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 VNET_DEFINE(int, lacp_debug);
 191 #define V_lacp_debug    VNET(lacp_debug)
 192 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
 193 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 194     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 195
 196 static VNET_DEFINE(int, lacp_default_strict_mode) = 1;
 197 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode, CTLFLAG_RWTUN,
 198     &VNET_NAME(lacp_default_strict_mode), 0,
 199     "LACP strict protocol compliance default");
 200
 201 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 202 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 203 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 204
 205 /*
 206  * partner administration variables.
 207  * XXX should be configurable.
 208  */
 209
 210 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 211         .lip_systemid = { .lsi_prio = 0xffff },
 212         .lip_portid = { .lpi_prio = 0xffff },
 213         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 214             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 215 };
 216
 217 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 218         .lip_systemid = { .lsi_prio = 0xffff },
 219         .lip_portid = { .lpi_prio = 0xffff },
 220         .lip_state = 0,
 221 };
 222
 223 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 224         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 225         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 226         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 227 };
 228
 229 struct mbuf *
 230 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 231 {
 232         struct lacp_port *lp = LACP_PORT(lgp);
 233         uint8_t subtype;
 234
 235         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 236                 m_freem(m);
 237                 return (NULL);
 238         }
 239
 240         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 241         switch (subtype) {
 242                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 243                         lacp_pdu_input(lp, m);
 244                         return (NULL);
 245
 246                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 247                         lacp_marker_input(lp, m);
 248                         return (NULL);
 249         }
 250
 251         /* Not a subtype we are interested in */
 252         return (m);
 253 }
 254
 255 /*
 256  * lacp_pdu_input: process lacpdu
 257  */
 258 static int
 259 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 260 {
 261         struct lacp_softc *lsc = lp->lp_lsc;
 262         struct lacpdu *du;
 263         int error = 0;
 264
 265         if (m->m_pkthdr.len != sizeof(*du)) {
 266                 goto bad;
 267         }
 268
 269         if ((m->m_flags & M_MCAST) == 0) {
 270                 goto bad;
 271         }
 272
 273         if (m->m_len < sizeof(*du)) {
 274                 m = m_pullup(m, sizeof(*du));
 275                 if (m == NULL) {
 276                         return (ENOMEM);
 277                 }
 278         }
 279
 280         du = mtod(m, struct lacpdu *);
 281
 282         if (memcmp(&du->ldu_eh.ether_dhost,
 283             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 284                 goto bad;
 285         }
 286
 287         /*
 288          * ignore the version for compatibility with
 289          * the future protocol revisions.
 290          */
 291 #if 0
 292         if (du->ldu_sph.sph_version != 1) {
 293                 goto bad;
 294         }
 295 #endif
 296
 297         /*
 298          * ignore tlv types for compatibility with
 299          * the future protocol revisions.
 300          */
 301         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 302             lacp_info_tlv_template, FALSE)) {
 303                 goto bad;
 304         }
 305
 306         if (V_lacp_debug > 0) {
 307                 lacp_dprintf(lp, "lacpdu receive\n");
 308                 lacp_dump_lacpdu(du);
 309         }
 310
 311         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 312                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 313                 goto bad;
 314         }
 315
 316         LACP_LOCK(lsc);
 317         lacp_sm_rx(lp, du);
 318         LACP_UNLOCK(lsc);
 319
 320         m_freem(m);
 321         return (error);
 322
 323 bad:
 324         m_freem(m);
 325         return (EINVAL);
 326 }
 327
 328 static void
 329 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 330 {
 331         struct lagg_port *lgp = lp->lp_lagg;
 332         struct lagg_softc *sc = lgp->lp_softc;
 333
 334         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 335         memcpy(&info->lip_systemid.lsi_mac,
 336             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 337         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 338         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 339         info->lip_state = lp->lp_state;
 340 }
 341
 342 static void
 343 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 344 {
 345         struct ifnet *ifp = lp->lp_ifp;
 346
 347         /* Fill in the port index and system id (encoded as the MAC) */
 348         info->mi_rq_port = htons(ifp->if_index);
 349         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 350         info->mi_rq_xid = htonl(0);
 351 }
 352
 353 static int
 354 lacp_xmit_lacpdu(struct lacp_port *lp)
 355 {
 356         struct lagg_port *lgp = lp->lp_lagg;
 357         struct mbuf *m;
 358         struct lacpdu *du;
 359         int error;
 360
 361         LACP_LOCK_ASSERT(lp->lp_lsc);
 362
 363         m = m_gethdr(M_NOWAIT, MT_DATA);
 364         if (m == NULL) {
 365                 return (ENOMEM);
 366         }
 367         m->m_len = m->m_pkthdr.len = sizeof(*du);
 368
 369         du = mtod(m, struct lacpdu *);
 370         memset(du, 0, sizeof(*du));
 371
 372         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 373             ETHER_ADDR_LEN);
 374         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 375         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 376
 377         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 378         du->ldu_sph.sph_version = 1;
 379
 380         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 381         du->ldu_actor = lp->lp_actor;
 382
 383         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 384             sizeof(du->ldu_partner));
 385         du->ldu_partner = lp->lp_partner;
 386
 387         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 388             sizeof(du->ldu_collector));
 389         du->ldu_collector.lci_maxdelay = 0;
 390
 391         if (V_lacp_debug > 0) {
 392                 lacp_dprintf(lp, "lacpdu transmit\n");
 393                 lacp_dump_lacpdu(du);
 394         }
 395
 396         m->m_flags |= M_MCAST;
 397
 398         /*
 399          * XXX should use higher priority queue.
 400          * otherwise network congestion can break aggregation.
 401          */
 402
 403         error = lagg_enqueue(lp->lp_ifp, m);
 404         return (error);
 405 }
 406
 407 static int
 408 lacp_xmit_marker(struct lacp_port *lp)
 409 {
 410         struct lagg_port *lgp = lp->lp_lagg;
 411         struct mbuf *m;
 412         struct markerdu *mdu;
 413         int error;
 414
 415         LACP_LOCK_ASSERT(lp->lp_lsc);
 416
 417         m = m_gethdr(M_NOWAIT, MT_DATA);
 418         if (m == NULL) {
 419                 return (ENOMEM);
 420         }
 421         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 422
 423         mdu = mtod(m, struct markerdu *);
 424         memset(mdu, 0, sizeof(*mdu));
 425
 426         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 427             ETHER_ADDR_LEN);
 428         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 429         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 430
 431         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 432         mdu->mdu_sph.sph_version = 1;
 433
 434         /* Bump the transaction id and copy over the marker info */
 435         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 436         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 437         mdu->mdu_info = lp->lp_marker;
 438
 439         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 440             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 441             ntohl(mdu->mdu_info.mi_rq_xid)));
 442
 443         m->m_flags |= M_MCAST;
 444         error = lagg_enqueue(lp->lp_ifp, m);
 445         return (error);
 446 }
 447
 448 void
 449 lacp_linkstate(struct lagg_port *lgp)
 450 {
 451         struct lacp_port *lp = LACP_PORT(lgp);
 452         struct lacp_softc *lsc = lp->lp_lsc;
 453         struct ifnet *ifp = lgp->lp_ifp;
 454         struct ifmediareq ifmr;
 455         int error = 0;
 456         u_int media;
 457         uint8_t old_state;
 458         uint16_t old_key;
 459
 460         bzero((char *)&ifmr, sizeof(ifmr));
 461         error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 462         if (error != 0)
 463                 return;
 464
 465         LACP_LOCK(lsc);
 466         media = ifmr.ifm_active;
 467         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 468             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 469             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 470         old_state = lp->lp_state;
 471         old_key = lp->lp_key;
 472
 473         lp->lp_media = media;
 474         /*
 475          * If the port is not an active full duplex Ethernet link then it can
 476          * not be aggregated.
 477          */
 478         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 479             ifp->if_link_state != LINK_STATE_UP) {
 480                 lacp_port_disable(lp);
 481         } else {
 482                 lacp_port_enable(lp);
 483         }
 484         lp->lp_key = lacp_compose_key(lp);
 485
 486         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 487                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 488                 lp->lp_selected = LACP_UNSELECTED;
 489         }
 490         LACP_UNLOCK(lsc);
 491 }
 492
 493 static void
 494 lacp_tick(void *arg)
 495 {
 496         struct lacp_softc *lsc = arg;
 497         struct lacp_port *lp;
 498
 499         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 500                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 501                         continue;
 502
 503                 CURVNET_SET(lp->lp_ifp->if_vnet);
 504                 lacp_run_timers(lp);
 505
 506                 lacp_select(lp);
 507                 lacp_sm_mux(lp);
 508                 lacp_sm_tx(lp);
 509                 lacp_sm_ptx_tx_schedule(lp);
 510                 CURVNET_RESTORE();
 511         }
 512         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 513 }
 514
 515 int
 516 lacp_port_create(struct lagg_port *lgp)
 517 {
 518         struct lagg_softc *sc = lgp->lp_softc;
 519         struct lacp_softc *lsc = LACP_SOFTC(sc);
 520         struct lacp_port *lp;
 521         struct ifnet *ifp = lgp->lp_ifp;
 522         struct sockaddr_dl sdl;
 523         struct ifmultiaddr *rifma = NULL;
 524         int error;
 525
 526         bzero((char *)&sdl, sizeof(sdl));
 527         sdl.sdl_len = sizeof(sdl);
 528         sdl.sdl_family = AF_LINK;
 529         sdl.sdl_index = ifp->if_index;
 530         sdl.sdl_type = IFT_ETHER;
 531         sdl.sdl_alen = ETHER_ADDR_LEN;
 532
 533         bcopy(&ethermulticastaddr_slowprotocols,
 534             LLADDR(&sdl), ETHER_ADDR_LEN);
 535         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 536         if (error) {
 537                 printf("%s: ADDMULTI failed on %s\n", __func__, lgp->lp_ifname);
 538                 return (error);
 539         }
 540
 541         lp = malloc(sizeof(struct lacp_port),
 542             M_DEVBUF, M_NOWAIT|M_ZERO);
 543         if (lp == NULL)
 544                 return (ENOMEM);
 545
 546         LACP_LOCK(lsc);
 547         lgp->lp_psc = (caddr_t)lp;
 548         lp->lp_ifp = ifp;
 549         lp->lp_lagg = lgp;
 550         lp->lp_lsc = lsc;
 551         lp->lp_ifma = rifma;
 552
 553         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 554
 555         lacp_fill_actorinfo(lp, &lp->lp_actor);
 556         lacp_fill_markerinfo(lp, &lp->lp_marker);
 557         lp->lp_state = LACP_STATE_ACTIVITY;
 558         lp->lp_aggregator = NULL;
 559         lacp_sm_rx_set_expired(lp);
 560         LACP_UNLOCK(lsc);
 561         lacp_linkstate(lgp);
 562
 563         return (0);
 564 }
 565
 566 void
 567 lacp_port_destroy(struct lagg_port *lgp)
 568 {
 569         struct lacp_port *lp = LACP_PORT(lgp);
 570         struct lacp_softc *lsc = lp->lp_lsc;
 571         int i;
 572
 573         LACP_LOCK(lsc);
 574         for (i = 0; i < LACP_NTIMER; i++) {
 575                 LACP_TIMER_DISARM(lp, i);
 576         }
 577
 578         lacp_disable_collecting(lp);
 579         lacp_disable_distributing(lp);
 580         lacp_unselect(lp);
 581
 582         LIST_REMOVE(lp, lp_next);
 583         LACP_UNLOCK(lsc);
 584
 585         /* The address may have already been removed by if_purgemaddrs() */
 586         if (!lgp->lp_detaching)
 587                 if_delmulti_ifma(lp->lp_ifma);
 588
 589         free(lp, M_DEVBUF);
 590 }
 591
 592 void
 593 lacp_req(struct lagg_softc *sc, caddr_t data)
 594 {
 595         struct lacp_opreq *req = (struct lacp_opreq *)data;
 596         struct lacp_softc *lsc = LACP_SOFTC(sc);
 597         struct lacp_aggregator *la;
 598
 599         bzero(req, sizeof(struct lacp_opreq));
 600
 601         /*
 602          * If the LACP softc is NULL, return with the opreq structure full of
 603          * zeros.  It is normal for the softc to be NULL while the lagg is
 604          * being destroyed.
 605          */
 606         if (NULL == lsc)
 607                 return;
 608
 609         la = lsc->lsc_active_aggregator;
 610         LACP_LOCK(lsc);
 611         if (la != NULL) {
 612                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 613                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 614                     ETHER_ADDR_LEN);
 615                 req->actor_key = ntohs(la->la_actor.lip_key);
 616                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 617                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 618                 req->actor_state = la->la_actor.lip_state;
 619
 620                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 621                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 622                     ETHER_ADDR_LEN);
 623                 req->partner_key = ntohs(la->la_partner.lip_key);
 624                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 625                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 626                 req->partner_state = la->la_partner.lip_state;
 627         }
 628         LACP_UNLOCK(lsc);
 629 }
 630
 631 void
 632 lacp_portreq(struct lagg_port *lgp, caddr_t data)
 633 {
 634         struct lacp_opreq *req = (struct lacp_opreq *)data;
 635         struct lacp_port *lp = LACP_PORT(lgp);
 636         struct lacp_softc *lsc = lp->lp_lsc;
 637
 638         LACP_LOCK(lsc);
 639         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 640         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 641             ETHER_ADDR_LEN);
 642         req->actor_key = ntohs(lp->lp_actor.lip_key);
 643         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 644         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 645         req->actor_state = lp->lp_actor.lip_state;
 646
 647         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 648         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 649             ETHER_ADDR_LEN);
 650         req->partner_key = ntohs(lp->lp_partner.lip_key);
 651         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 652         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 653         req->partner_state = lp->lp_partner.lip_state;
 654         LACP_UNLOCK(lsc);
 655 }
 656
 657 static void
 658 lacp_disable_collecting(struct lacp_port *lp)
 659 {
 660         LACP_DPRINTF((lp, "collecting disabled\n"));
 661         lp->lp_state &= ~LACP_STATE_COLLECTING;
 662 }
 663
 664 static void
 665 lacp_enable_collecting(struct lacp_port *lp)
 666 {
 667         LACP_DPRINTF((lp, "collecting enabled\n"));
 668         lp->lp_state |= LACP_STATE_COLLECTING;
 669 }
 670
 671 static void
 672 lacp_disable_distributing(struct lacp_port *lp)
 673 {
 674         struct lacp_aggregator *la = lp->lp_aggregator;
 675         struct lacp_softc *lsc = lp->lp_lsc;
 676         struct lagg_softc *sc = lsc->lsc_softc;
 677         char buf[LACP_LAGIDSTR_MAX+1];
 678
 679         LACP_LOCK_ASSERT(lsc);
 680
 681         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 682                 return;
 683         }
 684
 685         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 686         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 687         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 688
 689         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 690             "nports %d -> %d\n",
 691             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 692             la->la_nports, la->la_nports - 1));
 693
 694         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 695         la->la_nports--;
 696         sc->sc_active = la->la_nports;
 697
 698         if (lsc->lsc_active_aggregator == la) {
 699                 lacp_suppress_distributing(lsc, la);
 700                 lacp_select_active_aggregator(lsc);
 701                 /* regenerate the port map, the active aggregator has changed */
 702                 lacp_update_portmap(lsc);
 703         }
 704
 705         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 706 }
 707
 708 static void
 709 lacp_enable_distributing(struct lacp_port *lp)
 710 {
 711         struct lacp_aggregator *la = lp->lp_aggregator;
 712         struct lacp_softc *lsc = lp->lp_lsc;
 713         struct lagg_softc *sc = lsc->lsc_softc;
 714         char buf[LACP_LAGIDSTR_MAX+1];
 715
 716         LACP_LOCK_ASSERT(lsc);
 717
 718         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 719                 return;
 720         }
 721
 722         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 723             "nports %d -> %d\n",
 724             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 725             la->la_nports, la->la_nports + 1));
 726
 727         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 728         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 729         la->la_nports++;
 730         sc->sc_active = la->la_nports;
 731
 732         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 733
 734         if (lsc->lsc_active_aggregator == la) {
 735                 lacp_suppress_distributing(lsc, la);
 736                 lacp_update_portmap(lsc);
 737         } else
 738                 /* try to become the active aggregator */
 739                 lacp_select_active_aggregator(lsc);
 740 }
 741
 742 static void
 743 lacp_transit_expire(void *vp)
 744 {
 745         struct lacp_softc *lsc = vp;
 746
 747         LACP_LOCK_ASSERT(lsc);
 748
 749         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 750         LACP_TRACE(NULL);
 751         CURVNET_RESTORE();
 752
 753         lsc->lsc_suppress_distributing = FALSE;
 754 }
 755
 756 void
 757 lacp_attach(struct lagg_softc *sc)
 758 {
 759         struct lacp_softc *lsc;
 760
 761         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 762
 763         sc->sc_psc = (caddr_t)lsc;
 764         lsc->lsc_softc = sc;
 765
 766         lsc->lsc_hashkey = arc4random();
 767         lsc->lsc_active_aggregator = NULL;
 768         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 769         LACP_LOCK_INIT(lsc);
 770         TAILQ_INIT(&lsc->lsc_aggregators);
 771         LIST_INIT(&lsc->lsc_ports);
 772
 773         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 774         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 775
 776         /* if the lagg is already up then do the same */
 777         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 778                 lacp_init(sc);
 779 }
 780
 781 int
 782 lacp_detach(void *psc)
 783 {
 784         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 785
 786         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 787             ("aggregators still active"));
 788         KASSERT(lsc->lsc_active_aggregator == NULL,
 789             ("aggregator still attached"));
 790
 791         callout_drain(&lsc->lsc_transit_callout);
 792         callout_drain(&lsc->lsc_callout);
 793
 794         LACP_LOCK_DESTROY(lsc);
 795         free(lsc, M_DEVBUF);
 796         return (0);
 797 }
 798
 799 void
 800 lacp_init(struct lagg_softc *sc)
 801 {
 802         struct lacp_softc *lsc = LACP_SOFTC(sc);
 803
 804         LACP_LOCK(lsc);
 805         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 806         LACP_UNLOCK(lsc);
 807 }
 808
 809 void
 810 lacp_stop(struct lagg_softc *sc)
 811 {
 812         struct lacp_softc *lsc = LACP_SOFTC(sc);
 813
 814         LACP_LOCK(lsc);
 815         callout_stop(&lsc->lsc_transit_callout);
 816         callout_stop(&lsc->lsc_callout);
 817         LACP_UNLOCK(lsc);
 818 }
 819
 820 struct lagg_port *
 821 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 822 {
 823         struct lacp_softc *lsc = LACP_SOFTC(sc);
 824         struct lacp_portmap *pm;
 825         struct lacp_port *lp;
 826         uint32_t hash;
 827
 828         if (__predict_false(lsc->lsc_suppress_distributing)) {
 829                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 830                 return (NULL);
 831         }
 832
 833         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 834         if (pm->pm_count == 0) {
 835                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 836                 return (NULL);
 837         }
 838
 839         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 840             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 841                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 842         else
 843                 hash = lagg_hashmbuf(sc, m, lsc->lsc_hashkey);
 844         hash %= pm->pm_count;
 845         lp = pm->pm_map[hash];
 846
 847         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 848             ("aggregated port is not distributing"));
 849
 850         return (lp->lp_lagg);
 851 }
 852 /*
 853  * lacp_suppress_distributing: drop transmit packets for a while
 854  * to preserve packet ordering.
 855  */
 856
 857 static void
 858 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 859 {
 860         struct lacp_port *lp;
 861
 862         if (lsc->lsc_active_aggregator != la) {
 863                 return;
 864         }
 865
 866         LACP_TRACE(NULL);
 867
 868         lsc->lsc_suppress_distributing = TRUE;
 869
 870         /* send a marker frame down each port to verify the queues are empty */
 871         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 872                 lp->lp_flags |= LACP_PORT_MARK;
 873                 lacp_xmit_marker(lp);
 874         }
 875
 876         /* set a timeout for the marker frames */
 877         callout_reset(&lsc->lsc_transit_callout,
 878             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 879 }
 880
 881 static int
 882 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 883     const struct lacp_peerinfo *b)
 884 {
 885         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 886 }
 887
 888 static int
 889 lacp_compare_systemid(const struct lacp_systemid *a,
 890     const struct lacp_systemid *b)
 891 {
 892         return (memcmp(a, b, sizeof(*a)));
 893 }
 894
 895 #if 0   /* unused */
 896 static int
 897 lacp_compare_portid(const struct lacp_portid *a,
 898     const struct lacp_portid *b)
 899 {
 900         return (memcmp(a, b, sizeof(*a)));
 901 }
 902 #endif
 903
 904 static uint64_t
 905 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 906 {
 907         struct lacp_port *lp;
 908         uint64_t speed;
 909
 910         lp = TAILQ_FIRST(&la->la_ports);
 911         if (lp == NULL) {
 912                 return (0);
 913         }
 914
 915         speed = ifmedia_baudrate(lp->lp_media);
 916         speed *= la->la_nports;
 917         if (speed == 0) {
 918                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 919                     lp->lp_media, la->la_nports));
 920         }
 921
 922         return (speed);
 923 }
 924
 925 /*
 926  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 927  * packets from lagg(4) interface.
 928  */
 929
 930 static void
 931 lacp_select_active_aggregator(struct lacp_softc *lsc)
 932 {
 933         struct lacp_aggregator *la;
 934         struct lacp_aggregator *best_la = NULL;
 935         uint64_t best_speed = 0;
 936         char buf[LACP_LAGIDSTR_MAX+1];
 937
 938         LACP_TRACE(NULL);
 939
 940         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 941                 uint64_t speed;
 942
 943                 if (la->la_nports == 0) {
 944                         continue;
 945                 }
 946
 947                 speed = lacp_aggregator_bandwidth(la);
 948                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 949                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 950                     speed, la->la_nports));
 951
 952                 /* This aggregator is chosen if
 953                  *      the partner has a better system priority
 954                  *  or, the total aggregated speed is higher
 955                  *  or, it is already the chosen aggregator
 956                  */
 957                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 958                      LACP_SYS_PRI(best_la->la_partner)) ||
 959                     speed > best_speed ||
 960                     (speed == best_speed &&
 961                     la == lsc->lsc_active_aggregator)) {
 962                         best_la = la;
 963                         best_speed = speed;
 964                 }
 965         }
 966
 967         KASSERT(best_la == NULL || best_la->la_nports > 0,
 968             ("invalid aggregator refcnt"));
 969         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
 970             ("invalid aggregator list"));
 971
 972         if (lsc->lsc_active_aggregator != best_la) {
 973                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
 974                 LACP_DPRINTF((NULL, "old %s\n",
 975                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
 976                     buf, sizeof(buf))));
 977         } else {
 978                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
 979         }
 980         LACP_DPRINTF((NULL, "new %s\n",
 981             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
 982
 983         if (lsc->lsc_active_aggregator != best_la) {
 984                 lsc->lsc_active_aggregator = best_la;
 985                 lacp_update_portmap(lsc);
 986                 if (best_la) {
 987                         lacp_suppress_distributing(lsc, best_la);
 988                 }
 989         }
 990 }
 991
 992 /*
 993  * Updated the inactive portmap array with the new list of ports and
 994  * make it live.
 995  */
 996 static void
 997 lacp_update_portmap(struct lacp_softc *lsc)
 998 {
 999         struct lagg_softc *sc = lsc->lsc_softc;
1000         struct lacp_aggregator *la;
1001         struct lacp_portmap *p;
1002         struct lacp_port *lp;
1003         uint64_t speed;
1004         u_int newmap;
1005         int i;
1006
1007         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1008         p = &lsc->lsc_pmap[newmap];
1009         la = lsc->lsc_active_aggregator;
1010         speed = 0;
1011         bzero(p, sizeof(struct lacp_portmap));
1012
1013         if (la != NULL && la->la_nports > 0) {
1014                 p->pm_count = la->la_nports;
1015                 i = 0;
1016                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1017                         p->pm_map[i++] = lp;
1018                 KASSERT(i == p->pm_count, ("Invalid port count"));
1019                 speed = lacp_aggregator_bandwidth(la);
1020         }
1021         sc->sc_ifp->if_baudrate = speed;
1022
1023         /* switch the active portmap over */
1024         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1025         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1026                     lsc->lsc_activemap,
1027                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1028 }
1029
1030 static uint16_t
1031 lacp_compose_key(struct lacp_port *lp)
1032 {
1033         struct lagg_port *lgp = lp->lp_lagg;
1034         struct lagg_softc *sc = lgp->lp_softc;
1035         u_int media = lp->lp_media;
1036         uint16_t key;
1037
1038         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1039
1040                 /*
1041                  * non-aggregatable links should have unique keys.
1042                  *
1043                  * XXX this isn't really unique as if_index is 16 bit.
1044                  */
1045
1046                 /* bit 0..14:   (some bits of) if_index of this port */
1047                 key = lp->lp_ifp->if_index;
1048                 /* bit 15:      1 */
1049                 key |= 0x8000;
1050         } else {
1051                 u_int subtype = IFM_SUBTYPE(media);
1052
1053                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1054                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1055
1056                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1057                 switch (subtype) {
1058                 case IFM_10_T:
1059                 case IFM_10_2:
1060                 case IFM_10_5:
1061                 case IFM_10_STP:
1062                 case IFM_10_FL:
1063                         key = IFM_10_T;
1064                         break;
1065                 case IFM_100_TX:
1066                 case IFM_100_FX:
1067                 case IFM_100_T4:
1068                 case IFM_100_VG:
1069                 case IFM_100_T2:
1070                 case IFM_100_T:
1071                         key = IFM_100_TX;
1072                         break;
1073                 case IFM_1000_SX:
1074                 case IFM_1000_LX:
1075                 case IFM_1000_CX:
1076                 case IFM_1000_T:
1077                 case IFM_1000_KX:
1078                 case IFM_1000_SGMII:
1079                 case IFM_1000_CX_SGMII:
1080                         key = IFM_1000_SX;
1081                         break;
1082                 case IFM_10G_LR:
1083                 case IFM_10G_SR:
1084                 case IFM_10G_CX4:
1085                 case IFM_10G_TWINAX:
1086                 case IFM_10G_TWINAX_LONG:
1087                 case IFM_10G_LRM:
1088                 case IFM_10G_T:
1089                 case IFM_10G_KX4:
1090                 case IFM_10G_KR:
1091                 case IFM_10G_CR1:
1092                 case IFM_10G_ER:
1093                 case IFM_10G_SFI:
1094                 case IFM_10G_AOC:
1095                         key = IFM_10G_LR;
1096                         break;
1097                 case IFM_20G_KR2:
1098                         key = IFM_20G_KR2;
1099                         break;
1100                 case IFM_2500_KX:
1101                 case IFM_2500_T:
1102                         key = IFM_2500_KX;
1103                         break;
1104                 case IFM_5000_T:
1105                         key = IFM_5000_T;
1106                         break;
1107                 case IFM_50G_PCIE:
1108                 case IFM_50G_CR2:
1109                 case IFM_50G_KR2:
1110                         key = IFM_50G_PCIE;
1111                         break;
1112                 case IFM_56G_R4:
1113                         key = IFM_56G_R4;
1114                         break;
1115                 case IFM_25G_PCIE:
1116                 case IFM_25G_CR:
1117                 case IFM_25G_KR:
1118                 case IFM_25G_SR:
1119                 case IFM_25G_LR:
1120                 case IFM_25G_ACC:
1121                 case IFM_25G_AOC:
1122                         key = IFM_25G_PCIE;
1123                         break;
1124                 case IFM_40G_CR4:
1125                 case IFM_40G_SR4:
1126                 case IFM_40G_LR4:
1127                 case IFM_40G_XLPPI:
1128                 case IFM_40G_KR4:
1129                         key = IFM_40G_CR4;
1130                         break;
1131                 case IFM_100G_CR4:
1132                 case IFM_100G_SR4:
1133                 case IFM_100G_KR4:
1134                 case IFM_100G_LR4:
1135                         key = IFM_100G_CR4;
1136                         break;
1137                 default:
1138                         key = subtype;
1139                         break;
1140                 }
1141                 /* bit 5..14:   (some bits of) if_index of lagg device */
1142                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1143                 /* bit 15:      0 */
1144         }
1145         return (htons(key));
1146 }
1147
1148 static void
1149 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1150 {
1151         char buf[LACP_LAGIDSTR_MAX+1];
1152
1153         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1154             __func__,
1155             lacp_format_lagid(&la->la_actor, &la->la_partner,
1156             buf, sizeof(buf)),
1157             la->la_refcnt, la->la_refcnt + 1));
1158
1159         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1160         la->la_refcnt++;
1161         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1162 }
1163
1164 static void
1165 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1166 {
1167         char buf[LACP_LAGIDSTR_MAX+1];
1168
1169         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1170             __func__,
1171             lacp_format_lagid(&la->la_actor, &la->la_partner,
1172             buf, sizeof(buf)),
1173             la->la_refcnt, la->la_refcnt - 1));
1174
1175         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1176         la->la_refcnt--;
1177         if (la->la_refcnt > 0) {
1178                 return;
1179         }
1180
1181         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1182         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1183
1184         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1185
1186         free(la, M_DEVBUF);
1187 }
1188
1189 /*
1190  * lacp_aggregator_get: allocate an aggregator.
1191  */
1192
1193 static struct lacp_aggregator *
1194 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1195 {
1196         struct lacp_aggregator *la;
1197
1198         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1199         if (la) {
1200                 la->la_refcnt = 1;
1201                 la->la_nports = 0;
1202                 TAILQ_INIT(&la->la_ports);
1203                 la->la_pending = 0;
1204                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1205         }
1206
1207         return (la);
1208 }
1209
1210 /*
1211  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1212  */
1213
1214 static void
1215 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1216 {
1217         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1218         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1219
1220         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1221 }
1222
1223 static void
1224 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1225     const struct lacp_peerinfo *lpi_port)
1226 {
1227         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1228         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1229         lpi_aggr->lip_key = lpi_port->lip_key;
1230 }
1231
1232 /*
1233  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1234  */
1235
1236 static int
1237 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1238     const struct lacp_port *lp)
1239 {
1240         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1241             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1242                 return (0);
1243         }
1244
1245         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1246                 return (0);
1247         }
1248
1249         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1250                 return (0);
1251         }
1252
1253         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1254                 return (0);
1255         }
1256
1257         return (1);
1258 }
1259
1260 static int
1261 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1262     const struct lacp_peerinfo *b)
1263 {
1264         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1265             sizeof(a->lip_systemid))) {
1266                 return (0);
1267         }
1268
1269         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1270                 return (0);
1271         }
1272
1273         return (1);
1274 }
1275
1276 static void
1277 lacp_port_enable(struct lacp_port *lp)
1278 {
1279         lp->lp_state |= LACP_STATE_AGGREGATION;
1280 }
1281
1282 static void
1283 lacp_port_disable(struct lacp_port *lp)
1284 {
1285         lacp_set_mux(lp, LACP_MUX_DETACHED);
1286
1287         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1288         lp->lp_selected = LACP_UNSELECTED;
1289         lacp_sm_rx_record_default(lp);
1290         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1291         lp->lp_state &= ~LACP_STATE_EXPIRED;
1292 }
1293
1294 /*
1295  * lacp_select: select an aggregator.  create one if necessary.
1296  */
1297 static void
1298 lacp_select(struct lacp_port *lp)
1299 {
1300         struct lacp_softc *lsc = lp->lp_lsc;
1301         struct lacp_aggregator *la;
1302         char buf[LACP_LAGIDSTR_MAX+1];
1303
1304         if (lp->lp_aggregator) {
1305                 return;
1306         }
1307
1308         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1309             ("timer_wait_while still active"));
1310
1311         LACP_DPRINTF((lp, "port lagid=%s\n",
1312             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1313             buf, sizeof(buf))));
1314
1315         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1316                 if (lacp_aggregator_is_compatible(la, lp)) {
1317                         break;
1318                 }
1319         }
1320
1321         if (la == NULL) {
1322                 la = lacp_aggregator_get(lsc, lp);
1323                 if (la == NULL) {
1324                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1325
1326                         /*
1327                          * will retry on the next tick.
1328                          */
1329
1330                         return;
1331                 }
1332                 lacp_fill_aggregator_id(la, lp);
1333                 LACP_DPRINTF((lp, "aggregator created\n"));
1334         } else {
1335                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1336                 if (la->la_refcnt == LACP_MAX_PORTS)
1337                         return;
1338                 lacp_aggregator_addref(lsc, la);
1339         }
1340
1341         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1342             lacp_format_lagid(&la->la_actor, &la->la_partner,
1343             buf, sizeof(buf))));
1344
1345         lp->lp_aggregator = la;
1346         lp->lp_selected = LACP_SELECTED;
1347 }
1348
1349 /*
1350  * lacp_unselect: finish unselect/detach process.
1351  */
1352
1353 static void
1354 lacp_unselect(struct lacp_port *lp)
1355 {
1356         struct lacp_softc *lsc = lp->lp_lsc;
1357         struct lacp_aggregator *la = lp->lp_aggregator;
1358
1359         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1360             ("timer_wait_while still active"));
1361
1362         if (la == NULL) {
1363                 return;
1364         }
1365
1366         lp->lp_aggregator = NULL;
1367         lacp_aggregator_delref(lsc, la);
1368 }
1369
1370 /* mux machine */
1371
1372 static void
1373 lacp_sm_mux(struct lacp_port *lp)
1374 {
1375         struct lagg_port *lgp = lp->lp_lagg;
1376         struct lagg_softc *sc = lgp->lp_softc;
1377         enum lacp_mux_state new_state;
1378         boolean_t p_sync =
1379                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1380         boolean_t p_collecting =
1381             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1382         enum lacp_selected selected = lp->lp_selected;
1383         struct lacp_aggregator *la;
1384
1385         if (V_lacp_debug > 1)
1386                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1387                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1388                     lp->lp_mux_state, selected, p_sync, p_collecting);
1389
1390 re_eval:
1391         la = lp->lp_aggregator;
1392         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1393             ("MUX not detached"));
1394         new_state = lp->lp_mux_state;
1395         switch (lp->lp_mux_state) {
1396         case LACP_MUX_DETACHED:
1397                 if (selected != LACP_UNSELECTED) {
1398                         new_state = LACP_MUX_WAITING;
1399                 }
1400                 break;
1401         case LACP_MUX_WAITING:
1402                 KASSERT(la->la_pending > 0 ||
1403                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1404                     ("timer_wait_while still active"));
1405                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1406                         new_state = LACP_MUX_ATTACHED;
1407                 } else if (selected == LACP_UNSELECTED) {
1408                         new_state = LACP_MUX_DETACHED;
1409                 }
1410                 break;
1411         case LACP_MUX_ATTACHED:
1412                 if (selected == LACP_SELECTED && p_sync) {
1413                         new_state = LACP_MUX_COLLECTING;
1414                 } else if (selected != LACP_SELECTED) {
1415                         new_state = LACP_MUX_DETACHED;
1416                 }
1417                 break;
1418         case LACP_MUX_COLLECTING:
1419                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1420                         new_state = LACP_MUX_DISTRIBUTING;
1421                 } else if (selected != LACP_SELECTED || !p_sync) {
1422                         new_state = LACP_MUX_ATTACHED;
1423                 }
1424                 break;
1425         case LACP_MUX_DISTRIBUTING:
1426                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1427                         new_state = LACP_MUX_COLLECTING;
1428                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1429                         sc->sc_flapping++;
1430                 }
1431                 break;
1432         default:
1433                 panic("%s: unknown state", __func__);
1434         }
1435
1436         if (lp->lp_mux_state == new_state) {
1437                 return;
1438         }
1439
1440         lacp_set_mux(lp, new_state);
1441         goto re_eval;
1442 }
1443
1444 static void
1445 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1446 {
1447         struct lacp_aggregator *la = lp->lp_aggregator;
1448
1449         if (lp->lp_mux_state == new_state) {
1450                 return;
1451         }
1452
1453         switch (new_state) {
1454         case LACP_MUX_DETACHED:
1455                 lp->lp_state &= ~LACP_STATE_SYNC;
1456                 lacp_disable_distributing(lp);
1457                 lacp_disable_collecting(lp);
1458                 lacp_sm_assert_ntt(lp);
1459                 /* cancel timer */
1460                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1461                         KASSERT(la->la_pending > 0,
1462                             ("timer_wait_while not active"));
1463                         la->la_pending--;
1464                 }
1465                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1466                 lacp_unselect(lp);
1467                 break;
1468         case LACP_MUX_WAITING:
1469                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1470                     LACP_AGGREGATE_WAIT_TIME);
1471                 la->la_pending++;
1472                 break;
1473         case LACP_MUX_ATTACHED:
1474                 lp->lp_state |= LACP_STATE_SYNC;
1475                 lacp_disable_collecting(lp);
1476                 lacp_sm_assert_ntt(lp);
1477                 break;
1478         case LACP_MUX_COLLECTING:
1479                 lacp_enable_collecting(lp);
1480                 lacp_disable_distributing(lp);
1481                 lacp_sm_assert_ntt(lp);
1482                 break;
1483         case LACP_MUX_DISTRIBUTING:
1484                 lacp_enable_distributing(lp);
1485                 break;
1486         default:
1487                 panic("%s: unknown state", __func__);
1488         }
1489
1490         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1491
1492         lp->lp_mux_state = new_state;
1493 }
1494
1495 static void
1496 lacp_sm_mux_timer(struct lacp_port *lp)
1497 {
1498         struct lacp_aggregator *la = lp->lp_aggregator;
1499         char buf[LACP_LAGIDSTR_MAX+1];
1500
1501         KASSERT(la->la_pending > 0, ("no pending event"));
1502
1503         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1504             lacp_format_lagid(&la->la_actor, &la->la_partner,
1505             buf, sizeof(buf)),
1506             la->la_pending, la->la_pending - 1));
1507
1508         la->la_pending--;
1509 }
1510
1511 /* periodic transmit machine */
1512
1513 static void
1514 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1515 {
1516         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1517             LACP_STATE_TIMEOUT)) {
1518                 return;
1519         }
1520
1521         LACP_DPRINTF((lp, "partner timeout changed\n"));
1522
1523         /*
1524          * FAST_PERIODIC -> SLOW_PERIODIC
1525          * or
1526          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1527          *
1528          * let lacp_sm_ptx_tx_schedule to update timeout.
1529          */
1530
1531         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1532
1533         /*
1534          * if timeout has been shortened, assert NTT.
1535          */
1536
1537         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1538                 lacp_sm_assert_ntt(lp);
1539         }
1540 }
1541
1542 static void
1543 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1544 {
1545         int timeout;
1546
1547         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1548             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1549
1550                 /*
1551                  * NO_PERIODIC
1552                  */
1553
1554                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1555                 return;
1556         }
1557
1558         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1559                 return;
1560         }
1561
1562         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1563             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1564
1565         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1566 }
1567
1568 static void
1569 lacp_sm_ptx_timer(struct lacp_port *lp)
1570 {
1571         lacp_sm_assert_ntt(lp);
1572 }
1573
1574 static void
1575 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1576 {
1577         int timeout;
1578
1579         /*
1580          * check LACP_DISABLED first
1581          */
1582
1583         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1584                 return;
1585         }
1586
1587         /*
1588          * check loopback condition.
1589          */
1590
1591         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1592             &lp->lp_actor.lip_systemid)) {
1593                 return;
1594         }
1595
1596         /*
1597          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1598          */
1599
1600         lacp_sm_rx_update_selected(lp, du);
1601         lacp_sm_rx_update_ntt(lp, du);
1602         lacp_sm_rx_record_pdu(lp, du);
1603
1604         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1605             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1606         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1607
1608         lp->lp_state &= ~LACP_STATE_EXPIRED;
1609
1610         /*
1611          * kick transmit machine without waiting the next tick.
1612          */
1613
1614         lacp_sm_tx(lp);
1615 }
1616
1617 static void
1618 lacp_sm_rx_set_expired(struct lacp_port *lp)
1619 {
1620         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1621         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1622         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1623         lp->lp_state |= LACP_STATE_EXPIRED;
1624 }
1625
1626 static void
1627 lacp_sm_rx_timer(struct lacp_port *lp)
1628 {
1629         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1630                 /* CURRENT -> EXPIRED */
1631                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1632                 lacp_sm_rx_set_expired(lp);
1633         } else {
1634                 /* EXPIRED -> DEFAULTED */
1635                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1636                 lacp_sm_rx_update_default_selected(lp);
1637                 lacp_sm_rx_record_default(lp);
1638                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1639         }
1640 }
1641
1642 static void
1643 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1644 {
1645         boolean_t active;
1646         uint8_t oldpstate;
1647         char buf[LACP_STATESTR_MAX+1];
1648
1649         LACP_TRACE(lp);
1650
1651         oldpstate = lp->lp_partner.lip_state;
1652
1653         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1654             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1655             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1656
1657         lp->lp_partner = du->ldu_actor;
1658         if (active &&
1659             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1660             LACP_STATE_AGGREGATION) &&
1661             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1662             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1663                 /* XXX nothing? */
1664         } else {
1665                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1666         }
1667
1668         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1669
1670         if (oldpstate != lp->lp_partner.lip_state) {
1671                 LACP_DPRINTF((lp, "old pstate %s\n",
1672                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1673                 LACP_DPRINTF((lp, "new pstate %s\n",
1674                     lacp_format_state(lp->lp_partner.lip_state, buf,
1675                     sizeof(buf))));
1676         }
1677
1678         /* XXX Hack, still need to implement 5.4.9 para 2,3,4 */
1679         if (lp->lp_lsc->lsc_strict_mode)
1680                 lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1681
1682         lacp_sm_ptx_update_timeout(lp, oldpstate);
1683 }
1684
1685 static void
1686 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1687 {
1688
1689         LACP_TRACE(lp);
1690
1691         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1692             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1693             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1694                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1695                 lacp_sm_assert_ntt(lp);
1696         }
1697 }
1698
1699 static void
1700 lacp_sm_rx_record_default(struct lacp_port *lp)
1701 {
1702         uint8_t oldpstate;
1703
1704         LACP_TRACE(lp);
1705
1706         oldpstate = lp->lp_partner.lip_state;
1707         if (lp->lp_lsc->lsc_strict_mode)
1708                 lp->lp_partner = lacp_partner_admin_strict;
1709         else
1710                 lp->lp_partner = lacp_partner_admin_optimistic;
1711         lp->lp_state |= LACP_STATE_DEFAULTED;
1712         lacp_sm_ptx_update_timeout(lp, oldpstate);
1713 }
1714
1715 static void
1716 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1717     const struct lacp_peerinfo *info)
1718 {
1719
1720         LACP_TRACE(lp);
1721
1722         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1723             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1724             LACP_STATE_AGGREGATION)) {
1725                 lp->lp_selected = LACP_UNSELECTED;
1726                 /* mux machine will clean up lp->lp_aggregator */
1727         }
1728 }
1729
1730 static void
1731 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1732 {
1733
1734         LACP_TRACE(lp);
1735
1736         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1737 }
1738
1739 static void
1740 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1741 {
1742
1743         LACP_TRACE(lp);
1744
1745         if (lp->lp_lsc->lsc_strict_mode)
1746                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1747                     &lacp_partner_admin_strict);
1748         else
1749                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1750                     &lacp_partner_admin_optimistic);
1751 }
1752
1753 /* transmit machine */
1754
1755 static void
1756 lacp_sm_tx(struct lacp_port *lp)
1757 {
1758         int error = 0;
1759
1760         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1761 #if 1
1762             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1763             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1764 #endif
1765             ) {
1766                 lp->lp_flags &= ~LACP_PORT_NTT;
1767         }
1768
1769         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1770                 return;
1771         }
1772
1773         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1774         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1775                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1776                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1777                 return;
1778         }
1779
1780         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1781                 error = lacp_xmit_lacpdu(lp);
1782         } else {
1783                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1784         }
1785
1786         if (error == 0) {
1787                 lp->lp_flags &= ~LACP_PORT_NTT;
1788         } else {
1789                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1790                     error));
1791         }
1792 }
1793
1794 static void
1795 lacp_sm_assert_ntt(struct lacp_port *lp)
1796 {
1797
1798         lp->lp_flags |= LACP_PORT_NTT;
1799 }
1800
1801 static void
1802 lacp_run_timers(struct lacp_port *lp)
1803 {
1804         int i;
1805
1806         for (i = 0; i < LACP_NTIMER; i++) {
1807                 KASSERT(lp->lp_timer[i] >= 0,
1808                     ("invalid timer value %d", lp->lp_timer[i]));
1809                 if (lp->lp_timer[i] == 0) {
1810                         continue;
1811                 } else if (--lp->lp_timer[i] <= 0) {
1812                         if (lacp_timer_funcs[i]) {
1813                                 (*lacp_timer_funcs[i])(lp);
1814                         }
1815                 }
1816         }
1817 }
1818
1819 int
1820 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1821 {
1822         struct lacp_softc *lsc = lp->lp_lsc;
1823         struct lagg_port *lgp = lp->lp_lagg;
1824         struct lacp_port *lp2;
1825         struct markerdu *mdu;
1826         int error = 0;
1827         int pending = 0;
1828
1829         if (m->m_pkthdr.len != sizeof(*mdu)) {
1830                 goto bad;
1831         }
1832
1833         if ((m->m_flags & M_MCAST) == 0) {
1834                 goto bad;
1835         }
1836
1837         if (m->m_len < sizeof(*mdu)) {
1838                 m = m_pullup(m, sizeof(*mdu));
1839                 if (m == NULL) {
1840                         return (ENOMEM);
1841                 }
1842         }
1843
1844         mdu = mtod(m, struct markerdu *);
1845
1846         if (memcmp(&mdu->mdu_eh.ether_dhost,
1847             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1848                 goto bad;
1849         }
1850
1851         if (mdu->mdu_sph.sph_version != 1) {
1852                 goto bad;
1853         }
1854
1855         switch (mdu->mdu_tlv.tlv_type) {
1856         case MARKER_TYPE_INFO:
1857                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1858                     marker_info_tlv_template, TRUE)) {
1859                         goto bad;
1860                 }
1861                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1862                 memcpy(&mdu->mdu_eh.ether_dhost,
1863                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1864                 memcpy(&mdu->mdu_eh.ether_shost,
1865                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1866                 error = lagg_enqueue(lp->lp_ifp, m);
1867                 break;
1868
1869         case MARKER_TYPE_RESPONSE:
1870                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1871                     marker_response_tlv_template, TRUE)) {
1872                         goto bad;
1873                 }
1874                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1875                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1876                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1877
1878                 /* Verify that it is the last marker we sent out */
1879                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1880                     sizeof(struct lacp_markerinfo)))
1881                         goto bad;
1882
1883                 LACP_LOCK(lsc);
1884                 lp->lp_flags &= ~LACP_PORT_MARK;
1885
1886                 if (lsc->lsc_suppress_distributing) {
1887                         /* Check if any ports are waiting for a response */
1888                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1889                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1890                                         pending = 1;
1891                                         break;
1892                                 }
1893                         }
1894
1895                         if (pending == 0) {
1896                                 /* All interface queues are clear */
1897                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1898                                 lsc->lsc_suppress_distributing = FALSE;
1899                         }
1900                 }
1901                 LACP_UNLOCK(lsc);
1902                 m_freem(m);
1903                 break;
1904
1905         default:
1906                 goto bad;
1907         }
1908
1909         return (error);
1910
1911 bad:
1912         LACP_DPRINTF((lp, "bad marker frame\n"));
1913         m_freem(m);
1914         return (EINVAL);
1915 }
1916
1917 static int
1918 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1919     const struct tlv_template *tmpl, boolean_t check_type)
1920 {
1921         while (/* CONSTCOND */ 1) {
1922                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1923                         return (EINVAL);
1924                 }
1925                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1926                     tlv->tlv_length != tmpl->tmpl_length) {
1927                         return (EINVAL);
1928                 }
1929                 if (tmpl->tmpl_type == 0) {
1930                         break;
1931                 }
1932                 tlv = (const struct tlvhdr *)
1933                     ((const char *)tlv + tlv->tlv_length);
1934                 tmpl++;
1935         }
1936
1937         return (0);
1938 }
1939
1940 /* Debugging */
1941 const char *
1942 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1943 {
1944         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1945             (int)mac[0],
1946             (int)mac[1],
1947             (int)mac[2],
1948             (int)mac[3],
1949             (int)mac[4],
1950             (int)mac[5]);
1951
1952         return (buf);
1953 }
1954
1955 const char *
1956 lacp_format_systemid(const struct lacp_systemid *sysid,
1957     char *buf, size_t buflen)
1958 {
1959         char macbuf[LACP_MACSTR_MAX+1];
1960
1961         snprintf(buf, buflen, "%04X,%s",
1962             ntohs(sysid->lsi_prio),
1963             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1964
1965         return (buf);
1966 }
1967
1968 const char *
1969 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
1970 {
1971         snprintf(buf, buflen, "%04X,%04X",
1972             ntohs(portid->lpi_prio),
1973             ntohs(portid->lpi_portno));
1974
1975         return (buf);
1976 }
1977
1978 const char *
1979 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
1980 {
1981         char sysid[LACP_SYSTEMIDSTR_MAX+1];
1982         char portid[LACP_PORTIDSTR_MAX+1];
1983
1984         snprintf(buf, buflen, "(%s,%04X,%s)",
1985             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
1986             ntohs(peer->lip_key),
1987             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
1988
1989         return (buf);
1990 }
1991
1992 const char *
1993 lacp_format_lagid(const struct lacp_peerinfo *a,
1994     const struct lacp_peerinfo *b, char *buf, size_t buflen)
1995 {
1996         char astr[LACP_PARTNERSTR_MAX+1];
1997         char bstr[LACP_PARTNERSTR_MAX+1];
1998
1999 #if 0
2000         /*
2001          * there's a convention to display small numbered peer
2002          * in the left.
2003          */
2004
2005         if (lacp_compare_peerinfo(a, b) > 0) {
2006                 const struct lacp_peerinfo *t;
2007
2008                 t = a;
2009                 a = b;
2010                 b = t;
2011         }
2012 #endif
2013
2014         snprintf(buf, buflen, "[%s,%s]",
2015             lacp_format_partner(a, astr, sizeof(astr)),
2016             lacp_format_partner(b, bstr, sizeof(bstr)));
2017
2018         return (buf);
2019 }
2020
2021 const char *
2022 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2023     char *buf, size_t buflen)
2024 {
2025         if (la == NULL) {
2026                 return ("(none)");
2027         }
2028
2029         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2030 }
2031
2032 const char *
2033 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2034 {
2035         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2036         return (buf);
2037 }
2038
2039 static void
2040 lacp_dump_lacpdu(const struct lacpdu *du)
2041 {
2042         char buf[LACP_PARTNERSTR_MAX+1];
2043         char buf2[LACP_STATESTR_MAX+1];
2044
2045         printf("actor=%s\n",
2046             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2047         printf("actor.state=%s\n",
2048             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2049         printf("partner=%s\n",
2050             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2051         printf("partner.state=%s\n",
2052             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2053
2054         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2055 }
2056
2057 static void
2058 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2059 {
2060         va_list va;
2061
2062         if (lp) {
2063                 printf("%s: ", lp->lp_ifp->if_xname);
2064         }
2065
2066         va_start(va, fmt);
2067         vprintf(fmt, va);
2068         va_end(va);
2069 }