
1467                 return;
1468         }
1469
1470         lacp_set_mux(lp, new_state);
1471         goto re_eval;
1472 }
1473
1474 static void
1475 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1476 {
1477         struct lacp_aggregator *la = lp->lp_aggregator;
1478
1479         if (lp->lp_mux_state == new_state) {
1480                 return;
1481         }
1482
1483         switch (new_state) {
1484         case LACP_MUX_DETACHED:
1485                 lp->lp_state &= ~LACP_STATE_SYNC;
1486                 lacp_disable_distributing(lp);
1487                 lacp_disable_collecting(lp);
1488                 lacp_sm_assert_ntt(lp);
1489                 /* cancel timer */
1490                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1491                         KASSERT(la->la_pending > 0,
1492                             ("timer_wait_while not active"));
1493                         la->la_pending--;
1494                 }
1495                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1496                 lacp_unselect(lp);
1497                 break;
1498         case LACP_MUX_WAITING:
1499                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1500                     LACP_AGGREGATE_WAIT_TIME);
1501                 la->la_pending++;
1502                 break;
1503         case LACP_MUX_ATTACHED:
1504                 lp->lp_state |= LACP_STATE_SYNC;
1505                 lacp_disable_collecting(lp);
1506                 lacp_sm_assert_ntt(lp);
1507                 break;
1508         case LACP_MUX_COLLECTING:
1509                 lacp_enable_collecting(lp);
1510                 lacp_disable_distributing(lp);
1511                 lacp_sm_assert_ntt(lp);
1512                 break;
1513         case LACP_MUX_DISTRIBUTING:
1514                 lacp_enable_distributing(lp);
1515                 break;
1516         default:
1517                 panic("%s: unknown state", __func__);
1518         }
1519
1520         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1521
1522         lp->lp_mux_state = new_state;
1523 }
1524
1525 static void
1526 lacp_sm_mux_timer(struct lacp_port *lp)
1527 {
1528         struct lacp_aggregator *la = lp->lp_aggregator;
1529         char buf[LACP_LAGIDSTR_MAX+1];
1530
1531         KASSERT(la->la_pending > 0, ("no pending event"));
1532
1533         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1534             lacp_format_lagid(&la->la_actor, &la->la_partner,
1535             buf, sizeof(buf)),
1536             la->la_pending, la->la_pending - 1));
1537
1538         la->la_pending--;
1539 }
1540
1541 /* periodic transmit machine */
1542
1543 static void
1544 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1545 {
1546         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1547             LACP_STATE_TIMEOUT)) {
1548                 return;
1549         }
1550
1551         LACP_DPRINTF((lp, "partner timeout changed\n"));
1552
1553         /*
1554          * FAST_PERIODIC -> SLOW_PERIODIC
1555          * or
1556          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1557          *
1558          * let lacp_sm_ptx_tx_schedule to update timeout.
1559          */
1560
1561         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1562
1563         /*
1564          * if timeout has been shortened, assert NTT.
1565          */
1566
1567         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1568                 lacp_sm_assert_ntt(lp);
1569         }
1570 }
1571
1572 static void
1573 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1574 {
1575         int timeout;
1576
1577         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1578             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1579
1580                 /*
1581                  * NO_PERIODIC
1582                  */
1583
1584                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1585                 return;
1586         }
1587
1588         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1589                 return;
1590         }
1591
1592         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1593             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1594
1595         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1596 }
1597
1598 static void
1599 lacp_sm_ptx_timer(struct lacp_port *lp)
1600 {
1601         lacp_sm_assert_ntt(lp);
1602 }
1603
1604 static void
1605 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1606 {
1607         int timeout;
1608
1609         /*
1610          * check LACP_DISABLED first
1611          */
1612
1613         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1614                 return;
1615         }
1616
1617         /*
1618          * check loopback condition.
1619          */
1620
1621         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1622             &lp->lp_actor.lip_systemid)) {
1623                 return;
1624         }
1625
1626         /*
1627          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1628          */
1629
1630         lacp_sm_rx_update_selected(lp, du);
1631         lacp_sm_rx_update_ntt(lp, du);
1632         lacp_sm_rx_record_pdu(lp, du);
1633
1634         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1635             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1636         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1637
1638         lp->lp_state &= ~LACP_STATE_EXPIRED;
1639
1640         /*
1641          * kick transmit machine without waiting the next tick.
1642          */
1643
1644         lacp_sm_tx(lp);
1645 }
1646
1647 static void
1648 lacp_sm_rx_set_expired(struct lacp_port *lp)
1649 {
1650         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1651         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1652         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1653         lp->lp_state |= LACP_STATE_EXPIRED;
1654 }
1655
1656 static void
1657 lacp_sm_rx_timer(struct lacp_port *lp)
1658 {
1659         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1660                 /* CURRENT -> EXPIRED */
1661                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1662                 lacp_sm_rx_set_expired(lp);
1663         } else {
1664                 /* EXPIRED -> DEFAULTED */
1665                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1666                 lacp_sm_rx_update_default_selected(lp);
1667                 lacp_sm_rx_record_default(lp);
1668                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1669         }
1670 }
1671
1672 static void
1673 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1674 {
1675         boolean_t active;
1676         uint8_t oldpstate;
1677         char buf[LACP_STATESTR_MAX+1];
1678
1679         LACP_TRACE(lp);
1680
1681         oldpstate = lp->lp_partner.lip_state;
1682
1683         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1684             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1685             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1686
1687         lp->lp_partner = du->ldu_actor;
1688         if (active &&
1689             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1690             LACP_STATE_AGGREGATION) &&
1691             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1692             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1693                 /*
1694                  * XXX Maintain legacy behavior of leaving the
1695                  * LACP_STATE_SYNC bit unchanged from the partner's
1696                  * advertisement if lsc_strict_mode is false.
1697                  * TODO: We should re-examine the concept of the "strict mode"
1698                  * to ensure it makes sense to maintain a non-strict mode.
1699                  */
1700                 if (lp->lp_lsc->lsc_strict_mode)
1701                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1702         } else {
1703                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1704         }
1705
1706         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1707
1708         if (oldpstate != lp->lp_partner.lip_state) {
1709                 LACP_DPRINTF((lp, "old pstate %s\n",
1710                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1711                 LACP_DPRINTF((lp, "new pstate %s\n",
1712                     lacp_format_state(lp->lp_partner.lip_state, buf,
1713                     sizeof(buf))));
1714         }
1715
1716         lacp_sm_ptx_update_timeout(lp, oldpstate);
1717 }
1718
1719 static void
1720 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1721 {
1722
1723         LACP_TRACE(lp);
1724
1725         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1726             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1727             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1728                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1729                 lacp_sm_assert_ntt(lp);
1730         }
1731 }
1732
1733 static void
1734 lacp_sm_rx_record_default(struct lacp_port *lp)
1735 {
1736         uint8_t oldpstate;
1737
1738         LACP_TRACE(lp);
1739
1740         oldpstate = lp->lp_partner.lip_state;
1741         if (lp->lp_lsc->lsc_strict_mode)
1742                 lp->lp_partner = lacp_partner_admin_strict;
1743         else
1744                 lp->lp_partner = lacp_partner_admin_optimistic;
1745         lp->lp_state |= LACP_STATE_DEFAULTED;
1746         lacp_sm_ptx_update_timeout(lp, oldpstate);
1747 }
1748
1749 static void
1750 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1751     const struct lacp_peerinfo *info)
1752 {
1753
1754         LACP_TRACE(lp);
1755
1756         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1757             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1758             LACP_STATE_AGGREGATION)) {
1759                 lp->lp_selected = LACP_UNSELECTED;
1760                 /* mux machine will clean up lp->lp_aggregator */
1761         }
1762 }
1763
1764 static void
1765 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1766 {
1767
1768         LACP_TRACE(lp);
1769
1770         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1771 }
1772
1773 static void
1774 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1775 {
1776
1777         LACP_TRACE(lp);
1778
1779         if (lp->lp_lsc->lsc_strict_mode)
1780                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1781                     &lacp_partner_admin_strict);
1782         else
1783                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1784                     &lacp_partner_admin_optimistic);
1785 }
1786
1787 /* transmit machine */
1788
1789 static void
1790 lacp_sm_tx(struct lacp_port *lp)
1791 {
1792         int error = 0;
1793
1794         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1795 #if 1
1796             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1797             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1798 #endif
1799             ) {
1800                 lp->lp_flags &= ~LACP_PORT_NTT;
1801         }
1802
1803         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1804                 return;
1805         }
1806
1807         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1808         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1809                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1810                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1811                 return;
1812         }
1813
1814         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1815                 error = lacp_xmit_lacpdu(lp);
1816         } else {
1817                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1818         }
1819
1820         if (error == 0) {
1821                 lp->lp_flags &= ~LACP_PORT_NTT;
1822         } else {
1823                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1824                     error));
1825         }
1826 }
1827
1828 static void
1829 lacp_sm_assert_ntt(struct lacp_port *lp)
1830 {
1831
1832         lp->lp_flags |= LACP_PORT_NTT;
1833 }
1834
1835 static void
1836 lacp_run_timers(struct lacp_port *lp)
1837 {
1838         int i;
1839
1840         for (i = 0; i < LACP_NTIMER; i++) {
1841                 KASSERT(lp->lp_timer[i] >= 0,
1842                     ("invalid timer value %d", lp->lp_timer[i]));
1843                 if (lp->lp_timer[i] == 0) {
1844                         continue;
1845                 } else if (--lp->lp_timer[i] <= 0) {
1846                         if (lacp_timer_funcs[i]) {
1847                                 (*lacp_timer_funcs[i])(lp);
1848                         }
1849                 }
1850         }
1851 }
1852
1853 int
1854 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1855 {
1856         struct lacp_softc *lsc = lp->lp_lsc;
1857         struct lagg_port *lgp = lp->lp_lagg;
1858         struct lacp_port *lp2;
1859         struct markerdu *mdu;
1860         int error = 0;
1861         int pending = 0;
1862
1863         if (m->m_pkthdr.len != sizeof(*mdu)) {
1864                 goto bad;
1865         }
1866
1867         if ((m->m_flags & M_MCAST) == 0) {
1868                 goto bad;
1869         }
1870
1871         if (m->m_len < sizeof(*mdu)) {
1872                 m = m_pullup(m, sizeof(*mdu));
1873                 if (m == NULL) {
1874                         return (ENOMEM);
1875                 }
1876         }
1877
1878         mdu = mtod(m, struct markerdu *);
1879
1880         if (memcmp(&mdu->mdu_eh.ether_dhost,
1881             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1882                 goto bad;
1883         }
1884
1885         if (mdu->mdu_sph.sph_version != 1) {
1886                 goto bad;
1887         }
1888
1889         switch (mdu->mdu_tlv.tlv_type) {
1890         case MARKER_TYPE_INFO:
1891                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1892                     marker_info_tlv_template, TRUE)) {
1893                         goto bad;
1894                 }
1895                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1896                 memcpy(&mdu->mdu_eh.ether_dhost,
1897                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1898                 memcpy(&mdu->mdu_eh.ether_shost,
1899                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1900                 error = lagg_enqueue(lp->lp_ifp, m);
1901                 break;
1902
1903         case MARKER_TYPE_RESPONSE:
1904                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1905                     marker_response_tlv_template, TRUE)) {
1906                         goto bad;
1907                 }
1908                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1909                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1910                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1911
1912                 /* Verify that it is the last marker we sent out */
1913                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1914                     sizeof(struct lacp_markerinfo)))
1915                         goto bad;
1916
1917                 LACP_LOCK(lsc);
1918                 lp->lp_flags &= ~LACP_PORT_MARK;
1919
1920                 if (lsc->lsc_suppress_distributing) {
1921                         /* Check if any ports are waiting for a response */
1922                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1923                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1924                                         pending = 1;
1925                                         break;
1926                                 }
1927                         }
1928
1929                         if (pending == 0) {
1930                                 /* All interface queues are clear */
1931                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1932                                 lsc->lsc_suppress_distributing = FALSE;
1933                         }
1934                 }
1935                 LACP_UNLOCK(lsc);
1936                 m_freem(m);
1937                 break;
1938
1939         default:
1940                 goto bad;
1941         }
1942
1943         return (error);
1944
1945 bad:
1946         LACP_DPRINTF((lp, "bad marker frame\n"));
1947         m_freem(m);
1948         return (EINVAL);
1949 }
1950
1951 static int
1952 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1953     const struct tlv_template *tmpl, boolean_t check_type)
1954 {
1955         while (/* CONSTCOND */ 1) {
1956                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1957                         return (EINVAL);
1958                 }
1959                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1960                     tlv->tlv_length != tmpl->tmpl_length) {
1961                         return (EINVAL);
1962                 }
1963                 if (tmpl->tmpl_type == 0) {
1964                         break;
1965                 }
1966                 tlv = (const struct tlvhdr *)
1967                     ((const char *)tlv + tlv->tlv_length);
1968                 tmpl++;
1969         }
1970
1971         return (0);
1972 }
1973
1974 /* Debugging */
1975 const char *
1976 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1977 {
1978         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1979             (int)mac[0],
1980             (int)mac[1],
1981             (int)mac[2],
1982             (int)mac[3],
1983             (int)mac[4],
1984             (int)mac[5]);
1985
1986         return (buf);
1987 }
1988
1989 const char *
1990 lacp_format_systemid(const struct lacp_systemid *sysid,
1991     char *buf, size_t buflen)
1992 {
1993         char macbuf[LACP_MACSTR_MAX+1];
1994
1995         snprintf(buf, buflen, "%04X,%s",
1996             ntohs(sysid->lsi_prio),
1997             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1998
1999         return (buf);
2000 }
2001
2002 const char *
2003 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2004 {
2005         snprintf(buf, buflen, "%04X,%04X",
2006             ntohs(portid->lpi_prio),
2007             ntohs(portid->lpi_portno));
2008
2009         return (buf);
2010 }
2011
2012 const char *
2013 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2014 {
2015         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2016         char portid[LACP_PORTIDSTR_MAX+1];
2017
2018         snprintf(buf, buflen, "(%s,%04X,%s)",
2019             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2020             ntohs(peer->lip_key),
2021             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2022
2023         return (buf);
2024 }
2025
2026 const char *
2027 lacp_format_lagid(const struct lacp_peerinfo *a,
2028     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2029 {
2030         char astr[LACP_PARTNERSTR_MAX+1];
2031         char bstr[LACP_PARTNERSTR_MAX+1];
2032
2033 #if 0
2034         /*
2035          * there's a convention to display small numbered peer
2036          * in the left.
2037          */
2038
2039         if (lacp_compare_peerinfo(a, b) > 0) {
2040                 const struct lacp_peerinfo *t;
2041
2042                 t = a;
2043                 a = b;
2044                 b = t;
2045         }
2046 #endif
2047
2048         snprintf(buf, buflen, "[%s,%s]",
2049             lacp_format_partner(a, astr, sizeof(astr)),
2050             lacp_format_partner(b, bstr, sizeof(bstr)));
2051
2052         return (buf);
2053 }
2054
2055 const char *
2056 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2057     char *buf, size_t buflen)
2058 {
2059         if (la == NULL) {
2060                 return ("(none)");
2061         }
2062
2063         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2064 }
2065
2066 const char *
2067 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2068 {
2069         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2070         return (buf);
2071 }
2072
2073 static void
2074 lacp_dump_lacpdu(const struct lacpdu *du)
2075 {
2076         char buf[LACP_PARTNERSTR_MAX+1];
2077         char buf2[LACP_STATESTR_MAX+1];
2078
2079         printf("actor=%s\n",
2080             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2081         printf("actor.state=%s\n",
2082             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2083         printf("partner=%s\n",
2084             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2085         printf("partner.state=%s\n",
2086             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2087
2088         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2089 }
2090
2091 static void
2092 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2093 {
2094         va_list va;
2095
2096         if (lp) {
2097                 printf("%s: ", lp->lp_ifp->if_xname);
2098         }
2099
2100         va_start(va, fmt);
2101         vprintf(fmt, va);
2102         va_end(va);
2103 }