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         boolean_t active = TRUE; /* XXX should be configurable */
 527         boolean_t fast = FALSE; /* Configurable via ioctl */
 528
 529         bzero((char *)&sdl, sizeof(sdl));
 530         sdl.sdl_len = sizeof(sdl);
 531         sdl.sdl_family = AF_LINK;
 532         sdl.sdl_index = ifp->if_index;
 533         sdl.sdl_type = IFT_ETHER;
 534         sdl.sdl_alen = ETHER_ADDR_LEN;
 535
 536         bcopy(&ethermulticastaddr_slowprotocols,
 537             LLADDR(&sdl), ETHER_ADDR_LEN);
 538         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 539         if (error) {
 540                 printf("%s: ADDMULTI failed on %s\n", __func__, lgp->lp_ifname);
 541                 return (error);
 542         }
 543
 544         lp = malloc(sizeof(struct lacp_port),
 545             M_DEVBUF, M_NOWAIT|M_ZERO);
 546         if (lp == NULL)
 547                 return (ENOMEM);
 548
 549         LACP_LOCK(lsc);
 550         lgp->lp_psc = (caddr_t)lp;
 551         lp->lp_ifp = ifp;
 552         lp->lp_lagg = lgp;
 553         lp->lp_lsc = lsc;
 554         lp->lp_ifma = rifma;
 555
 556         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 557
 558         lacp_fill_actorinfo(lp, &lp->lp_actor);
 559         lacp_fill_markerinfo(lp, &lp->lp_marker);
 560         lp->lp_state =
 561             (active ? LACP_STATE_ACTIVITY : 0) |
 562             (fast ? LACP_STATE_TIMEOUT : 0);
 563         lp->lp_aggregator = NULL;
 564         lacp_sm_rx_set_expired(lp);
 565         LACP_UNLOCK(lsc);
 566         lacp_linkstate(lgp);
 567
 568         return (0);
 569 }
 570
 571 void
 572 lacp_port_destroy(struct lagg_port *lgp)
 573 {
 574         struct lacp_port *lp = LACP_PORT(lgp);
 575         struct lacp_softc *lsc = lp->lp_lsc;
 576         int i;
 577
 578         LACP_LOCK(lsc);
 579         for (i = 0; i < LACP_NTIMER; i++) {
 580                 LACP_TIMER_DISARM(lp, i);
 581         }
 582
 583         lacp_disable_collecting(lp);
 584         lacp_disable_distributing(lp);
 585         lacp_unselect(lp);
 586
 587         LIST_REMOVE(lp, lp_next);
 588         LACP_UNLOCK(lsc);
 589
 590         /* The address may have already been removed by if_purgemaddrs() */
 591         if (!lgp->lp_detaching)
 592                 if_delmulti_ifma(lp->lp_ifma);
 593
 594         free(lp, M_DEVBUF);
 595 }
 596
 597 void
 598 lacp_req(struct lagg_softc *sc, caddr_t data)
 599 {
 600         struct lacp_opreq *req = (struct lacp_opreq *)data;
 601         struct lacp_softc *lsc = LACP_SOFTC(sc);
 602         struct lacp_aggregator *la;
 603
 604         bzero(req, sizeof(struct lacp_opreq));
 605
 606         /*
 607          * If the LACP softc is NULL, return with the opreq structure full of
 608          * zeros.  It is normal for the softc to be NULL while the lagg is
 609          * being destroyed.
 610          */
 611         if (NULL == lsc)
 612                 return;
 613
 614         la = lsc->lsc_active_aggregator;
 615         LACP_LOCK(lsc);
 616         if (la != NULL) {
 617                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 618                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 619                     ETHER_ADDR_LEN);
 620                 req->actor_key = ntohs(la->la_actor.lip_key);
 621                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 622                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 623                 req->actor_state = la->la_actor.lip_state;
 624
 625                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 626                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 627                     ETHER_ADDR_LEN);
 628                 req->partner_key = ntohs(la->la_partner.lip_key);
 629                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 630                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 631                 req->partner_state = la->la_partner.lip_state;
 632         }
 633         LACP_UNLOCK(lsc);
 634 }
 635
 636 void
 637 lacp_portreq(struct lagg_port *lgp, caddr_t data)
 638 {
 639         struct lacp_opreq *req = (struct lacp_opreq *)data;
 640         struct lacp_port *lp = LACP_PORT(lgp);
 641         struct lacp_softc *lsc = lp->lp_lsc;
 642
 643         LACP_LOCK(lsc);
 644         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 645         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 646             ETHER_ADDR_LEN);
 647         req->actor_key = ntohs(lp->lp_actor.lip_key);
 648         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 649         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 650         req->actor_state = lp->lp_actor.lip_state;
 651
 652         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 653         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 654             ETHER_ADDR_LEN);
 655         req->partner_key = ntohs(lp->lp_partner.lip_key);
 656         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 657         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 658         req->partner_state = lp->lp_partner.lip_state;
 659         LACP_UNLOCK(lsc);
 660 }
 661
 662 static void
 663 lacp_disable_collecting(struct lacp_port *lp)
 664 {
 665         LACP_DPRINTF((lp, "collecting disabled\n"));
 666         lp->lp_state &= ~LACP_STATE_COLLECTING;
 667 }
 668
 669 static void
 670 lacp_enable_collecting(struct lacp_port *lp)
 671 {
 672         LACP_DPRINTF((lp, "collecting enabled\n"));
 673         lp->lp_state |= LACP_STATE_COLLECTING;
 674 }
 675
 676 static void
 677 lacp_disable_distributing(struct lacp_port *lp)
 678 {
 679         struct lacp_aggregator *la = lp->lp_aggregator;
 680         struct lacp_softc *lsc = lp->lp_lsc;
 681         struct lagg_softc *sc = lsc->lsc_softc;
 682         char buf[LACP_LAGIDSTR_MAX+1];
 683
 684         LACP_LOCK_ASSERT(lsc);
 685
 686         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 687                 return;
 688         }
 689
 690         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 691         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 692         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 693
 694         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 695             "nports %d -> %d\n",
 696             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 697             la->la_nports, la->la_nports - 1));
 698
 699         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 700         la->la_nports--;
 701         sc->sc_active = la->la_nports;
 702
 703         if (lsc->lsc_active_aggregator == la) {
 704                 lacp_suppress_distributing(lsc, la);
 705                 lacp_select_active_aggregator(lsc);
 706                 /* regenerate the port map, the active aggregator has changed */
 707                 lacp_update_portmap(lsc);
 708         }
 709
 710         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 711 }
 712
 713 static void
 714 lacp_enable_distributing(struct lacp_port *lp)
 715 {
 716         struct lacp_aggregator *la = lp->lp_aggregator;
 717         struct lacp_softc *lsc = lp->lp_lsc;
 718         struct lagg_softc *sc = lsc->lsc_softc;
 719         char buf[LACP_LAGIDSTR_MAX+1];
 720
 721         LACP_LOCK_ASSERT(lsc);
 722
 723         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 724                 return;
 725         }
 726
 727         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 728             "nports %d -> %d\n",
 729             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 730             la->la_nports, la->la_nports + 1));
 731
 732         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 733         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 734         la->la_nports++;
 735         sc->sc_active = la->la_nports;
 736
 737         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 738
 739         if (lsc->lsc_active_aggregator == la) {
 740                 lacp_suppress_distributing(lsc, la);
 741                 lacp_update_portmap(lsc);
 742         } else
 743                 /* try to become the active aggregator */
 744                 lacp_select_active_aggregator(lsc);
 745 }
 746
 747 static void
 748 lacp_transit_expire(void *vp)
 749 {
 750         struct lacp_softc *lsc = vp;
 751
 752         LACP_LOCK_ASSERT(lsc);
 753
 754         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 755         LACP_TRACE(NULL);
 756         CURVNET_RESTORE();
 757
 758         lsc->lsc_suppress_distributing = FALSE;
 759 }
 760
 761 void
 762 lacp_attach(struct lagg_softc *sc)
 763 {
 764         struct lacp_softc *lsc;
 765
 766         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 767
 768         sc->sc_psc = (caddr_t)lsc;
 769         lsc->lsc_softc = sc;
 770
 771         lsc->lsc_hashkey = arc4random();
 772         lsc->lsc_active_aggregator = NULL;
 773         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 774         LACP_LOCK_INIT(lsc);
 775         TAILQ_INIT(&lsc->lsc_aggregators);
 776         LIST_INIT(&lsc->lsc_ports);
 777
 778         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 779         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 780
 781         /* if the lagg is already up then do the same */
 782         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 783                 lacp_init(sc);
 784 }
 785
 786 int
 787 lacp_detach(void *psc)
 788 {
 789         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 790
 791         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 792             ("aggregators still active"));
 793         KASSERT(lsc->lsc_active_aggregator == NULL,
 794             ("aggregator still attached"));
 795
 796         callout_drain(&lsc->lsc_transit_callout);
 797         callout_drain(&lsc->lsc_callout);
 798
 799         LACP_LOCK_DESTROY(lsc);
 800         free(lsc, M_DEVBUF);
 801         return (0);
 802 }
 803
 804 void
 805 lacp_init(struct lagg_softc *sc)
 806 {
 807         struct lacp_softc *lsc = LACP_SOFTC(sc);
 808
 809         LACP_LOCK(lsc);
 810         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 811         LACP_UNLOCK(lsc);
 812 }
 813
 814 void
 815 lacp_stop(struct lagg_softc *sc)
 816 {
 817         struct lacp_softc *lsc = LACP_SOFTC(sc);
 818
 819         LACP_LOCK(lsc);
 820         callout_stop(&lsc->lsc_transit_callout);
 821         callout_stop(&lsc->lsc_callout);
 822         LACP_UNLOCK(lsc);
 823 }
 824
 825 struct lagg_port *
 826 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 827 {
 828         struct lacp_softc *lsc = LACP_SOFTC(sc);
 829         struct lacp_portmap *pm;
 830         struct lacp_port *lp;
 831         uint32_t hash;
 832
 833         if (__predict_false(lsc->lsc_suppress_distributing)) {
 834                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 835                 return (NULL);
 836         }
 837
 838         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 839         if (pm->pm_count == 0) {
 840                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 841                 return (NULL);
 842         }
 843
 844         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 845             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 846                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 847         else
 848                 hash = lagg_hashmbuf(sc, m, lsc->lsc_hashkey);
 849         hash %= pm->pm_count;
 850         lp = pm->pm_map[hash];
 851
 852         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 853             ("aggregated port is not distributing"));
 854
 855         return (lp->lp_lagg);
 856 }
 857 /*
 858  * lacp_suppress_distributing: drop transmit packets for a while
 859  * to preserve packet ordering.
 860  */
 861
 862 static void
 863 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 864 {
 865         struct lacp_port *lp;
 866
 867         if (lsc->lsc_active_aggregator != la) {
 868                 return;
 869         }
 870
 871         LACP_TRACE(NULL);
 872
 873         lsc->lsc_suppress_distributing = TRUE;
 874
 875         /* send a marker frame down each port to verify the queues are empty */
 876         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 877                 lp->lp_flags |= LACP_PORT_MARK;
 878                 lacp_xmit_marker(lp);
 879         }
 880
 881         /* set a timeout for the marker frames */
 882         callout_reset(&lsc->lsc_transit_callout,
 883             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 884 }
 885
 886 static int
 887 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 888     const struct lacp_peerinfo *b)
 889 {
 890         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 891 }
 892
 893 static int
 894 lacp_compare_systemid(const struct lacp_systemid *a,
 895     const struct lacp_systemid *b)
 896 {
 897         return (memcmp(a, b, sizeof(*a)));
 898 }
 899
 900 #if 0   /* unused */
 901 static int
 902 lacp_compare_portid(const struct lacp_portid *a,
 903     const struct lacp_portid *b)
 904 {
 905         return (memcmp(a, b, sizeof(*a)));
 906 }
 907 #endif
 908
 909 static uint64_t
 910 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 911 {
 912         struct lacp_port *lp;
 913         uint64_t speed;
 914
 915         lp = TAILQ_FIRST(&la->la_ports);
 916         if (lp == NULL) {
 917                 return (0);
 918         }
 919
 920         speed = ifmedia_baudrate(lp->lp_media);
 921         speed *= la->la_nports;
 922         if (speed == 0) {
 923                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 924                     lp->lp_media, la->la_nports));
 925         }
 926
 927         return (speed);
 928 }
 929
 930 /*
 931  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 932  * packets from lagg(4) interface.
 933  */
 934
 935 static void
 936 lacp_select_active_aggregator(struct lacp_softc *lsc)
 937 {
 938         struct lacp_aggregator *la;
 939         struct lacp_aggregator *best_la = NULL;
 940         uint64_t best_speed = 0;
 941         char buf[LACP_LAGIDSTR_MAX+1];
 942
 943         LACP_TRACE(NULL);
 944
 945         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 946                 uint64_t speed;
 947
 948                 if (la->la_nports == 0) {
 949                         continue;
 950                 }
 951
 952                 speed = lacp_aggregator_bandwidth(la);
 953                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 954                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 955                     speed, la->la_nports));
 956
 957                 /* This aggregator is chosen if
 958                  *      the partner has a better system priority
 959                  *  or, the total aggregated speed is higher
 960                  *  or, it is already the chosen aggregator
 961                  */
 962                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 963                      LACP_SYS_PRI(best_la->la_partner)) ||
 964                     speed > best_speed ||
 965                     (speed == best_speed &&
 966                     la == lsc->lsc_active_aggregator)) {
 967                         best_la = la;
 968                         best_speed = speed;
 969                 }
 970         }
 971
 972         KASSERT(best_la == NULL || best_la->la_nports > 0,
 973             ("invalid aggregator refcnt"));
 974         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
 975             ("invalid aggregator list"));
 976
 977         if (lsc->lsc_active_aggregator != best_la) {
 978                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
 979                 LACP_DPRINTF((NULL, "old %s\n",
 980                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
 981                     buf, sizeof(buf))));
 982         } else {
 983                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
 984         }
 985         LACP_DPRINTF((NULL, "new %s\n",
 986             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
 987
 988         if (lsc->lsc_active_aggregator != best_la) {
 989                 lsc->lsc_active_aggregator = best_la;
 990                 lacp_update_portmap(lsc);
 991                 if (best_la) {
 992                         lacp_suppress_distributing(lsc, best_la);
 993                 }
 994         }
 995 }
 996
 997 /*
 998  * Updated the inactive portmap array with the new list of ports and
 999  * make it live.
1000  */
1001 static void
1002 lacp_update_portmap(struct lacp_softc *lsc)
1003 {
1004         struct lagg_softc *sc = lsc->lsc_softc;
1005         struct lacp_aggregator *la;
1006         struct lacp_portmap *p;
1007         struct lacp_port *lp;
1008         uint64_t speed;
1009         u_int newmap;
1010         int i;
1011
1012         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1013         p = &lsc->lsc_pmap[newmap];
1014         la = lsc->lsc_active_aggregator;
1015         speed = 0;
1016         bzero(p, sizeof(struct lacp_portmap));
1017
1018         if (la != NULL && la->la_nports > 0) {
1019                 p->pm_count = la->la_nports;
1020                 i = 0;
1021                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1022                         p->pm_map[i++] = lp;
1023                 KASSERT(i == p->pm_count, ("Invalid port count"));
1024                 speed = lacp_aggregator_bandwidth(la);
1025         }
1026         sc->sc_ifp->if_baudrate = speed;
1027
1028         /* switch the active portmap over */
1029         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1030         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1031                     lsc->lsc_activemap,
1032                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1033 }
1034
1035 static uint16_t
1036 lacp_compose_key(struct lacp_port *lp)
1037 {
1038         struct lagg_port *lgp = lp->lp_lagg;
1039         struct lagg_softc *sc = lgp->lp_softc;
1040         u_int media = lp->lp_media;
1041         uint16_t key;
1042
1043         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1044
1045                 /*
1046                  * non-aggregatable links should have unique keys.
1047                  *
1048                  * XXX this isn't really unique as if_index is 16 bit.
1049                  */
1050
1051                 /* bit 0..14:   (some bits of) if_index of this port */
1052                 key = lp->lp_ifp->if_index;
1053                 /* bit 15:      1 */
1054                 key |= 0x8000;
1055         } else {
1056                 u_int subtype = IFM_SUBTYPE(media);
1057
1058                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1059                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1060
1061                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1062                 switch (subtype) {
1063                 case IFM_10_T:
1064                 case IFM_10_2:
1065                 case IFM_10_5:
1066                 case IFM_10_STP:
1067                 case IFM_10_FL:
1068                         key = IFM_10_T;
1069                         break;
1070                 case IFM_100_TX:
1071                 case IFM_100_FX:
1072                 case IFM_100_T4:
1073                 case IFM_100_VG:
1074                 case IFM_100_T2:
1075                 case IFM_100_T:
1076                         key = IFM_100_TX;
1077                         break;
1078                 case IFM_1000_SX:
1079                 case IFM_1000_LX:
1080                 case IFM_1000_CX:
1081                 case IFM_1000_T:
1082                 case IFM_1000_KX:
1083                 case IFM_1000_SGMII:
1084                 case IFM_1000_CX_SGMII:
1085                         key = IFM_1000_SX;
1086                         break;
1087                 case IFM_10G_LR:
1088                 case IFM_10G_SR:
1089                 case IFM_10G_CX4:
1090                 case IFM_10G_TWINAX:
1091                 case IFM_10G_TWINAX_LONG:
1092                 case IFM_10G_LRM:
1093                 case IFM_10G_T:
1094                 case IFM_10G_KX4:
1095                 case IFM_10G_KR:
1096                 case IFM_10G_CR1:
1097                 case IFM_10G_ER:
1098                 case IFM_10G_SFI:
1099                         key = IFM_10G_LR;
1100                         break;
1101                 case IFM_20G_KR2:
1102                         key = IFM_20G_KR2;
1103                         break;
1104                 case IFM_2500_KX:
1105                 case IFM_2500_T:
1106                         key = IFM_2500_KX;
1107                         break;
1108                 case IFM_5000_T:
1109                         key = IFM_5000_T;
1110                         break;
1111                 case IFM_50G_PCIE:
1112                 case IFM_50G_CR2:
1113                 case IFM_50G_KR2:
1114                         key = IFM_50G_PCIE;
1115                         break;
1116                 case IFM_56G_R4:
1117                         key = IFM_56G_R4;
1118                         break;
1119                 case IFM_25G_PCIE:
1120                 case IFM_25G_CR:
1121                 case IFM_25G_KR:
1122                 case IFM_25G_SR:
1123                         key = IFM_25G_PCIE;
1124                         break;
1125                 case IFM_40G_CR4:
1126                 case IFM_40G_SR4:
1127                 case IFM_40G_LR4:
1128                 case IFM_40G_XLPPI:
1129                 case IFM_40G_KR4:
1130                         key = IFM_40G_CR4;
1131                         break;
1132                 case IFM_100G_CR4:
1133                 case IFM_100G_SR4:
1134                 case IFM_100G_KR4:
1135                 case IFM_100G_LR4:
1136                         key = IFM_100G_CR4;
1137                         break;
1138                 default:
1139                         key = subtype;
1140                         break;
1141                 }
1142                 /* bit 5..14:   (some bits of) if_index of lagg device */
1143                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1144                 /* bit 15:      0 */
1145         }
1146         return (htons(key));
1147 }
1148
1149 static void
1150 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1151 {
1152         char buf[LACP_LAGIDSTR_MAX+1];
1153
1154         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1155             __func__,
1156             lacp_format_lagid(&la->la_actor, &la->la_partner,
1157             buf, sizeof(buf)),
1158             la->la_refcnt, la->la_refcnt + 1));
1159
1160         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1161         la->la_refcnt++;
1162         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1163 }
1164
1165 static void
1166 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1167 {
1168         char buf[LACP_LAGIDSTR_MAX+1];
1169
1170         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1171             __func__,
1172             lacp_format_lagid(&la->la_actor, &la->la_partner,
1173             buf, sizeof(buf)),
1174             la->la_refcnt, la->la_refcnt - 1));
1175
1176         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1177         la->la_refcnt--;
1178         if (la->la_refcnt > 0) {
1179                 return;
1180         }
1181
1182         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1183         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1184
1185         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1186
1187         free(la, M_DEVBUF);
1188 }
1189
1190 /*
1191  * lacp_aggregator_get: allocate an aggregator.
1192  */
1193
1194 static struct lacp_aggregator *
1195 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1196 {
1197         struct lacp_aggregator *la;
1198
1199         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1200         if (la) {
1201                 la->la_refcnt = 1;
1202                 la->la_nports = 0;
1203                 TAILQ_INIT(&la->la_ports);
1204                 la->la_pending = 0;
1205                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1206         }
1207
1208         return (la);
1209 }
1210
1211 /*
1212  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1213  */
1214
1215 static void
1216 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1217 {
1218         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1219         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1220
1221         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1222 }
1223
1224 static void
1225 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1226     const struct lacp_peerinfo *lpi_port)
1227 {
1228         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1229         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1230         lpi_aggr->lip_key = lpi_port->lip_key;
1231 }
1232
1233 /*
1234  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1235  */
1236
1237 static int
1238 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1239     const struct lacp_port *lp)
1240 {
1241         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1242             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1243                 return (0);
1244         }
1245
1246         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1247                 return (0);
1248         }
1249
1250         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1251                 return (0);
1252         }
1253
1254         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1255                 return (0);
1256         }
1257
1258         return (1);
1259 }
1260
1261 static int
1262 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1263     const struct lacp_peerinfo *b)
1264 {
1265         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1266             sizeof(a->lip_systemid))) {
1267                 return (0);
1268         }
1269
1270         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1271                 return (0);
1272         }
1273
1274         return (1);
1275 }
1276
1277 static void
1278 lacp_port_enable(struct lacp_port *lp)
1279 {
1280         lp->lp_state |= LACP_STATE_AGGREGATION;
1281 }
1282
1283 static void
1284 lacp_port_disable(struct lacp_port *lp)
1285 {
1286         lacp_set_mux(lp, LACP_MUX_DETACHED);
1287
1288         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1289         lp->lp_selected = LACP_UNSELECTED;
1290         lacp_sm_rx_record_default(lp);
1291         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1292         lp->lp_state &= ~LACP_STATE_EXPIRED;
1293 }
1294
1295 /*
1296  * lacp_select: select an aggregator.  create one if necessary.
1297  */
1298 static void
1299 lacp_select(struct lacp_port *lp)
1300 {
1301         struct lacp_softc *lsc = lp->lp_lsc;
1302         struct lacp_aggregator *la;
1303         char buf[LACP_LAGIDSTR_MAX+1];
1304
1305         if (lp->lp_aggregator) {
1306                 return;
1307         }
1308
1309         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1310             ("timer_wait_while still active"));
1311
1312         LACP_DPRINTF((lp, "port lagid=%s\n",
1313             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1314             buf, sizeof(buf))));
1315
1316         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1317                 if (lacp_aggregator_is_compatible(la, lp)) {
1318                         break;
1319                 }
1320         }
1321
1322         if (la == NULL) {
1323                 la = lacp_aggregator_get(lsc, lp);
1324                 if (la == NULL) {
1325                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1326
1327                         /*
1328                          * will retry on the next tick.
1329                          */
1330
1331                         return;
1332                 }
1333                 lacp_fill_aggregator_id(la, lp);
1334                 LACP_DPRINTF((lp, "aggregator created\n"));
1335         } else {
1336                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1337                 if (la->la_refcnt == LACP_MAX_PORTS)
1338                         return;
1339                 lacp_aggregator_addref(lsc, la);
1340         }
1341
1342         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1343             lacp_format_lagid(&la->la_actor, &la->la_partner,
1344             buf, sizeof(buf))));
1345
1346         lp->lp_aggregator = la;
1347         lp->lp_selected = LACP_SELECTED;
1348 }
1349
1350 /*
1351  * lacp_unselect: finish unselect/detach process.
1352  */
1353
1354 static void
1355 lacp_unselect(struct lacp_port *lp)
1356 {
1357         struct lacp_softc *lsc = lp->lp_lsc;
1358         struct lacp_aggregator *la = lp->lp_aggregator;
1359
1360         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1361             ("timer_wait_while still active"));
1362
1363         if (la == NULL) {
1364                 return;
1365         }
1366
1367         lp->lp_aggregator = NULL;
1368         lacp_aggregator_delref(lsc, la);
1369 }
1370
1371 /* mux machine */
1372
1373 static void
1374 lacp_sm_mux(struct lacp_port *lp)
1375 {
1376         struct lagg_port *lgp = lp->lp_lagg;
1377         struct lagg_softc *sc = lgp->lp_softc;
1378         enum lacp_mux_state new_state;
1379         boolean_t p_sync =
1380                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1381         boolean_t p_collecting =
1382             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1383         enum lacp_selected selected = lp->lp_selected;
1384         struct lacp_aggregator *la;
1385
1386         if (V_lacp_debug > 1)
1387                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1388                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1389                     lp->lp_mux_state, selected, p_sync, p_collecting);
1390
1391 re_eval:
1392         la = lp->lp_aggregator;
1393         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1394             ("MUX not detached"));
1395         new_state = lp->lp_mux_state;
1396         switch (lp->lp_mux_state) {
1397         case LACP_MUX_DETACHED:
1398                 if (selected != LACP_UNSELECTED) {
1399                         new_state = LACP_MUX_WAITING;
1400                 }
1401                 break;
1402         case LACP_MUX_WAITING:
1403                 KASSERT(la->la_pending > 0 ||
1404                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1405                     ("timer_wait_while still active"));
1406                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1407                         new_state = LACP_MUX_ATTACHED;
1408                 } else if (selected == LACP_UNSELECTED) {
1409                         new_state = LACP_MUX_DETACHED;
1410                 }
1411                 break;
1412         case LACP_MUX_ATTACHED:
1413                 if (selected == LACP_SELECTED && p_sync) {
1414                         new_state = LACP_MUX_COLLECTING;
1415                 } else if (selected != LACP_SELECTED) {
1416                         new_state = LACP_MUX_DETACHED;
1417                 }
1418                 break;
1419         case LACP_MUX_COLLECTING:
1420                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1421                         new_state = LACP_MUX_DISTRIBUTING;
1422                 } else if (selected != LACP_SELECTED || !p_sync) {
1423                         new_state = LACP_MUX_ATTACHED;
1424                 }
1425                 break;
1426         case LACP_MUX_DISTRIBUTING:
1427                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1428                         new_state = LACP_MUX_COLLECTING;
1429                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1430                         sc->sc_flapping++;
1431                 }
1432                 break;
1433         default:
1434                 panic("%s: unknown state", __func__);
1435         }
1436
1437         if (lp->lp_mux_state == new_state) {
1438                 return;
1439         }
1440
1441         lacp_set_mux(lp, new_state);
1442         goto re_eval;
1443 }
1444
1445 static void
1446 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1447 {
1448         struct lacp_aggregator *la = lp->lp_aggregator;
1449
1450         if (lp->lp_mux_state == new_state) {
1451                 return;
1452         }
1453
1454         switch (new_state) {
1455         case LACP_MUX_DETACHED:
1456                 lp->lp_state &= ~LACP_STATE_SYNC;
1457                 lacp_disable_distributing(lp);
1458                 lacp_disable_collecting(lp);
1459                 lacp_sm_assert_ntt(lp);
1460                 /* cancel timer */
1461                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1462                         KASSERT(la->la_pending > 0,
1463                             ("timer_wait_while not active"));
1464                         la->la_pending--;
1465                 }
1466                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1467                 lacp_unselect(lp);
1468                 break;
1469         case LACP_MUX_WAITING:
1470                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1471                     LACP_AGGREGATE_WAIT_TIME);
1472                 la->la_pending++;
1473                 break;
1474         case LACP_MUX_ATTACHED:
1475                 lp->lp_state |= LACP_STATE_SYNC;
1476                 lacp_disable_collecting(lp);
1477                 lacp_sm_assert_ntt(lp);
1478                 break;
1479         case LACP_MUX_COLLECTING:
1480                 lacp_enable_collecting(lp);
1481                 lacp_disable_distributing(lp);
1482                 lacp_sm_assert_ntt(lp);
1483                 break;
1484         case LACP_MUX_DISTRIBUTING:
1485                 lacp_enable_distributing(lp);
1486                 break;
1487         default:
1488                 panic("%s: unknown state", __func__);
1489         }
1490
1491         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1492
1493         lp->lp_mux_state = new_state;
1494 }
1495
1496 static void
1497 lacp_sm_mux_timer(struct lacp_port *lp)
1498 {
1499         struct lacp_aggregator *la = lp->lp_aggregator;
1500         char buf[LACP_LAGIDSTR_MAX+1];
1501
1502         KASSERT(la->la_pending > 0, ("no pending event"));
1503
1504         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1505             lacp_format_lagid(&la->la_actor, &la->la_partner,
1506             buf, sizeof(buf)),
1507             la->la_pending, la->la_pending - 1));
1508
1509         la->la_pending--;
1510 }
1511
1512 /* periodic transmit machine */
1513
1514 static void
1515 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1516 {
1517         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1518             LACP_STATE_TIMEOUT)) {
1519                 return;
1520         }
1521
1522         LACP_DPRINTF((lp, "partner timeout changed\n"));
1523
1524         /*
1525          * FAST_PERIODIC -> SLOW_PERIODIC
1526          * or
1527          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1528          *
1529          * let lacp_sm_ptx_tx_schedule to update timeout.
1530          */
1531
1532         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1533
1534         /*
1535          * if timeout has been shortened, assert NTT.
1536          */
1537
1538         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1539                 lacp_sm_assert_ntt(lp);
1540         }
1541 }
1542
1543 static void
1544 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1545 {
1546         int timeout;
1547
1548         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1549             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1550
1551                 /*
1552                  * NO_PERIODIC
1553                  */
1554
1555                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1556                 return;
1557         }
1558
1559         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1560                 return;
1561         }
1562
1563         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1564             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1565
1566         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1567 }
1568
1569 static void
1570 lacp_sm_ptx_timer(struct lacp_port *lp)
1571 {
1572         lacp_sm_assert_ntt(lp);
1573 }
1574
1575 static void
1576 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1577 {
1578         int timeout;
1579
1580         /*
1581          * check LACP_DISABLED first
1582          */
1583
1584         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1585                 return;
1586         }
1587
1588         /*
1589          * check loopback condition.
1590          */
1591
1592         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1593             &lp->lp_actor.lip_systemid)) {
1594                 return;
1595         }
1596
1597         /*
1598          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1599          */
1600
1601         lacp_sm_rx_update_selected(lp, du);
1602         lacp_sm_rx_update_ntt(lp, du);
1603         lacp_sm_rx_record_pdu(lp, du);
1604
1605         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1606             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1607         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1608
1609         lp->lp_state &= ~LACP_STATE_EXPIRED;
1610
1611         /*
1612          * kick transmit machine without waiting the next tick.
1613          */
1614
1615         lacp_sm_tx(lp);
1616 }
1617
1618 static void
1619 lacp_sm_rx_set_expired(struct lacp_port *lp)
1620 {
1621         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1622         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1623         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1624         lp->lp_state |= LACP_STATE_EXPIRED;
1625 }
1626
1627 static void
1628 lacp_sm_rx_timer(struct lacp_port *lp)
1629 {
1630         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1631                 /* CURRENT -> EXPIRED */
1632                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1633                 lacp_sm_rx_set_expired(lp);
1634         } else {
1635                 /* EXPIRED -> DEFAULTED */
1636                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1637                 lacp_sm_rx_update_default_selected(lp);
1638                 lacp_sm_rx_record_default(lp);
1639                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1640         }
1641 }
1642
1643 static void
1644 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1645 {
1646         boolean_t active;
1647         uint8_t oldpstate;
1648         char buf[LACP_STATESTR_MAX+1];
1649
1650         LACP_TRACE(lp);
1651
1652         oldpstate = lp->lp_partner.lip_state;
1653
1654         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1655             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1656             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1657
1658         lp->lp_partner = du->ldu_actor;
1659         if (active &&
1660             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1661             LACP_STATE_AGGREGATION) &&
1662             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1663             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1664                 /* XXX nothing? */
1665         } else {
1666                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1667         }
1668
1669         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1670
1671         if (oldpstate != lp->lp_partner.lip_state) {
1672                 LACP_DPRINTF((lp, "old pstate %s\n",
1673                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1674                 LACP_DPRINTF((lp, "new pstate %s\n",
1675                     lacp_format_state(lp->lp_partner.lip_state, buf,
1676                     sizeof(buf))));
1677         }
1678
1679         /* XXX Hack, still need to implement 5.4.9 para 2,3,4 */
1680         if (lp->lp_lsc->lsc_strict_mode)
1681                 lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1682
1683         lacp_sm_ptx_update_timeout(lp, oldpstate);
1684 }
1685
1686 static void
1687 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1688 {
1689
1690         LACP_TRACE(lp);
1691
1692         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1693             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1694             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1695                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1696                 lacp_sm_assert_ntt(lp);
1697         }
1698 }
1699
1700 static void
1701 lacp_sm_rx_record_default(struct lacp_port *lp)
1702 {
1703         uint8_t oldpstate;
1704
1705         LACP_TRACE(lp);
1706
1707         oldpstate = lp->lp_partner.lip_state;
1708         if (lp->lp_lsc->lsc_strict_mode)
1709                 lp->lp_partner = lacp_partner_admin_strict;
1710         else
1711                 lp->lp_partner = lacp_partner_admin_optimistic;
1712         lp->lp_state |= LACP_STATE_DEFAULTED;
1713         lacp_sm_ptx_update_timeout(lp, oldpstate);
1714 }
1715
1716 static void
1717 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1718     const struct lacp_peerinfo *info)
1719 {
1720
1721         LACP_TRACE(lp);
1722
1723         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1724             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1725             LACP_STATE_AGGREGATION)) {
1726                 lp->lp_selected = LACP_UNSELECTED;
1727                 /* mux machine will clean up lp->lp_aggregator */
1728         }
1729 }
1730
1731 static void
1732 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1733 {
1734
1735         LACP_TRACE(lp);
1736
1737         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1738 }
1739
1740 static void
1741 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1742 {
1743
1744         LACP_TRACE(lp);
1745
1746         if (lp->lp_lsc->lsc_strict_mode)
1747                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1748                     &lacp_partner_admin_strict);
1749         else
1750                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1751                     &lacp_partner_admin_optimistic);
1752 }
1753
1754 /* transmit machine */
1755
1756 static void
1757 lacp_sm_tx(struct lacp_port *lp)
1758 {
1759         int error = 0;
1760
1761         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1762 #if 1
1763             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1764             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1765 #endif
1766             ) {
1767                 lp->lp_flags &= ~LACP_PORT_NTT;
1768         }
1769
1770         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1771                 return;
1772         }
1773
1774         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1775         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1776                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1777                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1778                 return;
1779         }
1780
1781         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1782                 error = lacp_xmit_lacpdu(lp);
1783         } else {
1784                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1785         }
1786
1787         if (error == 0) {
1788                 lp->lp_flags &= ~LACP_PORT_NTT;
1789         } else {
1790                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1791                     error));
1792         }
1793 }
1794
1795 static void
1796 lacp_sm_assert_ntt(struct lacp_port *lp)
1797 {
1798
1799         lp->lp_flags |= LACP_PORT_NTT;
1800 }
1801
1802 static void
1803 lacp_run_timers(struct lacp_port *lp)
1804 {
1805         int i;
1806
1807         for (i = 0; i < LACP_NTIMER; i++) {
1808                 KASSERT(lp->lp_timer[i] >= 0,
1809                     ("invalid timer value %d", lp->lp_timer[i]));
1810                 if (lp->lp_timer[i] == 0) {
1811                         continue;
1812                 } else if (--lp->lp_timer[i] <= 0) {
1813                         if (lacp_timer_funcs[i]) {
1814                                 (*lacp_timer_funcs[i])(lp);
1815                         }
1816                 }
1817         }
1818 }
1819
1820 int
1821 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1822 {
1823         struct lacp_softc *lsc = lp->lp_lsc;
1824         struct lagg_port *lgp = lp->lp_lagg;
1825         struct lacp_port *lp2;
1826         struct markerdu *mdu;
1827         int error = 0;
1828         int pending = 0;
1829
1830         if (m->m_pkthdr.len != sizeof(*mdu)) {
1831                 goto bad;
1832         }
1833
1834         if ((m->m_flags & M_MCAST) == 0) {
1835                 goto bad;
1836         }
1837
1838         if (m->m_len < sizeof(*mdu)) {
1839                 m = m_pullup(m, sizeof(*mdu));
1840                 if (m == NULL) {
1841                         return (ENOMEM);
1842                 }
1843         }
1844
1845         mdu = mtod(m, struct markerdu *);
1846
1847         if (memcmp(&mdu->mdu_eh.ether_dhost,
1848             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1849                 goto bad;
1850         }
1851
1852         if (mdu->mdu_sph.sph_version != 1) {
1853                 goto bad;
1854         }
1855
1856         switch (mdu->mdu_tlv.tlv_type) {
1857         case MARKER_TYPE_INFO:
1858                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1859                     marker_info_tlv_template, TRUE)) {
1860                         goto bad;
1861                 }
1862                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1863                 memcpy(&mdu->mdu_eh.ether_dhost,
1864                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1865                 memcpy(&mdu->mdu_eh.ether_shost,
1866                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1867                 error = lagg_enqueue(lp->lp_ifp, m);
1868                 break;
1869
1870         case MARKER_TYPE_RESPONSE:
1871                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1872                     marker_response_tlv_template, TRUE)) {
1873                         goto bad;
1874                 }
1875                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1876                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1877                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1878
1879                 /* Verify that it is the last marker we sent out */
1880                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1881                     sizeof(struct lacp_markerinfo)))
1882                         goto bad;
1883
1884                 LACP_LOCK(lsc);
1885                 lp->lp_flags &= ~LACP_PORT_MARK;
1886
1887                 if (lsc->lsc_suppress_distributing) {
1888                         /* Check if any ports are waiting for a response */
1889                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1890                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1891                                         pending = 1;
1892                                         break;
1893                                 }
1894                         }
1895
1896                         if (pending == 0) {
1897                                 /* All interface queues are clear */
1898                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1899                                 lsc->lsc_suppress_distributing = FALSE;
1900                         }
1901                 }
1902                 LACP_UNLOCK(lsc);
1903                 m_freem(m);
1904                 break;
1905
1906         default:
1907                 goto bad;
1908         }
1909
1910         return (error);
1911
1912 bad:
1913         LACP_DPRINTF((lp, "bad marker frame\n"));
1914         m_freem(m);
1915         return (EINVAL);
1916 }
1917
1918 static int
1919 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1920     const struct tlv_template *tmpl, boolean_t check_type)
1921 {
1922         while (/* CONSTCOND */ 1) {
1923                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1924                         return (EINVAL);
1925                 }
1926                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1927                     tlv->tlv_length != tmpl->tmpl_length) {
1928                         return (EINVAL);
1929                 }
1930                 if (tmpl->tmpl_type == 0) {
1931                         break;
1932                 }
1933                 tlv = (const struct tlvhdr *)
1934                     ((const char *)tlv + tlv->tlv_length);
1935                 tmpl++;
1936         }
1937
1938         return (0);
1939 }
1940
1941 /* Debugging */
1942 const char *
1943 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1944 {
1945         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1946             (int)mac[0],
1947             (int)mac[1],
1948             (int)mac[2],
1949             (int)mac[3],
1950             (int)mac[4],
1951             (int)mac[5]);
1952
1953         return (buf);
1954 }
1955
1956 const char *
1957 lacp_format_systemid(const struct lacp_systemid *sysid,
1958     char *buf, size_t buflen)
1959 {
1960         char macbuf[LACP_MACSTR_MAX+1];
1961
1962         snprintf(buf, buflen, "%04X,%s",
1963             ntohs(sysid->lsi_prio),
1964             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1965
1966         return (buf);
1967 }
1968
1969 const char *
1970 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
1971 {
1972         snprintf(buf, buflen, "%04X,%04X",
1973             ntohs(portid->lpi_prio),
1974             ntohs(portid->lpi_portno));
1975
1976         return (buf);
1977 }
1978
1979 const char *
1980 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
1981 {
1982         char sysid[LACP_SYSTEMIDSTR_MAX+1];
1983         char portid[LACP_PORTIDSTR_MAX+1];
1984
1985         snprintf(buf, buflen, "(%s,%04X,%s)",
1986             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
1987             ntohs(peer->lip_key),
1988             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
1989
1990         return (buf);
1991 }
1992
1993 const char *
1994 lacp_format_lagid(const struct lacp_peerinfo *a,
1995     const struct lacp_peerinfo *b, char *buf, size_t buflen)
1996 {
1997         char astr[LACP_PARTNERSTR_MAX+1];
1998         char bstr[LACP_PARTNERSTR_MAX+1];
1999
2000 #if 0
2001         /*
2002          * there's a convention to display small numbered peer
2003          * in the left.
2004          */
2005
2006         if (lacp_compare_peerinfo(a, b) > 0) {
2007                 const struct lacp_peerinfo *t;
2008
2009                 t = a;
2010                 a = b;
2011                 b = t;
2012         }
2013 #endif
2014
2015         snprintf(buf, buflen, "[%s,%s]",
2016             lacp_format_partner(a, astr, sizeof(astr)),
2017             lacp_format_partner(b, bstr, sizeof(bstr)));
2018
2019         return (buf);
2020 }
2021
2022 const char *
2023 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2024     char *buf, size_t buflen)
2025 {
2026         if (la == NULL) {
2027                 return ("(none)");
2028         }
2029
2030         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2031 }
2032
2033 const char *
2034 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2035 {
2036         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2037         return (buf);
2038 }
2039
2040 static void
2041 lacp_dump_lacpdu(const struct lacpdu *du)
2042 {
2043         char buf[LACP_PARTNERSTR_MAX+1];
2044         char buf2[LACP_STATESTR_MAX+1];
2045
2046         printf("actor=%s\n",
2047             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2048         printf("actor.state=%s\n",
2049             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2050         printf("partner=%s\n",
2051             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2052         printf("partner.state=%s\n",
2053             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2054
2055         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2056 }
2057
2058 static void
2059 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2060 {
2061         va_list va;
2062
2063         if (lp) {
2064                 printf("%s: ", lp->lp_ifp->if_xname);
2065         }
2066
2067         va_start(va, fmt);
2068         vprintf(fmt, va);
2069         va_end(va);
2070 }