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/mbuf.h>
  36 #include <sys/systm.h>
  37 #include <sys/malloc.h>
  38 #include <sys/kernel.h> /* hz */
  39 #include <sys/socket.h> /* for net/if.h */
  40 #include <sys/sockio.h>
  41 #include <sys/sysctl.h>
  42 #include <machine/stdarg.h>
  43 #include <sys/lock.h>
  44 #include <sys/rwlock.h>
  45
  46 #include <net/if.h>
  47 #include <net/if_dl.h>
  48 #include <net/ethernet.h>
  49 #include <net/if_media.h>
  50 #include <net/if_types.h>
  51
  52 #include <net/if_lagg.h>
  53 #include <net/ieee8023ad_lacp.h>
  54
  55 /*
  56  * actor system priority and port priority.
  57  * XXX should be configurable.
  58  */
  59
  60 #define LACP_SYSTEM_PRIO        0x8000
  61 #define LACP_PORT_PRIO          0x8000
  62
  63 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  64     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  65
  66 static const struct tlv_template lacp_info_tlv_template[] = {
  67         { LACP_TYPE_ACTORINFO,
  68             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  69         { LACP_TYPE_PARTNERINFO,
  70             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  71         { LACP_TYPE_COLLECTORINFO,
  72             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  73         { 0, 0 },
  74 };
  75
  76 static const struct tlv_template marker_info_tlv_template[] = {
  77         { MARKER_TYPE_INFO,
  78             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  79         { 0, 0 },
  80 };
  81
  82 static const struct tlv_template marker_response_tlv_template[] = {
  83         { MARKER_TYPE_RESPONSE,
  84             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  85         { 0, 0 },
  86 };
  87
  88 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  89
  90 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
  91 static void     lacp_fill_markerinfo(struct lacp_port *,
  92                     struct lacp_markerinfo *);
  93
  94 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
  95 static void     lacp_suppress_distributing(struct lacp_softc *,
  96                     struct lacp_aggregator *);
  97 static void     lacp_transit_expire(void *);
  98 static void     lacp_update_portmap(struct lacp_softc *);
  99 static void     lacp_select_active_aggregator(struct lacp_softc *);
 100 static uint16_t lacp_compose_key(struct lacp_port *);
 101 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 102                     const struct tlv_template *, boolean_t);
 103 static void     lacp_tick(void *);
 104
 105 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 106                     const struct lacp_port *);
 107 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 108                     const struct lacp_peerinfo *);
 109 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 110                     const struct lacp_port *);
 111 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 112                     const struct lacp_peerinfo *);
 113
 114 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 115                     struct lacp_port *);
 116 static void     lacp_aggregator_addref(struct lacp_softc *,
 117                     struct lacp_aggregator *);
 118 static void     lacp_aggregator_delref(struct lacp_softc *,
 119                     struct lacp_aggregator *);
 120
 121 /* receive machine */
 122
 123 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 124 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 125 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 126 static void     lacp_sm_rx_timer(struct lacp_port *);
 127 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 128 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 129                     const struct lacpdu *);
 130 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 131                     const struct lacpdu *);
 132 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 133                     const struct lacpdu *);
 134 static void     lacp_sm_rx_record_default(struct lacp_port *);
 135 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 136 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 137                     const struct lacp_peerinfo *);
 138
 139 /* mux machine */
 140
 141 static void     lacp_sm_mux(struct lacp_port *);
 142 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 143 static void     lacp_sm_mux_timer(struct lacp_port *);
 144
 145 /* periodic transmit machine */
 146
 147 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 148 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 149 static void     lacp_sm_ptx_timer(struct lacp_port *);
 150
 151 /* transmit machine */
 152
 153 static void     lacp_sm_tx(struct lacp_port *);
 154 static void     lacp_sm_assert_ntt(struct lacp_port *);
 155
 156 static void     lacp_run_timers(struct lacp_port *);
 157 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 158                     const struct lacp_peerinfo *);
 159 static int      lacp_compare_systemid(const struct lacp_systemid *,
 160                     const struct lacp_systemid *);
 161 static void     lacp_port_enable(struct lacp_port *);
 162 static void     lacp_port_disable(struct lacp_port *);
 163 static void     lacp_select(struct lacp_port *);
 164 static void     lacp_unselect(struct lacp_port *);
 165 static void     lacp_disable_collecting(struct lacp_port *);
 166 static void     lacp_enable_collecting(struct lacp_port *);
 167 static void     lacp_disable_distributing(struct lacp_port *);
 168 static void     lacp_enable_distributing(struct lacp_port *);
 169 static int      lacp_xmit_lacpdu(struct lacp_port *);
 170 static int      lacp_xmit_marker(struct lacp_port *);
 171
 172 /* Debugging */
 173
 174 static void     lacp_dump_lacpdu(const struct lacpdu *);
 175 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 176                     size_t);
 177 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 178                     const struct lacp_peerinfo *, char *, size_t);
 179 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 180                     char *, size_t);
 181 static const char *lacp_format_state(uint8_t, char *, size_t);
 182 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 183 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 184                     size_t);
 185 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 186                     size_t);
 187 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 188                     __attribute__((__format__(__printf__, 2, 3)));
 189
 190 static int lacp_debug = 0;
 191 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
 192 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_TUN,
 193     &lacp_debug, 0, "Enable LACP debug logging (1=debug, 2=trace)");
 194 TUNABLE_INT("net.link.lagg.lacp.debug", &lacp_debug);
 195
 196 #define LACP_DPRINTF(a) if (lacp_debug & 0x01) { lacp_dprintf a ; }
 197 #define LACP_TRACE(a) if (lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 198 #define LACP_TPRINTF(a) if (lacp_debug & 0x04) { lacp_dprintf a ; }
 199
 200 /*
 201  * partner administration variables.
 202  * XXX should be configurable.
 203  */
 204
 205 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 206         .lip_systemid = { .lsi_prio = 0xffff },
 207         .lip_portid = { .lpi_prio = 0xffff },
 208         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 209             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 210 };
 211
 212 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 213         .lip_systemid = { .lsi_prio = 0xffff },
 214         .lip_portid = { .lpi_prio = 0xffff },
 215         .lip_state = 0,
 216 };
 217
 218 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 219         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 220         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 221         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 222 };
 223
 224 struct mbuf *
 225 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 226 {
 227         struct lacp_port *lp = LACP_PORT(lgp);
 228         uint8_t subtype;
 229
 230         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 231                 m_freem(m);
 232                 return (NULL);
 233         }
 234
 235         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 236         switch (subtype) {
 237                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 238                         lacp_pdu_input(lp, m);
 239                         return (NULL);
 240
 241                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 242                         lacp_marker_input(lp, m);
 243                         return (NULL);
 244         }
 245
 246         /* Not a subtype we are interested in */
 247         return (m);
 248 }
 249
 250 /*
 251  * lacp_pdu_input: process lacpdu
 252  */
 253 static int
 254 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 255 {
 256         struct lacp_softc *lsc = lp->lp_lsc;
 257         struct lacpdu *du;
 258         int error = 0;
 259
 260         if (m->m_pkthdr.len != sizeof(*du)) {
 261                 goto bad;
 262         }
 263
 264         if ((m->m_flags & M_MCAST) == 0) {
 265                 goto bad;
 266         }
 267
 268         if (m->m_len < sizeof(*du)) {
 269                 m = m_pullup(m, sizeof(*du));
 270                 if (m == NULL) {
 271                         return (ENOMEM);
 272                 }
 273         }
 274
 275         du = mtod(m, struct lacpdu *);
 276
 277         if (memcmp(&du->ldu_eh.ether_dhost,
 278             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 279                 goto bad;
 280         }
 281
 282         /*
 283          * ignore the version for compatibility with
 284          * the future protocol revisions.
 285          */
 286 #if 0
 287         if (du->ldu_sph.sph_version != 1) {
 288                 goto bad;
 289         }
 290 #endif
 291
 292         /*
 293          * ignore tlv types for compatibility with
 294          * the future protocol revisions.
 295          */
 296         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 297             lacp_info_tlv_template, FALSE)) {
 298                 goto bad;
 299         }
 300
 301         if (lacp_debug > 0) {
 302                 lacp_dprintf(lp, "lacpdu receive\n");
 303                 lacp_dump_lacpdu(du);
 304         }
 305
 306         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 307                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 308                 goto bad;
 309         }
 310
 311         LACP_LOCK(lsc);
 312         lacp_sm_rx(lp, du);
 313         LACP_UNLOCK(lsc);
 314
 315         m_freem(m);
 316         return (error);
 317
 318 bad:
 319         m_freem(m);
 320         return (EINVAL);
 321 }
 322
 323 static void
 324 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 325 {
 326         struct lagg_port *lgp = lp->lp_lagg;
 327         struct lagg_softc *sc = lgp->lp_softc;
 328
 329         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 330         memcpy(&info->lip_systemid.lsi_mac,
 331             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 332         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 333         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 334         info->lip_state = lp->lp_state;
 335 }
 336
 337 static void
 338 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 339 {
 340         struct ifnet *ifp = lp->lp_ifp;
 341
 342         /* Fill in the port index and system id (encoded as the MAC) */
 343         info->mi_rq_port = htons(ifp->if_index);
 344         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 345         info->mi_rq_xid = htonl(0);
 346 }
 347
 348 static int
 349 lacp_xmit_lacpdu(struct lacp_port *lp)
 350 {
 351         struct lagg_port *lgp = lp->lp_lagg;
 352         struct mbuf *m;
 353         struct lacpdu *du;
 354         int error;
 355
 356         LACP_LOCK_ASSERT(lp->lp_lsc);
 357
 358         m = m_gethdr(M_NOWAIT, MT_DATA);
 359         if (m == NULL) {
 360                 return (ENOMEM);
 361         }
 362         m->m_len = m->m_pkthdr.len = sizeof(*du);
 363
 364         du = mtod(m, struct lacpdu *);
 365         memset(du, 0, sizeof(*du));
 366
 367         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 368             ETHER_ADDR_LEN);
 369         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 370         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 371
 372         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 373         du->ldu_sph.sph_version = 1;
 374
 375         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 376         du->ldu_actor = lp->lp_actor;
 377
 378         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 379             sizeof(du->ldu_partner));
 380         du->ldu_partner = lp->lp_partner;
 381
 382         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 383             sizeof(du->ldu_collector));
 384         du->ldu_collector.lci_maxdelay = 0;
 385
 386         if (lacp_debug > 0) {
 387                 lacp_dprintf(lp, "lacpdu transmit\n");
 388                 lacp_dump_lacpdu(du);
 389         }
 390
 391         m->m_flags |= M_MCAST;
 392
 393         /*
 394          * XXX should use higher priority queue.
 395          * otherwise network congestion can break aggregation.
 396          */
 397
 398         error = lagg_enqueue(lp->lp_ifp, m);
 399         return (error);
 400 }
 401
 402 static int
 403 lacp_xmit_marker(struct lacp_port *lp)
 404 {
 405         struct lagg_port *lgp = lp->lp_lagg;
 406         struct mbuf *m;
 407         struct markerdu *mdu;
 408         int error;
 409
 410         LACP_LOCK_ASSERT(lp->lp_lsc);
 411
 412         m = m_gethdr(M_NOWAIT, MT_DATA);
 413         if (m == NULL) {
 414                 return (ENOMEM);
 415         }
 416         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 417
 418         mdu = mtod(m, struct markerdu *);
 419         memset(mdu, 0, sizeof(*mdu));
 420
 421         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 422             ETHER_ADDR_LEN);
 423         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 424         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 425
 426         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 427         mdu->mdu_sph.sph_version = 1;
 428
 429         /* Bump the transaction id and copy over the marker info */
 430         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 431         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 432         mdu->mdu_info = lp->lp_marker;
 433
 434         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 435             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 436             ntohl(mdu->mdu_info.mi_rq_xid)));
 437
 438         m->m_flags |= M_MCAST;
 439         error = lagg_enqueue(lp->lp_ifp, m);
 440         return (error);
 441 }
 442
 443 void
 444 lacp_linkstate(struct lagg_port *lgp)
 445 {
 446         struct lacp_port *lp = LACP_PORT(lgp);
 447         struct lacp_softc *lsc = lp->lp_lsc;
 448         struct ifnet *ifp = lgp->lp_ifp;
 449         struct ifmediareq ifmr;
 450         int error = 0;
 451         u_int media;
 452         uint8_t old_state;
 453         uint16_t old_key;
 454
 455         bzero((char *)&ifmr, sizeof(ifmr));
 456         error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 457         if (error != 0)
 458                 return;
 459
 460         LACP_LOCK(lsc);
 461         media = ifmr.ifm_active;
 462         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 463             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 464             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 465         old_state = lp->lp_state;
 466         old_key = lp->lp_key;
 467
 468         lp->lp_media = media;
 469         /*
 470          * If the port is not an active full duplex Ethernet link then it can
 471          * not be aggregated.
 472          */
 473         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 474             ifp->if_link_state != LINK_STATE_UP) {
 475                 lacp_port_disable(lp);
 476         } else {
 477                 lacp_port_enable(lp);
 478         }
 479         lp->lp_key = lacp_compose_key(lp);
 480
 481         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 482                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 483                 lp->lp_selected = LACP_UNSELECTED;
 484         }
 485         LACP_UNLOCK(lsc);
 486 }
 487
 488 static void
 489 lacp_tick(void *arg)
 490 {
 491         struct lacp_softc *lsc = arg;
 492         struct lacp_port *lp;
 493
 494         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 495                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 496                         continue;
 497
 498                 lacp_run_timers(lp);
 499
 500                 lacp_select(lp);
 501                 lacp_sm_mux(lp);
 502                 lacp_sm_tx(lp);
 503                 lacp_sm_ptx_tx_schedule(lp);
 504         }
 505         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 506 }
 507
 508 int
 509 lacp_port_create(struct lagg_port *lgp)
 510 {
 511         struct lagg_softc *sc = lgp->lp_softc;
 512         struct lacp_softc *lsc = LACP_SOFTC(sc);
 513         struct lacp_port *lp;
 514         struct ifnet *ifp = lgp->lp_ifp;
 515         struct sockaddr_dl sdl;
 516         struct ifmultiaddr *rifma = NULL;
 517         int error;
 518
 519         boolean_t active = TRUE; /* XXX should be configurable */
 520         boolean_t fast = FALSE; /* XXX should be configurable */
 521
 522         bzero((char *)&sdl, sizeof(sdl));
 523         sdl.sdl_len = sizeof(sdl);
 524         sdl.sdl_family = AF_LINK;
 525         sdl.sdl_index = ifp->if_index;
 526         sdl.sdl_type = IFT_ETHER;
 527         sdl.sdl_alen = ETHER_ADDR_LEN;
 528
 529         bcopy(&ethermulticastaddr_slowprotocols,
 530             LLADDR(&sdl), ETHER_ADDR_LEN);
 531         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 532         if (error) {
 533                 printf("%s: ADDMULTI failed on %s\n", __func__, lgp->lp_ifname);
 534                 return (error);
 535         }
 536
 537         lp = malloc(sizeof(struct lacp_port),
 538             M_DEVBUF, M_NOWAIT|M_ZERO);
 539         if (lp == NULL)
 540                 return (ENOMEM);
 541
 542         LACP_LOCK(lsc);
 543         lgp->lp_psc = (caddr_t)lp;
 544         lp->lp_ifp = ifp;
 545         lp->lp_lagg = lgp;
 546         lp->lp_lsc = lsc;
 547         lp->lp_ifma = rifma;
 548
 549         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 550
 551         lacp_fill_actorinfo(lp, &lp->lp_actor);
 552         lacp_fill_markerinfo(lp, &lp->lp_marker);
 553         lp->lp_state =
 554             (active ? LACP_STATE_ACTIVITY : 0) |
 555             (fast ? LACP_STATE_TIMEOUT : 0);
 556         lp->lp_aggregator = NULL;
 557         lacp_sm_rx_set_expired(lp);
 558         LACP_UNLOCK(lsc);
 559         lacp_linkstate(lgp);
 560
 561         return (0);
 562 }
 563
 564 void
 565 lacp_port_destroy(struct lagg_port *lgp)
 566 {
 567         struct lacp_port *lp = LACP_PORT(lgp);
 568         struct lacp_softc *lsc = lp->lp_lsc;
 569         int i;
 570
 571         LACP_LOCK(lsc);
 572         for (i = 0; i < LACP_NTIMER; i++) {
 573                 LACP_TIMER_DISARM(lp, i);
 574         }
 575
 576         lacp_disable_collecting(lp);
 577         lacp_disable_distributing(lp);
 578         lacp_unselect(lp);
 579
 580         /* The address may have already been removed by if_purgemaddrs() */
 581         if (!lgp->lp_detaching)
 582                 if_delmulti_ifma(lp->lp_ifma);
 583
 584         LIST_REMOVE(lp, lp_next);
 585         LACP_UNLOCK(lsc);
 586         free(lp, M_DEVBUF);
 587 }
 588
 589 void
 590 lacp_req(struct lagg_softc *sc, caddr_t data)
 591 {
 592         struct lacp_opreq *req = (struct lacp_opreq *)data;
 593         struct lacp_softc *lsc = LACP_SOFTC(sc);
 594         struct lacp_aggregator *la;
 595
 596         bzero(req, sizeof(struct lacp_opreq));
 597
 598         /*
 599          * If the LACP softc is NULL, return with the opreq structure full of
 600          * zeros.  It is normal for the softc to be NULL while the lagg is
 601          * being destroyed.
 602          */
 603         if (NULL == lsc)
 604                 return;
 605
 606         la = lsc->lsc_active_aggregator;
 607         LACP_LOCK(lsc);
 608         if (la != NULL) {
 609                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 610                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 611                     ETHER_ADDR_LEN);
 612                 req->actor_key = ntohs(la->la_actor.lip_key);
 613                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 614                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 615                 req->actor_state = la->la_actor.lip_state;
 616
 617                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 618                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 619                     ETHER_ADDR_LEN);
 620                 req->partner_key = ntohs(la->la_partner.lip_key);
 621                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 622                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 623                 req->partner_state = la->la_partner.lip_state;
 624         }
 625         LACP_UNLOCK(lsc);
 626 }
 627
 628 void
 629 lacp_portreq(struct lagg_port *lgp, caddr_t data)
 630 {
 631         struct lacp_opreq *req = (struct lacp_opreq *)data;
 632         struct lacp_port *lp = LACP_PORT(lgp);
 633         struct lacp_softc *lsc = lp->lp_lsc;
 634
 635         LACP_LOCK(lsc);
 636         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 637         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 638             ETHER_ADDR_LEN);
 639         req->actor_key = ntohs(lp->lp_actor.lip_key);
 640         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 641         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 642         req->actor_state = lp->lp_actor.lip_state;
 643
 644         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 645         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 646             ETHER_ADDR_LEN);
 647         req->partner_key = ntohs(lp->lp_partner.lip_key);
 648         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 649         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 650         req->partner_state = lp->lp_partner.lip_state;
 651         LACP_UNLOCK(lsc);
 652 }
 653
 654 static void
 655 lacp_disable_collecting(struct lacp_port *lp)
 656 {
 657         LACP_DPRINTF((lp, "collecting disabled\n"));
 658         lp->lp_state &= ~LACP_STATE_COLLECTING;
 659 }
 660
 661 static void
 662 lacp_enable_collecting(struct lacp_port *lp)
 663 {
 664         LACP_DPRINTF((lp, "collecting enabled\n"));
 665         lp->lp_state |= LACP_STATE_COLLECTING;
 666 }
 667
 668 static void
 669 lacp_disable_distributing(struct lacp_port *lp)
 670 {
 671         struct lacp_aggregator *la = lp->lp_aggregator;
 672         struct lacp_softc *lsc = lp->lp_lsc;
 673         struct lagg_softc *sc = lsc->lsc_softc;
 674         char buf[LACP_LAGIDSTR_MAX+1];
 675
 676         LACP_LOCK_ASSERT(lsc);
 677
 678         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 679                 return;
 680         }
 681
 682         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 683         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 684         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 685
 686         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 687             "nports %d -> %d\n",
 688             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 689             la->la_nports, la->la_nports - 1));
 690
 691         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 692         la->la_nports--;
 693         sc->sc_active = la->la_nports;
 694
 695         if (lsc->lsc_active_aggregator == la) {
 696                 lacp_suppress_distributing(lsc, la);
 697                 lacp_select_active_aggregator(lsc);
 698                 /* regenerate the port map, the active aggregator has changed */
 699                 lacp_update_portmap(lsc);
 700         }
 701
 702         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 703 }
 704
 705 static void
 706 lacp_enable_distributing(struct lacp_port *lp)
 707 {
 708         struct lacp_aggregator *la = lp->lp_aggregator;
 709         struct lacp_softc *lsc = lp->lp_lsc;
 710         struct lagg_softc *sc = lsc->lsc_softc;
 711         char buf[LACP_LAGIDSTR_MAX+1];
 712
 713         LACP_LOCK_ASSERT(lsc);
 714
 715         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 716                 return;
 717         }
 718
 719         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 720             "nports %d -> %d\n",
 721             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 722             la->la_nports, la->la_nports + 1));
 723
 724         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 725         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 726         la->la_nports++;
 727         sc->sc_active = la->la_nports;
 728
 729         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 730
 731         if (lsc->lsc_active_aggregator == la) {
 732                 lacp_suppress_distributing(lsc, la);
 733                 lacp_update_portmap(lsc);
 734         } else
 735                 /* try to become the active aggregator */
 736                 lacp_select_active_aggregator(lsc);
 737 }
 738
 739 static void
 740 lacp_transit_expire(void *vp)
 741 {
 742         struct lacp_softc *lsc = vp;
 743
 744         LACP_LOCK_ASSERT(lsc);
 745
 746         LACP_TRACE(NULL);
 747
 748         lsc->lsc_suppress_distributing = FALSE;
 749 }
 750
 751 static void
 752 lacp_attach_sysctl(struct lacp_softc *lsc, struct sysctl_oid *p_oid)
 753 {
 754         struct lagg_softc *sc = lsc->lsc_softc;
 755
 756         SYSCTL_ADD_UINT(&sc->ctx, SYSCTL_CHILDREN(p_oid), OID_AUTO,
 757             "lacp_strict_mode",
 758             CTLFLAG_RW,
 759             &lsc->lsc_strict_mode,
 760             lsc->lsc_strict_mode,
 761             "Enable LACP strict mode");
 762 }
 763
 764 static void
 765 lacp_attach_sysctl_debug(struct lacp_softc *lsc, struct sysctl_oid *p_oid)
 766 {
 767         struct lagg_softc *sc = lsc->lsc_softc;
 768         struct sysctl_oid *oid;
 769
 770         /* Create a child of the parent lagg interface */
 771         oid = SYSCTL_ADD_NODE(&sc->ctx, SYSCTL_CHILDREN(p_oid),
 772             OID_AUTO, "debug", CTLFLAG_RD, NULL, "DEBUG");
 773
 774         SYSCTL_ADD_UINT(&sc->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
 775             "rx_test",
 776             CTLFLAG_RW,
 777             &lsc->lsc_debug.lsc_rx_test,
 778             lsc->lsc_debug.lsc_rx_test,
 779             "Bitmap of if_dunit entries to drop RX frames for");
 780         SYSCTL_ADD_UINT(&sc->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
 781             "tx_test",
 782             CTLFLAG_RW,
 783             &lsc->lsc_debug.lsc_tx_test,
 784             lsc->lsc_debug.lsc_tx_test,
 785             "Bitmap of if_dunit entries to drop TX frames for");
 786 }
 787
 788 int
 789 lacp_attach(struct lagg_softc *sc)
 790 {
 791         struct lacp_softc *lsc;
 792         struct sysctl_oid *oid;
 793
 794         lsc = malloc(sizeof(struct lacp_softc),
 795             M_DEVBUF, M_NOWAIT|M_ZERO);
 796         if (lsc == NULL)
 797                 return (ENOMEM);
 798
 799         sc->sc_psc = (caddr_t)lsc;
 800         lsc->lsc_softc = sc;
 801
 802         lsc->lsc_hashkey = arc4random();
 803         lsc->lsc_active_aggregator = NULL;
 804         lsc->lsc_strict_mode = 1;
 805         LACP_LOCK_INIT(lsc);
 806         TAILQ_INIT(&lsc->lsc_aggregators);
 807         LIST_INIT(&lsc->lsc_ports);
 808
 809         /* Create a child of the parent lagg interface */
 810         oid = SYSCTL_ADD_NODE(&sc->ctx, SYSCTL_CHILDREN(sc->sc_oid),
 811             OID_AUTO, "lacp", CTLFLAG_RD, NULL, "LACP");
 812
 813         /* Attach sysctl nodes */
 814         lacp_attach_sysctl(lsc, oid);
 815         lacp_attach_sysctl_debug(lsc, oid);
 816
 817         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 818         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 819
 820         /* if the lagg is already up then do the same */
 821         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 822                 lacp_init(sc);
 823
 824         return (0);
 825 }
 826
 827 int
 828 lacp_detach(struct lagg_softc *sc)
 829 {
 830         struct lacp_softc *lsc = LACP_SOFTC(sc);
 831
 832         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 833             ("aggregators still active"));
 834         KASSERT(lsc->lsc_active_aggregator == NULL,
 835             ("aggregator still attached"));
 836
 837         sc->sc_psc = NULL;
 838         callout_drain(&lsc->lsc_transit_callout);
 839         callout_drain(&lsc->lsc_callout);
 840
 841         LACP_LOCK_DESTROY(lsc);
 842         free(lsc, M_DEVBUF);
 843         return (0);
 844 }
 845
 846 void
 847 lacp_init(struct lagg_softc *sc)
 848 {
 849         struct lacp_softc *lsc = LACP_SOFTC(sc);
 850
 851         LACP_LOCK(lsc);
 852         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 853         LACP_UNLOCK(lsc);
 854 }
 855
 856 void
 857 lacp_stop(struct lagg_softc *sc)
 858 {
 859         struct lacp_softc *lsc = LACP_SOFTC(sc);
 860
 861         LACP_LOCK(lsc);
 862         callout_stop(&lsc->lsc_transit_callout);
 863         callout_stop(&lsc->lsc_callout);
 864         LACP_UNLOCK(lsc);
 865 }
 866
 867 struct lagg_port *
 868 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 869 {
 870         struct lacp_softc *lsc = LACP_SOFTC(sc);
 871         struct lacp_portmap *pm;
 872         struct lacp_port *lp;
 873         uint32_t hash;
 874
 875         if (__predict_false(lsc->lsc_suppress_distributing)) {
 876                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 877                 return (NULL);
 878         }
 879
 880         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 881         if (pm->pm_count == 0) {
 882                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 883                 return (NULL);
 884         }
 885
 886         if (sc->use_flowid &&
 887             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 888                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 889         else
 890                 hash = lagg_hashmbuf(sc, m, lsc->lsc_hashkey);
 891         hash %= pm->pm_count;
 892         lp = pm->pm_map[hash];
 893
 894         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 895             ("aggregated port is not distributing"));
 896
 897         return (lp->lp_lagg);
 898 }
 899 /*
 900  * lacp_suppress_distributing: drop transmit packets for a while
 901  * to preserve packet ordering.
 902  */
 903
 904 static void
 905 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 906 {
 907         struct lacp_port *lp;
 908
 909         if (lsc->lsc_active_aggregator != la) {
 910                 return;
 911         }
 912
 913         LACP_TRACE(NULL);
 914
 915         lsc->lsc_suppress_distributing = TRUE;
 916
 917         /* send a marker frame down each port to verify the queues are empty */
 918         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 919                 lp->lp_flags |= LACP_PORT_MARK;
 920                 lacp_xmit_marker(lp);
 921         }
 922
 923         /* set a timeout for the marker frames */
 924         callout_reset(&lsc->lsc_transit_callout,
 925             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 926 }
 927
 928 static int
 929 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 930     const struct lacp_peerinfo *b)
 931 {
 932         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 933 }
 934
 935 static int
 936 lacp_compare_systemid(const struct lacp_systemid *a,
 937     const struct lacp_systemid *b)
 938 {
 939         return (memcmp(a, b, sizeof(*a)));
 940 }
 941
 942 #if 0   /* unused */
 943 static int
 944 lacp_compare_portid(const struct lacp_portid *a,
 945     const struct lacp_portid *b)
 946 {
 947         return (memcmp(a, b, sizeof(*a)));
 948 }
 949 #endif
 950
 951 static uint64_t
 952 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 953 {
 954         struct lacp_port *lp;
 955         uint64_t speed;
 956
 957         lp = TAILQ_FIRST(&la->la_ports);
 958         if (lp == NULL) {
 959                 return (0);
 960         }
 961
 962         speed = ifmedia_baudrate(lp->lp_media);
 963         speed *= la->la_nports;
 964         if (speed == 0) {
 965                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 966                     lp->lp_media, la->la_nports));
 967         }
 968
 969         return (speed);
 970 }
 971
 972 /*
 973  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 974  * packets from lagg(4) interface.
 975  */
 976
 977 static void
 978 lacp_select_active_aggregator(struct lacp_softc *lsc)
 979 {
 980         struct lacp_aggregator *la;
 981         struct lacp_aggregator *best_la = NULL;
 982         uint64_t best_speed = 0;
 983         char buf[LACP_LAGIDSTR_MAX+1];
 984
 985         LACP_TRACE(NULL);
 986
 987         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 988                 uint64_t speed;
 989
 990                 if (la->la_nports == 0) {
 991                         continue;
 992                 }
 993
 994                 speed = lacp_aggregator_bandwidth(la);
 995                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 996                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 997                     speed, la->la_nports));
 998
 999                 /* This aggregator is chosen if
1000                  *      the partner has a better system priority
1001                  *  or, the total aggregated speed is higher
1002                  *  or, it is already the chosen aggregator
1003                  */
1004                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
1005                      LACP_SYS_PRI(best_la->la_partner)) ||
1006                     speed > best_speed ||
1007                     (speed == best_speed &&
1008                     la == lsc->lsc_active_aggregator)) {
1009                         best_la = la;
1010                         best_speed = speed;
1011                 }
1012         }
1013
1014         KASSERT(best_la == NULL || best_la->la_nports > 0,
1015             ("invalid aggregator refcnt"));
1016         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1017             ("invalid aggregator list"));
1018
1019         if (lsc->lsc_active_aggregator != best_la) {
1020                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1021                 LACP_DPRINTF((NULL, "old %s\n",
1022                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1023                     buf, sizeof(buf))));
1024         } else {
1025                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1026         }
1027         LACP_DPRINTF((NULL, "new %s\n",
1028             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1029
1030         if (lsc->lsc_active_aggregator != best_la) {
1031                 lsc->lsc_active_aggregator = best_la;
1032                 lacp_update_portmap(lsc);
1033                 if (best_la) {
1034                         lacp_suppress_distributing(lsc, best_la);
1035                 }
1036         }
1037 }
1038
1039 /*
1040  * Updated the inactive portmap array with the new list of ports and
1041  * make it live.
1042  */
1043 static void
1044 lacp_update_portmap(struct lacp_softc *lsc)
1045 {
1046         struct lagg_softc *sc = lsc->lsc_softc;
1047         struct lacp_aggregator *la;
1048         struct lacp_portmap *p;
1049         struct lacp_port *lp;
1050         uint64_t speed;
1051         u_int newmap;
1052         int i;
1053
1054         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1055         p = &lsc->lsc_pmap[newmap];
1056         la = lsc->lsc_active_aggregator;
1057         speed = 0;
1058         bzero(p, sizeof(struct lacp_portmap));
1059
1060         if (la != NULL && la->la_nports > 0) {
1061                 p->pm_count = la->la_nports;
1062                 i = 0;
1063                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1064                         p->pm_map[i++] = lp;
1065                 KASSERT(i == p->pm_count, ("Invalid port count"));
1066                 speed = lacp_aggregator_bandwidth(la);
1067         }
1068         sc->sc_ifp->if_baudrate = speed;
1069
1070         /* switch the active portmap over */
1071         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1072         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1073                     lsc->lsc_activemap,
1074                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1075 }
1076
1077 static uint16_t
1078 lacp_compose_key(struct lacp_port *lp)
1079 {
1080         struct lagg_port *lgp = lp->lp_lagg;
1081         struct lagg_softc *sc = lgp->lp_softc;
1082         u_int media = lp->lp_media;
1083         uint16_t key;
1084
1085         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1086
1087                 /*
1088                  * non-aggregatable links should have unique keys.
1089                  *
1090                  * XXX this isn't really unique as if_index is 16 bit.
1091                  */
1092
1093                 /* bit 0..14:   (some bits of) if_index of this port */
1094                 key = lp->lp_ifp->if_index;
1095                 /* bit 15:      1 */
1096                 key |= 0x8000;
1097         } else {
1098                 u_int subtype = IFM_SUBTYPE(media);
1099
1100                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1101                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1102
1103                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1104                 switch (subtype) {
1105                 case IFM_10_T:
1106                 case IFM_10_2:
1107                 case IFM_10_5:
1108                 case IFM_10_STP:
1109                 case IFM_10_FL:
1110                         key = IFM_10_T;
1111                         break;
1112                 case IFM_100_TX:
1113                 case IFM_100_FX:
1114                 case IFM_100_T4:
1115                 case IFM_100_VG:
1116                 case IFM_100_T2:
1117                 case IFM_100_T:
1118                         key = IFM_100_TX;
1119                         break;
1120                 case IFM_1000_SX:
1121                 case IFM_1000_LX:
1122                 case IFM_1000_CX:
1123                 case IFM_1000_T:
1124                 case IFM_1000_KX:
1125                 case IFM_1000_SGMII:
1126                 case IFM_1000_CX_SGMII:
1127                         key = IFM_1000_SX;
1128                         break;
1129                 case IFM_10G_LR:
1130                 case IFM_10G_SR:
1131                 case IFM_10G_CX4:
1132                 case IFM_10G_TWINAX:
1133                 case IFM_10G_TWINAX_LONG:
1134                 case IFM_10G_LRM:
1135                 case IFM_10G_T:
1136                 case IFM_10G_KX4:
1137                 case IFM_10G_KR:
1138                 case IFM_10G_CR1:
1139                 case IFM_10G_ER:
1140                 case IFM_10G_SFI:
1141                         key = IFM_10G_LR;
1142                         break;
1143                 case IFM_20G_KR2:
1144                         key = IFM_20G_KR2;
1145                         break;
1146                 case IFM_2500_KX:
1147                 case IFM_2500_T:
1148                         key = IFM_2500_KX;
1149                         break;
1150                 case IFM_5000_T:
1151                         key = IFM_5000_T;
1152                         break;
1153                 case IFM_50G_PCIE:
1154                 case IFM_50G_CR2:
1155                 case IFM_50G_KR2:
1156                         key = IFM_50G_PCIE;
1157                         break;
1158                 case IFM_56G_R4:
1159                         key = IFM_56G_R4;
1160                         break;
1161                 case IFM_25G_PCIE:
1162                 case IFM_25G_CR:
1163                 case IFM_25G_KR:
1164                 case IFM_25G_SR:
1165                         key = IFM_25G_PCIE;
1166                         break;
1167                 case IFM_40G_CR4:
1168                 case IFM_40G_SR4:
1169                 case IFM_40G_LR4:
1170                 case IFM_40G_XLPPI:
1171                 case IFM_40G_KR4:
1172                         key = IFM_40G_CR4;
1173                         break;
1174                 case IFM_100G_CR4:
1175                 case IFM_100G_SR4:
1176                 case IFM_100G_KR4:
1177                 case IFM_100G_LR4:
1178                         key = IFM_100G_CR4;
1179                         break;
1180                 default:
1181                         key = subtype;
1182                         break;
1183                 }
1184                 /* bit 5..14:   (some bits of) if_index of lagg device */
1185                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1186                 /* bit 15:      0 */
1187         }
1188         return (htons(key));
1189 }
1190
1191 static void
1192 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1193 {
1194         char buf[LACP_LAGIDSTR_MAX+1];
1195
1196         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1197             __func__,
1198             lacp_format_lagid(&la->la_actor, &la->la_partner,
1199             buf, sizeof(buf)),
1200             la->la_refcnt, la->la_refcnt + 1));
1201
1202         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1203         la->la_refcnt++;
1204         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1205 }
1206
1207 static void
1208 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1209 {
1210         char buf[LACP_LAGIDSTR_MAX+1];
1211
1212         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1213             __func__,
1214             lacp_format_lagid(&la->la_actor, &la->la_partner,
1215             buf, sizeof(buf)),
1216             la->la_refcnt, la->la_refcnt - 1));
1217
1218         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1219         la->la_refcnt--;
1220         if (la->la_refcnt > 0) {
1221                 return;
1222         }
1223
1224         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1225         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1226
1227         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1228
1229         free(la, M_DEVBUF);
1230 }
1231
1232 /*
1233  * lacp_aggregator_get: allocate an aggregator.
1234  */
1235
1236 static struct lacp_aggregator *
1237 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1238 {
1239         struct lacp_aggregator *la;
1240
1241         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1242         if (la) {
1243                 la->la_refcnt = 1;
1244                 la->la_nports = 0;
1245                 TAILQ_INIT(&la->la_ports);
1246                 la->la_pending = 0;
1247                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1248         }
1249
1250         return (la);
1251 }
1252
1253 /*
1254  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1255  */
1256
1257 static void
1258 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1259 {
1260         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1261         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1262
1263         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1264 }
1265
1266 static void
1267 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1268     const struct lacp_peerinfo *lpi_port)
1269 {
1270         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1271         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1272         lpi_aggr->lip_key = lpi_port->lip_key;
1273 }
1274
1275 /*
1276  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1277  */
1278
1279 static int
1280 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1281     const struct lacp_port *lp)
1282 {
1283         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1284             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1285                 return (0);
1286         }
1287
1288         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1289                 return (0);
1290         }
1291
1292         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1293                 return (0);
1294         }
1295
1296         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1297                 return (0);
1298         }
1299
1300         return (1);
1301 }
1302
1303 static int
1304 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1305     const struct lacp_peerinfo *b)
1306 {
1307         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1308             sizeof(a->lip_systemid))) {
1309                 return (0);
1310         }
1311
1312         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1313                 return (0);
1314         }
1315
1316         return (1);
1317 }
1318
1319 static void
1320 lacp_port_enable(struct lacp_port *lp)
1321 {
1322         lp->lp_state |= LACP_STATE_AGGREGATION;
1323 }
1324
1325 static void
1326 lacp_port_disable(struct lacp_port *lp)
1327 {
1328         lacp_set_mux(lp, LACP_MUX_DETACHED);
1329
1330         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1331         lp->lp_selected = LACP_UNSELECTED;
1332         lacp_sm_rx_record_default(lp);
1333         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1334         lp->lp_state &= ~LACP_STATE_EXPIRED;
1335 }
1336
1337 /*
1338  * lacp_select: select an aggregator.  create one if necessary.
1339  */
1340 static void
1341 lacp_select(struct lacp_port *lp)
1342 {
1343         struct lacp_softc *lsc = lp->lp_lsc;
1344         struct lacp_aggregator *la;
1345         char buf[LACP_LAGIDSTR_MAX+1];
1346
1347         if (lp->lp_aggregator) {
1348                 return;
1349         }
1350
1351         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1352             ("timer_wait_while still active"));
1353
1354         LACP_DPRINTF((lp, "port lagid=%s\n",
1355             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1356             buf, sizeof(buf))));
1357
1358         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1359                 if (lacp_aggregator_is_compatible(la, lp)) {
1360                         break;
1361                 }
1362         }
1363
1364         if (la == NULL) {
1365                 la = lacp_aggregator_get(lsc, lp);
1366                 if (la == NULL) {
1367                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1368
1369                         /*
1370                          * will retry on the next tick.
1371                          */
1372
1373                         return;
1374                 }
1375                 lacp_fill_aggregator_id(la, lp);
1376                 LACP_DPRINTF((lp, "aggregator created\n"));
1377         } else {
1378                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1379                 if (la->la_refcnt == LACP_MAX_PORTS)
1380                         return;
1381                 lacp_aggregator_addref(lsc, la);
1382         }
1383
1384         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1385             lacp_format_lagid(&la->la_actor, &la->la_partner,
1386             buf, sizeof(buf))));
1387
1388         lp->lp_aggregator = la;
1389         lp->lp_selected = LACP_SELECTED;
1390 }
1391
1392 /*
1393  * lacp_unselect: finish unselect/detach process.
1394  */
1395
1396 static void
1397 lacp_unselect(struct lacp_port *lp)
1398 {
1399         struct lacp_softc *lsc = lp->lp_lsc;
1400         struct lacp_aggregator *la = lp->lp_aggregator;
1401
1402         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1403             ("timer_wait_while still active"));
1404
1405         if (la == NULL) {
1406                 return;
1407         }
1408
1409         lp->lp_aggregator = NULL;
1410         lacp_aggregator_delref(lsc, la);
1411 }
1412
1413 /* mux machine */
1414
1415 static void
1416 lacp_sm_mux(struct lacp_port *lp)
1417 {
1418         struct lagg_port *lgp = lp->lp_lagg;
1419         struct lagg_softc *sc = lgp->lp_softc;
1420         enum lacp_mux_state new_state;
1421         boolean_t p_sync =
1422                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1423         boolean_t p_collecting =
1424             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1425         enum lacp_selected selected = lp->lp_selected;
1426         struct lacp_aggregator *la;
1427
1428         if (lacp_debug > 1)
1429                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1430                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1431                     lp->lp_mux_state, selected, p_sync, p_collecting);
1432
1433 re_eval:
1434         la = lp->lp_aggregator;
1435         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1436             ("MUX not detached"));
1437         new_state = lp->lp_mux_state;
1438         switch (lp->lp_mux_state) {
1439         case LACP_MUX_DETACHED:
1440                 if (selected != LACP_UNSELECTED) {
1441                         new_state = LACP_MUX_WAITING;
1442                 }
1443                 break;
1444         case LACP_MUX_WAITING:
1445                 KASSERT(la->la_pending > 0 ||
1446                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1447                     ("timer_wait_while still active"));
1448                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1449                         new_state = LACP_MUX_ATTACHED;
1450                 } else if (selected == LACP_UNSELECTED) {
1451                         new_state = LACP_MUX_DETACHED;
1452                 }
1453                 break;
1454         case LACP_MUX_ATTACHED:
1455                 if (selected == LACP_SELECTED && p_sync) {
1456                         new_state = LACP_MUX_COLLECTING;
1457                 } else if (selected != LACP_SELECTED) {
1458                         new_state = LACP_MUX_DETACHED;
1459                 }
1460                 break;
1461         case LACP_MUX_COLLECTING:
1462                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1463                         new_state = LACP_MUX_DISTRIBUTING;
1464                 } else if (selected != LACP_SELECTED || !p_sync) {
1465                         new_state = LACP_MUX_ATTACHED;
1466                 }
1467                 break;
1468         case LACP_MUX_DISTRIBUTING:
1469                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1470                         new_state = LACP_MUX_COLLECTING;
1471                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1472                         sc->sc_flapping++;
1473                 }
1474                 break;
1475         default:
1476                 panic("%s: unknown state", __func__);
1477         }
1478
1479         if (lp->lp_mux_state == new_state) {
1480                 return;
1481         }
1482
1483         lacp_set_mux(lp, new_state);
1484         goto re_eval;
1485 }
1486
1487 static void
1488 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1489 {
1490         struct lacp_aggregator *la = lp->lp_aggregator;
1491
1492         if (lp->lp_mux_state == new_state) {
1493                 return;
1494         }
1495
1496         switch (new_state) {
1497         case LACP_MUX_DETACHED:
1498                 lp->lp_state &= ~LACP_STATE_SYNC;
1499                 lacp_disable_distributing(lp);
1500                 lacp_disable_collecting(lp);
1501                 lacp_sm_assert_ntt(lp);
1502                 /* cancel timer */
1503                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1504                         KASSERT(la->la_pending > 0,
1505                             ("timer_wait_while not active"));
1506                         la->la_pending--;
1507                 }
1508                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1509                 lacp_unselect(lp);
1510                 break;
1511         case LACP_MUX_WAITING:
1512                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1513                     LACP_AGGREGATE_WAIT_TIME);
1514                 la->la_pending++;
1515                 break;
1516         case LACP_MUX_ATTACHED:
1517                 lp->lp_state |= LACP_STATE_SYNC;
1518                 lacp_disable_collecting(lp);
1519                 lacp_sm_assert_ntt(lp);
1520                 break;
1521         case LACP_MUX_COLLECTING:
1522                 lacp_enable_collecting(lp);
1523                 lacp_disable_distributing(lp);
1524                 lacp_sm_assert_ntt(lp);
1525                 break;
1526         case LACP_MUX_DISTRIBUTING:
1527                 lacp_enable_distributing(lp);
1528                 break;
1529         default:
1530                 panic("%s: unknown state", __func__);
1531         }
1532
1533         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1534
1535         lp->lp_mux_state = new_state;
1536 }
1537
1538 static void
1539 lacp_sm_mux_timer(struct lacp_port *lp)
1540 {
1541         struct lacp_aggregator *la = lp->lp_aggregator;
1542         char buf[LACP_LAGIDSTR_MAX+1];
1543
1544         KASSERT(la->la_pending > 0, ("no pending event"));
1545
1546         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1547             lacp_format_lagid(&la->la_actor, &la->la_partner,
1548             buf, sizeof(buf)),
1549             la->la_pending, la->la_pending - 1));
1550
1551         la->la_pending--;
1552 }
1553
1554 /* periodic transmit machine */
1555
1556 static void
1557 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1558 {
1559         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1560             LACP_STATE_TIMEOUT)) {
1561                 return;
1562         }
1563
1564         LACP_DPRINTF((lp, "partner timeout changed\n"));
1565
1566         /*
1567          * FAST_PERIODIC -> SLOW_PERIODIC
1568          * or
1569          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1570          *
1571          * let lacp_sm_ptx_tx_schedule to update timeout.
1572          */
1573
1574         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1575
1576         /*
1577          * if timeout has been shortened, assert NTT.
1578          */
1579
1580         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1581                 lacp_sm_assert_ntt(lp);
1582         }
1583 }
1584
1585 static void
1586 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1587 {
1588         int timeout;
1589
1590         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1591             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1592
1593                 /*
1594                  * NO_PERIODIC
1595                  */
1596
1597                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1598                 return;
1599         }
1600
1601         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1602                 return;
1603         }
1604
1605         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1606             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1607
1608         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1609 }
1610
1611 static void
1612 lacp_sm_ptx_timer(struct lacp_port *lp)
1613 {
1614         lacp_sm_assert_ntt(lp);
1615 }
1616
1617 static void
1618 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1619 {
1620         int timeout;
1621
1622         /*
1623          * check LACP_DISABLED first
1624          */
1625
1626         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1627                 return;
1628         }
1629
1630         /*
1631          * check loopback condition.
1632          */
1633
1634         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1635             &lp->lp_actor.lip_systemid)) {
1636                 return;
1637         }
1638
1639         /*
1640          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1641          */
1642
1643         lacp_sm_rx_update_selected(lp, du);
1644         lacp_sm_rx_update_ntt(lp, du);
1645         lacp_sm_rx_record_pdu(lp, du);
1646
1647         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1648             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1649         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1650
1651         lp->lp_state &= ~LACP_STATE_EXPIRED;
1652
1653         /*
1654          * kick transmit machine without waiting the next tick.
1655          */
1656
1657         lacp_sm_tx(lp);
1658 }
1659
1660 static void
1661 lacp_sm_rx_set_expired(struct lacp_port *lp)
1662 {
1663         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1664         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1665         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1666         lp->lp_state |= LACP_STATE_EXPIRED;
1667 }
1668
1669 static void
1670 lacp_sm_rx_timer(struct lacp_port *lp)
1671 {
1672         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1673                 /* CURRENT -> EXPIRED */
1674                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1675                 lacp_sm_rx_set_expired(lp);
1676         } else {
1677                 /* EXPIRED -> DEFAULTED */
1678                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1679                 lacp_sm_rx_update_default_selected(lp);
1680                 lacp_sm_rx_record_default(lp);
1681                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1682         }
1683 }
1684
1685 static void
1686 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1687 {
1688         boolean_t active;
1689         uint8_t oldpstate;
1690         char buf[LACP_STATESTR_MAX+1];
1691
1692         LACP_TRACE(lp);
1693
1694         oldpstate = lp->lp_partner.lip_state;
1695
1696         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1697             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1698             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1699
1700         lp->lp_partner = du->ldu_actor;
1701         if (active &&
1702             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1703             LACP_STATE_AGGREGATION) &&
1704             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1705             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1706                 /* XXX nothing? */
1707         } else {
1708                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1709         }
1710
1711         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1712
1713         if (oldpstate != lp->lp_partner.lip_state) {
1714                 LACP_DPRINTF((lp, "old pstate %s\n",
1715                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1716                 LACP_DPRINTF((lp, "new pstate %s\n",
1717                     lacp_format_state(lp->lp_partner.lip_state, buf,
1718                     sizeof(buf))));
1719         }
1720
1721         /* XXX Hack, still need to implement 5.4.9 para 2,3,4 */
1722         if (lp->lp_lsc->lsc_strict_mode)
1723                 lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1724
1725         lacp_sm_ptx_update_timeout(lp, oldpstate);
1726 }
1727
1728 static void
1729 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1730 {
1731
1732         LACP_TRACE(lp);
1733
1734         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1735             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1736             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1737                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1738                 lacp_sm_assert_ntt(lp);
1739         }
1740 }
1741
1742 static void
1743 lacp_sm_rx_record_default(struct lacp_port *lp)
1744 {
1745         uint8_t oldpstate;
1746
1747         LACP_TRACE(lp);
1748
1749         oldpstate = lp->lp_partner.lip_state;
1750         if (lp->lp_lsc->lsc_strict_mode)
1751                 lp->lp_partner = lacp_partner_admin_strict;
1752         else
1753                 lp->lp_partner = lacp_partner_admin_optimistic;;
1754         lp->lp_state |= LACP_STATE_DEFAULTED;
1755         lacp_sm_ptx_update_timeout(lp, oldpstate);
1756 }
1757
1758 static void
1759 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1760     const struct lacp_peerinfo *info)
1761 {
1762
1763         LACP_TRACE(lp);
1764
1765         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1766             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1767             LACP_STATE_AGGREGATION)) {
1768                 lp->lp_selected = LACP_UNSELECTED;
1769                 /* mux machine will clean up lp->lp_aggregator */
1770         }
1771 }
1772
1773 static void
1774 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1775 {
1776
1777         LACP_TRACE(lp);
1778
1779         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1780 }
1781
1782 static void
1783 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1784 {
1785
1786         LACP_TRACE(lp);
1787
1788         if (lp->lp_lsc->lsc_strict_mode)
1789                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1790                     &lacp_partner_admin_strict);
1791         else
1792                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1793                     &lacp_partner_admin_optimistic);
1794 }
1795
1796 /* transmit machine */
1797
1798 static void
1799 lacp_sm_tx(struct lacp_port *lp)
1800 {
1801         int error = 0;
1802
1803         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1804 #if 1
1805             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1806             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1807 #endif
1808             ) {
1809                 lp->lp_flags &= ~LACP_PORT_NTT;
1810         }
1811
1812         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1813                 return;
1814         }
1815
1816         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1817         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1818                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1819                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1820                 return;
1821         }
1822
1823         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1824                 error = lacp_xmit_lacpdu(lp);
1825         } else {
1826                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1827         }
1828
1829         if (error == 0) {
1830                 lp->lp_flags &= ~LACP_PORT_NTT;
1831         } else {
1832                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1833                     error));
1834         }
1835 }
1836
1837 static void
1838 lacp_sm_assert_ntt(struct lacp_port *lp)
1839 {
1840
1841         lp->lp_flags |= LACP_PORT_NTT;
1842 }
1843
1844 static void
1845 lacp_run_timers(struct lacp_port *lp)
1846 {
1847         int i;
1848
1849         for (i = 0; i < LACP_NTIMER; i++) {
1850                 KASSERT(lp->lp_timer[i] >= 0,
1851                     ("invalid timer value %d", lp->lp_timer[i]));
1852                 if (lp->lp_timer[i] == 0) {
1853                         continue;
1854                 } else if (--lp->lp_timer[i] <= 0) {
1855                         if (lacp_timer_funcs[i]) {
1856                                 (*lacp_timer_funcs[i])(lp);
1857                         }
1858                 }
1859         }
1860 }
1861
1862 int
1863 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1864 {
1865         struct lacp_softc *lsc = lp->lp_lsc;
1866         struct lagg_port *lgp = lp->lp_lagg;
1867         struct lacp_port *lp2;
1868         struct markerdu *mdu;
1869         int error = 0;
1870         int pending = 0;
1871
1872         if (m->m_pkthdr.len != sizeof(*mdu)) {
1873                 goto bad;
1874         }
1875
1876         if ((m->m_flags & M_MCAST) == 0) {
1877                 goto bad;
1878         }
1879
1880         if (m->m_len < sizeof(*mdu)) {
1881                 m = m_pullup(m, sizeof(*mdu));
1882                 if (m == NULL) {
1883                         return (ENOMEM);
1884                 }
1885         }
1886
1887         mdu = mtod(m, struct markerdu *);
1888
1889         if (memcmp(&mdu->mdu_eh.ether_dhost,
1890             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1891                 goto bad;
1892         }
1893
1894         if (mdu->mdu_sph.sph_version != 1) {
1895                 goto bad;
1896         }
1897
1898         switch (mdu->mdu_tlv.tlv_type) {
1899         case MARKER_TYPE_INFO:
1900                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1901                     marker_info_tlv_template, TRUE)) {
1902                         goto bad;
1903                 }
1904                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1905                 memcpy(&mdu->mdu_eh.ether_dhost,
1906                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1907                 memcpy(&mdu->mdu_eh.ether_shost,
1908                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1909                 error = lagg_enqueue(lp->lp_ifp, m);
1910                 break;
1911
1912         case MARKER_TYPE_RESPONSE:
1913                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1914                     marker_response_tlv_template, TRUE)) {
1915                         goto bad;
1916                 }
1917                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1918                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1919                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1920
1921                 /* Verify that it is the last marker we sent out */
1922                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1923                     sizeof(struct lacp_markerinfo)))
1924                         goto bad;
1925
1926                 LACP_LOCK(lsc);
1927                 lp->lp_flags &= ~LACP_PORT_MARK;
1928
1929                 if (lsc->lsc_suppress_distributing) {
1930                         /* Check if any ports are waiting for a response */
1931                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1932                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1933                                         pending = 1;
1934                                         break;
1935                                 }
1936                         }
1937
1938                         if (pending == 0) {
1939                                 /* All interface queues are clear */
1940                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1941                                 lsc->lsc_suppress_distributing = FALSE;
1942                         }
1943                 }
1944                 LACP_UNLOCK(lsc);
1945                 m_freem(m);
1946                 break;
1947
1948         default:
1949                 goto bad;
1950         }
1951
1952         return (error);
1953
1954 bad:
1955         LACP_DPRINTF((lp, "bad marker frame\n"));
1956         m_freem(m);
1957         return (EINVAL);
1958 }
1959
1960 static int
1961 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1962     const struct tlv_template *tmpl, boolean_t check_type)
1963 {
1964         while (/* CONSTCOND */ 1) {
1965                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1966                         return (EINVAL);
1967                 }
1968                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1969                     tlv->tlv_length != tmpl->tmpl_length) {
1970                         return (EINVAL);
1971                 }
1972                 if (tmpl->tmpl_type == 0) {
1973                         break;
1974                 }
1975                 tlv = (const struct tlvhdr *)
1976                     ((const char *)tlv + tlv->tlv_length);
1977                 tmpl++;
1978         }
1979
1980         return (0);
1981 }
1982
1983 /* Debugging */
1984 const char *
1985 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1986 {
1987         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1988             (int)mac[0],
1989             (int)mac[1],
1990             (int)mac[2],
1991             (int)mac[3],
1992             (int)mac[4],
1993             (int)mac[5]);
1994
1995         return (buf);
1996 }
1997
1998 const char *
1999 lacp_format_systemid(const struct lacp_systemid *sysid,
2000     char *buf, size_t buflen)
2001 {
2002         char macbuf[LACP_MACSTR_MAX+1];
2003
2004         snprintf(buf, buflen, "%04X,%s",
2005             ntohs(sysid->lsi_prio),
2006             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2007
2008         return (buf);
2009 }
2010
2011 const char *
2012 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2013 {
2014         snprintf(buf, buflen, "%04X,%04X",
2015             ntohs(portid->lpi_prio),
2016             ntohs(portid->lpi_portno));
2017
2018         return (buf);
2019 }
2020
2021 const char *
2022 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2023 {
2024         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2025         char portid[LACP_PORTIDSTR_MAX+1];
2026
2027         snprintf(buf, buflen, "(%s,%04X,%s)",
2028             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2029             ntohs(peer->lip_key),
2030             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2031
2032         return (buf);
2033 }
2034
2035 const char *
2036 lacp_format_lagid(const struct lacp_peerinfo *a,
2037     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2038 {
2039         char astr[LACP_PARTNERSTR_MAX+1];
2040         char bstr[LACP_PARTNERSTR_MAX+1];
2041
2042 #if 0
2043         /*
2044          * there's a convention to display small numbered peer
2045          * in the left.
2046          */
2047
2048         if (lacp_compare_peerinfo(a, b) > 0) {
2049                 const struct lacp_peerinfo *t;
2050
2051                 t = a;
2052                 a = b;
2053                 b = t;
2054         }
2055 #endif
2056
2057         snprintf(buf, buflen, "[%s,%s]",
2058             lacp_format_partner(a, astr, sizeof(astr)),
2059             lacp_format_partner(b, bstr, sizeof(bstr)));
2060
2061         return (buf);
2062 }
2063
2064 const char *
2065 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2066     char *buf, size_t buflen)
2067 {
2068         if (la == NULL) {
2069                 return ("(none)");
2070         }
2071
2072         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2073 }
2074
2075 const char *
2076 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2077 {
2078         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2079         return (buf);
2080 }
2081
2082 static void
2083 lacp_dump_lacpdu(const struct lacpdu *du)
2084 {
2085         char buf[LACP_PARTNERSTR_MAX+1];
2086         char buf2[LACP_STATESTR_MAX+1];
2087
2088         printf("actor=%s\n",
2089             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2090         printf("actor.state=%s\n",
2091             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2092         printf("partner=%s\n",
2093             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2094         printf("partner.state=%s\n",
2095             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2096
2097         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2098 }
2099
2100 static void
2101 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2102 {
2103         va_list va;
2104
2105         if (lp) {
2106                 printf("%s: ", lp->lp_ifp->if_xname);
2107         }
2108
2109         va_start(va, fmt);
2110         vprintf(fmt, va);
2111         va_end(va);
2112 }