sys/net/ieee8023ad_lacp.c

   1 /*      $NetBSD: ieee8023ad_lacp.c,v 1.3 2005/12/11 12:24:54 christos Exp $     */
   2
   3 /*-
   4  * Copyright (c)2005 YAMAMOTO Takashi,
   5  * Copyright (c)2008 Andrew Thompson <thompsa@FreeBSD.org>
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 #include <sys/cdefs.h>
  31 __FBSDID("$FreeBSD$");
  32
  33 #include <sys/param.h>
  34 #include <sys/callout.h>
  35 #include <sys/eventhandler.h>
  36 #include <sys/mbuf.h>
  37 #include <sys/systm.h>
  38 #include <sys/malloc.h>
  39 #include <sys/kernel.h> /* hz */
  40 #include <sys/socket.h> /* for net/if.h */
  41 #include <sys/sockio.h>
  42 #include <sys/sysctl.h>
  43 #include <machine/stdarg.h>
  44 #include <sys/lock.h>
  45 #include <sys/rwlock.h>
  46 #include <sys/taskqueue.h>
  47
  48 #include <net/if.h>
  49 #include <net/if_var.h>
  50 #include <net/if_dl.h>
  51 #include <net/ethernet.h>
  52 #include <net/if_media.h>
  53 #include <net/if_types.h>
  54
  55 #include <net/if_lagg.h>
  56 #include <net/ieee8023ad_lacp.h>
  57
  58 /*
  59  * actor system priority and port priority.
  60  * XXX should be configurable.
  61  */
  62
  63 #define LACP_SYSTEM_PRIO        0x8000
  64 #define LACP_PORT_PRIO          0x8000
  65
  66 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  67     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  68
  69 static const struct tlv_template lacp_info_tlv_template[] = {
  70         { LACP_TYPE_ACTORINFO,
  71             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  72         { LACP_TYPE_PARTNERINFO,
  73             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  74         { LACP_TYPE_COLLECTORINFO,
  75             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  76         { 0, 0 },
  77 };
  78
  79 static const struct tlv_template marker_info_tlv_template[] = {
  80         { MARKER_TYPE_INFO,
  81             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  82         { 0, 0 },
  83 };
  84
  85 static const struct tlv_template marker_response_tlv_template[] = {
  86         { MARKER_TYPE_RESPONSE,
  87             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  88         { 0, 0 },
  89 };
  90
  91 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  92
  93 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
  94 static void     lacp_fill_markerinfo(struct lacp_port *,
  95                     struct lacp_markerinfo *);
  96
  97 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
  98 static void     lacp_suppress_distributing(struct lacp_softc *,
  99                     struct lacp_aggregator *);
 100 static void     lacp_transit_expire(void *);
 101 static void     lacp_update_portmap(struct lacp_softc *);
 102 static void     lacp_select_active_aggregator(struct lacp_softc *);
 103 static uint16_t lacp_compose_key(struct lacp_port *);
 104 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 105                     const struct tlv_template *, boolean_t);
 106 static void     lacp_tick(void *);
 107
 108 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 109                     const struct lacp_port *);
 110 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 111                     const struct lacp_peerinfo *);
 112 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 113                     const struct lacp_port *);
 114 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 115                     const struct lacp_peerinfo *);
 116
 117 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 118                     struct lacp_port *);
 119 static void     lacp_aggregator_addref(struct lacp_softc *,
 120                     struct lacp_aggregator *);
 121 static void     lacp_aggregator_delref(struct lacp_softc *,
 122                     struct lacp_aggregator *);
 123
 124 /* receive machine */
 125
 126 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 127 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 128 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 129 static void     lacp_sm_rx_timer(struct lacp_port *);
 130 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 131 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 132                     const struct lacpdu *);
 133 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 134                     const struct lacpdu *);
 135 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 136                     const struct lacpdu *);
 137 static void     lacp_sm_rx_record_default(struct lacp_port *);
 138 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 139 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 140                     const struct lacp_peerinfo *);
 141
 142 /* mux machine */
 143
 144 static void     lacp_sm_mux(struct lacp_port *);
 145 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 146 static void     lacp_sm_mux_timer(struct lacp_port *);
 147
 148 /* periodic transmit machine */
 149
 150 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 151 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 152 static void     lacp_sm_ptx_timer(struct lacp_port *);
 153
 154 /* transmit machine */
 155
 156 static void     lacp_sm_tx(struct lacp_port *);
 157 static void     lacp_sm_assert_ntt(struct lacp_port *);
 158
 159 static void     lacp_run_timers(struct lacp_port *);
 160 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 161                     const struct lacp_peerinfo *);
 162 static int      lacp_compare_systemid(const struct lacp_systemid *,
 163                     const struct lacp_systemid *);
 164 static void     lacp_port_enable(struct lacp_port *);
 165 static void     lacp_port_disable(struct lacp_port *);
 166 static void     lacp_select(struct lacp_port *);
 167 static void     lacp_unselect(struct lacp_port *);
 168 static void     lacp_disable_collecting(struct lacp_port *);
 169 static void     lacp_enable_collecting(struct lacp_port *);
 170 static void     lacp_disable_distributing(struct lacp_port *);
 171 static void     lacp_enable_distributing(struct lacp_port *);
 172 static int      lacp_xmit_lacpdu(struct lacp_port *);
 173 static int      lacp_xmit_marker(struct lacp_port *);
 174
 175 /* Debugging */
 176
 177 static void     lacp_dump_lacpdu(const struct lacpdu *);
 178 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 179                     size_t);
 180 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 181                     const struct lacp_peerinfo *, char *, size_t);
 182 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 183                     char *, size_t);
 184 static const char *lacp_format_state(uint8_t, char *, size_t);
 185 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 186 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 187                     size_t);
 188 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 189                     size_t);
 190 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 191                     __attribute__((__format__(__printf__, 2, 3)));
 192
 193 static VNET_DEFINE(int, lacp_debug);
 194 #define V_lacp_debug    VNET(lacp_debug)
 195 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
 196 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 197     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 198
 199 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 200 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 201 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 202
 203 /*
 204  * partner administration variables.
 205  * XXX should be configurable.
 206  */
 207
 208 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 209         .lip_systemid = { .lsi_prio = 0xffff },
 210         .lip_portid = { .lpi_prio = 0xffff },
 211         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 212             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 213 };
 214
 215 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 216         .lip_systemid = { .lsi_prio = 0xffff },
 217         .lip_portid = { .lpi_prio = 0xffff },
 218         .lip_state = 0,
 219 };
 220
 221 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 222         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 223         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 224         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 225 };
 226
 227 struct mbuf *
 228 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 229 {
 230         struct lacp_port *lp = LACP_PORT(lgp);
 231         uint8_t subtype;
 232
 233         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 234                 m_freem(m);
 235                 return (NULL);
 236         }
 237
 238         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 239         switch (subtype) {
 240                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 241                         lacp_pdu_input(lp, m);
 242                         return (NULL);
 243
 244                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 245                         lacp_marker_input(lp, m);
 246                         return (NULL);
 247         }
 248
 249         /* Not a subtype we are interested in */
 250         return (m);
 251 }
 252
 253 /*
 254  * lacp_pdu_input: process lacpdu
 255  */
 256 static int
 257 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 258 {
 259         struct lacp_softc *lsc = lp->lp_lsc;
 260         struct lacpdu *du;
 261         int error = 0;
 262
 263         if (m->m_pkthdr.len != sizeof(*du)) {
 264                 goto bad;
 265         }
 266
 267         if ((m->m_flags & M_MCAST) == 0) {
 268                 goto bad;
 269         }
 270
 271         if (m->m_len < sizeof(*du)) {
 272                 m = m_pullup(m, sizeof(*du));
 273                 if (m == NULL) {
 274                         return (ENOMEM);
 275                 }
 276         }
 277
 278         du = mtod(m, struct lacpdu *);
 279
 280         if (memcmp(&du->ldu_eh.ether_dhost,
 281             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 282                 goto bad;
 283         }
 284
 285         /*
 286          * ignore the version for compatibility with
 287          * the future protocol revisions.
 288          */
 289 #if 0
 290         if (du->ldu_sph.sph_version != 1) {
 291                 goto bad;
 292         }
 293 #endif
 294
 295         /*
 296          * ignore tlv types for compatibility with
 297          * the future protocol revisions.
 298          */
 299         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 300             lacp_info_tlv_template, FALSE)) {
 301                 goto bad;
 302         }
 303
 304         if (V_lacp_debug > 0) {
 305                 lacp_dprintf(lp, "lacpdu receive\n");
 306                 lacp_dump_lacpdu(du);
 307         }
 308
 309         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 310                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 311                 goto bad;
 312         }
 313
 314         LACP_LOCK(lsc);
 315         lacp_sm_rx(lp, du);
 316         LACP_UNLOCK(lsc);
 317
 318         m_freem(m);
 319         return (error);
 320
 321 bad:
 322         m_freem(m);
 323         return (EINVAL);
 324 }
 325
 326 static void
 327 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 328 {
 329         struct lagg_port *lgp = lp->lp_lagg;
 330         struct lagg_softc *sc = lgp->lp_softc;
 331
 332         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 333         memcpy(&info->lip_systemid.lsi_mac,
 334             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 335         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 336         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 337         info->lip_state = lp->lp_state;
 338 }
 339
 340 static void
 341 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 342 {
 343         struct ifnet *ifp = lp->lp_ifp;
 344
 345         /* Fill in the port index and system id (encoded as the MAC) */
 346         info->mi_rq_port = htons(ifp->if_index);
 347         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 348         info->mi_rq_xid = htonl(0);
 349 }
 350
 351 static int
 352 lacp_xmit_lacpdu(struct lacp_port *lp)
 353 {
 354         struct lagg_port *lgp = lp->lp_lagg;
 355         struct mbuf *m;
 356         struct lacpdu *du;
 357         int error;
 358
 359         LACP_LOCK_ASSERT(lp->lp_lsc);
 360
 361         m = m_gethdr(M_NOWAIT, MT_DATA);
 362         if (m == NULL) {
 363                 return (ENOMEM);
 364         }
 365         m->m_len = m->m_pkthdr.len = sizeof(*du);
 366
 367         du = mtod(m, struct lacpdu *);
 368         memset(du, 0, sizeof(*du));
 369
 370         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 371             ETHER_ADDR_LEN);
 372         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 373         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 374
 375         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 376         du->ldu_sph.sph_version = 1;
 377
 378         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 379         du->ldu_actor = lp->lp_actor;
 380
 381         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 382             sizeof(du->ldu_partner));
 383         du->ldu_partner = lp->lp_partner;
 384
 385         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 386             sizeof(du->ldu_collector));
 387         du->ldu_collector.lci_maxdelay = 0;
 388
 389         if (V_lacp_debug > 0) {
 390                 lacp_dprintf(lp, "lacpdu transmit\n");
 391                 lacp_dump_lacpdu(du);
 392         }
 393
 394         m->m_flags |= M_MCAST;
 395
 396         /*
 397          * XXX should use higher priority queue.
 398          * otherwise network congestion can break aggregation.
 399          */
 400
 401         error = lagg_enqueue(lp->lp_ifp, m);
 402         return (error);
 403 }
 404
 405 static int
 406 lacp_xmit_marker(struct lacp_port *lp)
 407 {
 408         struct lagg_port *lgp = lp->lp_lagg;
 409         struct mbuf *m;
 410         struct markerdu *mdu;
 411         int error;
 412
 413         LACP_LOCK_ASSERT(lp->lp_lsc);
 414
 415         m = m_gethdr(M_NOWAIT, MT_DATA);
 416         if (m == NULL) {
 417                 return (ENOMEM);
 418         }
 419         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 420
 421         mdu = mtod(m, struct markerdu *);
 422         memset(mdu, 0, sizeof(*mdu));
 423
 424         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 425             ETHER_ADDR_LEN);
 426         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 427         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 428
 429         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 430         mdu->mdu_sph.sph_version = 1;
 431
 432         /* Bump the transaction id and copy over the marker info */
 433         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 434         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 435         mdu->mdu_info = lp->lp_marker;
 436
 437         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 438             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 439             ntohl(mdu->mdu_info.mi_rq_xid)));
 440
 441         m->m_flags |= M_MCAST;
 442         error = lagg_enqueue(lp->lp_ifp, m);
 443         return (error);
 444 }
 445
 446 void
 447 lacp_linkstate(struct lagg_port *lgp)
 448 {
 449         struct lacp_port *lp = LACP_PORT(lgp);
 450         struct lacp_softc *lsc = lp->lp_lsc;
 451         struct ifnet *ifp = lgp->lp_ifp;
 452         struct ifmediareq ifmr;
 453         int error = 0;
 454         u_int media;
 455         uint8_t old_state;
 456         uint16_t old_key;
 457
 458         bzero((char *)&ifmr, sizeof(ifmr));
 459         error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 460         if (error != 0)
 461                 return;
 462
 463         LACP_LOCK(lsc);
 464         media = ifmr.ifm_active;
 465         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 466             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 467             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 468         old_state = lp->lp_state;
 469         old_key = lp->lp_key;
 470
 471         lp->lp_media = media;
 472         /*
 473          * If the port is not an active full duplex Ethernet link then it can
 474          * not be aggregated.
 475          */
 476         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 477             ifp->if_link_state != LINK_STATE_UP) {
 478                 lacp_port_disable(lp);
 479         } else {
 480                 lacp_port_enable(lp);
 481         }
 482         lp->lp_key = lacp_compose_key(lp);
 483
 484         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 485                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 486                 lp->lp_selected = LACP_UNSELECTED;
 487         }
 488         LACP_UNLOCK(lsc);
 489 }
 490
 491 static void
 492 lacp_tick(void *arg)
 493 {
 494         struct lacp_softc *lsc = arg;
 495         struct lacp_port *lp;
 496
 497         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 498                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 499                         continue;
 500
 501                 CURVNET_SET(lp->lp_ifp->if_vnet);
 502                 lacp_run_timers(lp);
 503
 504                 lacp_select(lp);
 505                 lacp_sm_mux(lp);
 506                 lacp_sm_tx(lp);
 507                 lacp_sm_ptx_tx_schedule(lp);
 508                 CURVNET_RESTORE();
 509         }
 510         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 511 }
 512
 513 int
 514 lacp_port_create(struct lagg_port *lgp)
 515 {
 516         struct lagg_softc *sc = lgp->lp_softc;
 517         struct lacp_softc *lsc = LACP_SOFTC(sc);
 518         struct lacp_port *lp;
 519         struct ifnet *ifp = lgp->lp_ifp;
 520         struct sockaddr_dl sdl;
 521         struct ifmultiaddr *rifma = NULL;
 522         int error;
 523
 524         boolean_t active = TRUE; /* XXX should be configurable */
 525         boolean_t fast = FALSE; /* Configurable via ioctl */
 526
 527         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 528         sdl.sdl_alen = ETHER_ADDR_LEN;
 529
 530         bcopy(&ethermulticastaddr_slowprotocols,
 531             LLADDR(&sdl), ETHER_ADDR_LEN);
 532         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 533         if (error) {
 534                 printf("%s: ADDMULTI failed on %s\n", __func__,
 535                     lgp->lp_ifp->if_xname);
 536                 return (error);
 537         }
 538
 539         lp = malloc(sizeof(struct lacp_port),
 540             M_DEVBUF, M_NOWAIT|M_ZERO);
 541         if (lp == NULL)
 542                 return (ENOMEM);
 543
 544         LACP_LOCK(lsc);
 545         lgp->lp_psc = lp;
 546         lp->lp_ifp = ifp;
 547         lp->lp_lagg = lgp;
 548         lp->lp_lsc = lsc;
 549         lp->lp_ifma = rifma;
 550
 551         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 552
 553         lacp_fill_actorinfo(lp, &lp->lp_actor);
 554         lacp_fill_markerinfo(lp, &lp->lp_marker);
 555         lp->lp_state =
 556             (active ? LACP_STATE_ACTIVITY : 0) |
 557             (fast ? LACP_STATE_TIMEOUT : 0);
 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, void *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, void *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 = lsc;
 764         lsc->lsc_softc = sc;
 765
 766         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 767         lsc->lsc_active_aggregator = NULL;
 768         lsc->lsc_strict_mode = 1;
 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 void
 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 }
 797
 798 void
 799 lacp_init(struct lagg_softc *sc)
 800 {
 801         struct lacp_softc *lsc = LACP_SOFTC(sc);
 802
 803         LACP_LOCK(lsc);
 804         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 805         LACP_UNLOCK(lsc);
 806 }
 807
 808 void
 809 lacp_stop(struct lagg_softc *sc)
 810 {
 811         struct lacp_softc *lsc = LACP_SOFTC(sc);
 812
 813         LACP_LOCK(lsc);
 814         callout_stop(&lsc->lsc_transit_callout);
 815         callout_stop(&lsc->lsc_callout);
 816         LACP_UNLOCK(lsc);
 817 }
 818
 819 struct lagg_port *
 820 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 821 {
 822         struct lacp_softc *lsc = LACP_SOFTC(sc);
 823         struct lacp_portmap *pm;
 824         struct lacp_port *lp;
 825         uint32_t hash;
 826
 827         if (__predict_false(lsc->lsc_suppress_distributing)) {
 828                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 829                 return (NULL);
 830         }
 831
 832         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 833         if (pm->pm_count == 0) {
 834                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 835                 return (NULL);
 836         }
 837
 838         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 839             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 840                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 841         else
 842                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 843         hash %= pm->pm_count;
 844         lp = pm->pm_map[hash];
 845
 846         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 847             ("aggregated port is not distributing"));
 848
 849         return (lp->lp_lagg);
 850 }
 851 /*
 852  * lacp_suppress_distributing: drop transmit packets for a while
 853  * to preserve packet ordering.
 854  */
 855
 856 static void
 857 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 858 {
 859         struct lacp_port *lp;
 860
 861         if (lsc->lsc_active_aggregator != la) {
 862                 return;
 863         }
 864
 865         LACP_TRACE(NULL);
 866
 867         lsc->lsc_suppress_distributing = TRUE;
 868
 869         /* send a marker frame down each port to verify the queues are empty */
 870         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 871                 lp->lp_flags |= LACP_PORT_MARK;
 872                 lacp_xmit_marker(lp);
 873         }
 874
 875         /* set a timeout for the marker frames */
 876         callout_reset(&lsc->lsc_transit_callout,
 877             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 878 }
 879
 880 static int
 881 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 882     const struct lacp_peerinfo *b)
 883 {
 884         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 885 }
 886
 887 static int
 888 lacp_compare_systemid(const struct lacp_systemid *a,
 889     const struct lacp_systemid *b)
 890 {
 891         return (memcmp(a, b, sizeof(*a)));
 892 }
 893
 894 #if 0   /* unused */
 895 static int
 896 lacp_compare_portid(const struct lacp_portid *a,
 897     const struct lacp_portid *b)
 898 {
 899         return (memcmp(a, b, sizeof(*a)));
 900 }
 901 #endif
 902
 903 static uint64_t
 904 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 905 {
 906         struct lacp_port *lp;
 907         uint64_t speed;
 908
 909         lp = TAILQ_FIRST(&la->la_ports);
 910         if (lp == NULL) {
 911                 return (0);
 912         }
 913
 914         speed = ifmedia_baudrate(lp->lp_media);
 915         speed *= la->la_nports;
 916         if (speed == 0) {
 917                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 918                     lp->lp_media, la->la_nports));
 919         }
 920
 921         return (speed);
 922 }
 923
 924 /*
 925  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 926  * packets from lagg(4) interface.
 927  */
 928
 929 static void
 930 lacp_select_active_aggregator(struct lacp_softc *lsc)
 931 {
 932         struct lacp_aggregator *la;
 933         struct lacp_aggregator *best_la = NULL;
 934         uint64_t best_speed = 0;
 935         char buf[LACP_LAGIDSTR_MAX+1];
 936
 937         LACP_TRACE(NULL);
 938
 939         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 940                 uint64_t speed;
 941
 942                 if (la->la_nports == 0) {
 943                         continue;
 944                 }
 945
 946                 speed = lacp_aggregator_bandwidth(la);
 947                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 948                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 949                     speed, la->la_nports));
 950
 951                 /*
 952                  * This aggregator is chosen if the partner has a better
 953                  * system priority or, the total aggregated speed is higher
 954                  * or, it is already the chosen aggregator
 955                  */
 956                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 957                     LACP_SYS_PRI(best_la->la_partner)) ||
 958                     speed > best_speed ||
 959                     (speed == best_speed &&
 960                     la == lsc->lsc_active_aggregator)) {
 961                         best_la = la;
 962                         best_speed = speed;
 963                 }
 964         }
 965
 966         KASSERT(best_la == NULL || best_la->la_nports > 0,
 967             ("invalid aggregator refcnt"));
 968         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
 969             ("invalid aggregator list"));
 970
 971         if (lsc->lsc_active_aggregator != best_la) {
 972                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
 973                 LACP_DPRINTF((NULL, "old %s\n",
 974                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
 975                     buf, sizeof(buf))));
 976         } else {
 977                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
 978         }
 979         LACP_DPRINTF((NULL, "new %s\n",
 980             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
 981
 982         if (lsc->lsc_active_aggregator != best_la) {
 983                 lsc->lsc_active_aggregator = best_la;
 984                 lacp_update_portmap(lsc);
 985                 if (best_la) {
 986                         lacp_suppress_distributing(lsc, best_la);
 987                 }
 988         }
 989 }
 990
 991 /*
 992  * Updated the inactive portmap array with the new list of ports and
 993  * make it live.
 994  */
 995 static void
 996 lacp_update_portmap(struct lacp_softc *lsc)
 997 {
 998         struct lagg_softc *sc = lsc->lsc_softc;
 999         struct lacp_aggregator *la;
1000         struct lacp_portmap *p;
1001         struct lacp_port *lp;
1002         uint64_t speed;
1003         u_int newmap;
1004         int i;
1005
1006         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1007         p = &lsc->lsc_pmap[newmap];
1008         la = lsc->lsc_active_aggregator;
1009         speed = 0;
1010         bzero(p, sizeof(struct lacp_portmap));
1011
1012         if (la != NULL && la->la_nports > 0) {
1013                 p->pm_count = la->la_nports;
1014                 i = 0;
1015                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1016                         p->pm_map[i++] = lp;
1017                 KASSERT(i == p->pm_count, ("Invalid port count"));
1018                 speed = lacp_aggregator_bandwidth(la);
1019         }
1020         sc->sc_ifp->if_baudrate = speed;
1021
1022         /* switch the active portmap over */
1023         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1024         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1025                     lsc->lsc_activemap,
1026                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1027 }
1028
1029 static uint16_t
1030 lacp_compose_key(struct lacp_port *lp)
1031 {
1032         struct lagg_port *lgp = lp->lp_lagg;
1033         struct lagg_softc *sc = lgp->lp_softc;
1034         u_int media = lp->lp_media;
1035         uint16_t key;
1036
1037         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1038
1039                 /*
1040                  * non-aggregatable links should have unique keys.
1041                  *
1042                  * XXX this isn't really unique as if_index is 16 bit.
1043                  */
1044
1045                 /* bit 0..14:   (some bits of) if_index of this port */
1046                 key = lp->lp_ifp->if_index;
1047                 /* bit 15:      1 */
1048                 key |= 0x8000;
1049         } else {
1050                 u_int subtype = IFM_SUBTYPE(media);
1051
1052                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1053                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1054
1055                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1056                 switch (subtype) {
1057                 case IFM_10_T:
1058                 case IFM_10_2:
1059                 case IFM_10_5:
1060                 case IFM_10_STP:
1061                 case IFM_10_FL:
1062                         key = IFM_10_T;
1063                         break;
1064                 case IFM_100_TX:
1065                 case IFM_100_FX:
1066                 case IFM_100_T4:
1067                 case IFM_100_VG:
1068                 case IFM_100_T2:
1069                 case IFM_100_T:
1070                         key = IFM_100_TX;
1071                         break;
1072                 case IFM_1000_SX:
1073                 case IFM_1000_LX:
1074                 case IFM_1000_CX:
1075                 case IFM_1000_T:
1076                 case IFM_1000_KX:
1077                 case IFM_1000_SGMII:
1078                 case IFM_1000_CX_SGMII:
1079                         key = IFM_1000_SX;
1080                         break;
1081                 case IFM_10G_LR:
1082                 case IFM_10G_SR:
1083                 case IFM_10G_CX4:
1084                 case IFM_10G_TWINAX:
1085                 case IFM_10G_TWINAX_LONG:
1086                 case IFM_10G_LRM:
1087                 case IFM_10G_T:
1088                 case IFM_10G_KX4:
1089                 case IFM_10G_KR:
1090                 case IFM_10G_CR1:
1091                 case IFM_10G_ER:
1092                 case IFM_10G_SFI:
1093                         key = IFM_10G_LR;
1094                         break;
1095                 case IFM_20G_KR2:
1096                         key = IFM_20G_KR2;
1097                         break;
1098                 case IFM_2500_KX:
1099                 case IFM_2500_T:
1100                         key = IFM_2500_KX;
1101                         break;
1102                 case IFM_5000_T:
1103                         key = IFM_5000_T;
1104                         break;
1105                 case IFM_50G_PCIE:
1106                 case IFM_50G_CR2:
1107                 case IFM_50G_KR2:
1108                         key = IFM_50G_PCIE;
1109                         break;
1110                 case IFM_56G_R4:
1111                         key = IFM_56G_R4;
1112                         break;
1113                 case IFM_25G_PCIE:
1114                 case IFM_25G_CR:
1115                 case IFM_25G_KR:
1116                 case IFM_25G_SR:
1117                         key = IFM_25G_PCIE;
1118                         break;
1119                 case IFM_40G_CR4:
1120                 case IFM_40G_SR4:
1121                 case IFM_40G_LR4:
1122                 case IFM_40G_XLPPI:
1123                 case IFM_40G_KR4:
1124                         key = IFM_40G_CR4;
1125                         break;
1126                 case IFM_100G_CR4:
1127                 case IFM_100G_SR4:
1128                 case IFM_100G_KR4:
1129                 case IFM_100G_LR4:
1130                         key = IFM_100G_CR4;
1131                         break;
1132                 default:
1133                         key = subtype;
1134                         break;
1135                 }
1136                 /* bit 5..14:   (some bits of) if_index of lagg device */
1137                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1138                 /* bit 15:      0 */
1139         }
1140         return (htons(key));
1141 }
1142
1143 static void
1144 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1145 {
1146         char buf[LACP_LAGIDSTR_MAX+1];
1147
1148         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1149             __func__,
1150             lacp_format_lagid(&la->la_actor, &la->la_partner,
1151             buf, sizeof(buf)),
1152             la->la_refcnt, la->la_refcnt + 1));
1153
1154         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1155         la->la_refcnt++;
1156         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1157 }
1158
1159 static void
1160 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1161 {
1162         char buf[LACP_LAGIDSTR_MAX+1];
1163
1164         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1165             __func__,
1166             lacp_format_lagid(&la->la_actor, &la->la_partner,
1167             buf, sizeof(buf)),
1168             la->la_refcnt, la->la_refcnt - 1));
1169
1170         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1171         la->la_refcnt--;
1172         if (la->la_refcnt > 0) {
1173                 return;
1174         }
1175
1176         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1177         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1178
1179         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1180
1181         free(la, M_DEVBUF);
1182 }
1183
1184 /*
1185  * lacp_aggregator_get: allocate an aggregator.
1186  */
1187
1188 static struct lacp_aggregator *
1189 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1190 {
1191         struct lacp_aggregator *la;
1192
1193         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1194         if (la) {
1195                 la->la_refcnt = 1;
1196                 la->la_nports = 0;
1197                 TAILQ_INIT(&la->la_ports);
1198                 la->la_pending = 0;
1199                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1200         }
1201
1202         return (la);
1203 }
1204
1205 /*
1206  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1207  */
1208
1209 static void
1210 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1211 {
1212         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1213         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1214
1215         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1216 }
1217
1218 static void
1219 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1220     const struct lacp_peerinfo *lpi_port)
1221 {
1222         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1223         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1224         lpi_aggr->lip_key = lpi_port->lip_key;
1225 }
1226
1227 /*
1228  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1229  */
1230
1231 static int
1232 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1233     const struct lacp_port *lp)
1234 {
1235         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1236             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1237                 return (0);
1238         }
1239
1240         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1241                 return (0);
1242         }
1243
1244         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1245                 return (0);
1246         }
1247
1248         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1249                 return (0);
1250         }
1251
1252         return (1);
1253 }
1254
1255 static int
1256 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1257     const struct lacp_peerinfo *b)
1258 {
1259         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1260             sizeof(a->lip_systemid))) {
1261                 return (0);
1262         }
1263
1264         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1265                 return (0);
1266         }
1267
1268         return (1);
1269 }
1270
1271 static void
1272 lacp_port_enable(struct lacp_port *lp)
1273 {
1274         lp->lp_state |= LACP_STATE_AGGREGATION;
1275 }
1276
1277 static void
1278 lacp_port_disable(struct lacp_port *lp)
1279 {
1280         lacp_set_mux(lp, LACP_MUX_DETACHED);
1281
1282         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1283         lp->lp_selected = LACP_UNSELECTED;
1284         lacp_sm_rx_record_default(lp);
1285         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1286         lp->lp_state &= ~LACP_STATE_EXPIRED;
1287 }
1288
1289 /*
1290  * lacp_select: select an aggregator.  create one if necessary.
1291  */
1292 static void
1293 lacp_select(struct lacp_port *lp)
1294 {
1295         struct lacp_softc *lsc = lp->lp_lsc;
1296         struct lacp_aggregator *la;
1297         char buf[LACP_LAGIDSTR_MAX+1];
1298
1299         if (lp->lp_aggregator) {
1300                 return;
1301         }
1302
1303         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1304             ("timer_wait_while still active"));
1305
1306         LACP_DPRINTF((lp, "port lagid=%s\n",
1307             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1308             buf, sizeof(buf))));
1309
1310         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1311                 if (lacp_aggregator_is_compatible(la, lp)) {
1312                         break;
1313                 }
1314         }
1315
1316         if (la == NULL) {
1317                 la = lacp_aggregator_get(lsc, lp);
1318                 if (la == NULL) {
1319                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1320
1321                         /*
1322                          * will retry on the next tick.
1323                          */
1324
1325                         return;
1326                 }
1327                 lacp_fill_aggregator_id(la, lp);
1328                 LACP_DPRINTF((lp, "aggregator created\n"));
1329         } else {
1330                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1331                 if (la->la_refcnt == LACP_MAX_PORTS)
1332                         return;
1333                 lacp_aggregator_addref(lsc, la);
1334         }
1335
1336         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1337             lacp_format_lagid(&la->la_actor, &la->la_partner,
1338             buf, sizeof(buf))));
1339
1340         lp->lp_aggregator = la;
1341         lp->lp_selected = LACP_SELECTED;
1342 }
1343
1344 /*
1345  * lacp_unselect: finish unselect/detach process.
1346  */
1347
1348 static void
1349 lacp_unselect(struct lacp_port *lp)
1350 {
1351         struct lacp_softc *lsc = lp->lp_lsc;
1352         struct lacp_aggregator *la = lp->lp_aggregator;
1353
1354         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1355             ("timer_wait_while still active"));
1356
1357         if (la == NULL) {
1358                 return;
1359         }
1360
1361         lp->lp_aggregator = NULL;
1362         lacp_aggregator_delref(lsc, la);
1363 }
1364
1365 /* mux machine */
1366
1367 static void
1368 lacp_sm_mux(struct lacp_port *lp)
1369 {
1370         struct lagg_port *lgp = lp->lp_lagg;
1371         struct lagg_softc *sc = lgp->lp_softc;
1372         enum lacp_mux_state new_state;
1373         boolean_t p_sync =
1374                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1375         boolean_t p_collecting =
1376             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1377         enum lacp_selected selected = lp->lp_selected;
1378         struct lacp_aggregator *la;
1379
1380         if (V_lacp_debug > 1)
1381                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1382                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1383                     lp->lp_mux_state, selected, p_sync, p_collecting);
1384
1385 re_eval:
1386         la = lp->lp_aggregator;
1387         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1388             ("MUX not detached"));
1389         new_state = lp->lp_mux_state;
1390         switch (lp->lp_mux_state) {
1391         case LACP_MUX_DETACHED:
1392                 if (selected != LACP_UNSELECTED) {
1393                         new_state = LACP_MUX_WAITING;
1394                 }
1395                 break;
1396         case LACP_MUX_WAITING:
1397                 KASSERT(la->la_pending > 0 ||
1398                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1399                     ("timer_wait_while still active"));
1400                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1401                         new_state = LACP_MUX_ATTACHED;
1402                 } else if (selected == LACP_UNSELECTED) {
1403                         new_state = LACP_MUX_DETACHED;
1404                 }
1405                 break;
1406         case LACP_MUX_ATTACHED:
1407                 if (selected == LACP_SELECTED && p_sync) {
1408                         new_state = LACP_MUX_COLLECTING;
1409                 } else if (selected != LACP_SELECTED) {
1410                         new_state = LACP_MUX_DETACHED;
1411                 }
1412                 break;
1413         case LACP_MUX_COLLECTING:
1414                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1415                         new_state = LACP_MUX_DISTRIBUTING;
1416                 } else if (selected != LACP_SELECTED || !p_sync) {
1417                         new_state = LACP_MUX_ATTACHED;
1418                 }
1419                 break;
1420         case LACP_MUX_DISTRIBUTING:
1421                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1422                         new_state = LACP_MUX_COLLECTING;
1423                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1424                         sc->sc_flapping++;
1425                 }
1426                 break;
1427         default:
1428                 panic("%s: unknown state", __func__);
1429         }
1430
1431         if (lp->lp_mux_state == new_state) {
1432                 return;
1433         }
1434
1435         lacp_set_mux(lp, new_state);
1436         goto re_eval;
1437 }
1438
1439 static void
1440 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1441 {
1442         struct lacp_aggregator *la = lp->lp_aggregator;
1443
1444         if (lp->lp_mux_state == new_state) {
1445                 return;
1446         }
1447
1448         switch (new_state) {
1449         case LACP_MUX_DETACHED:
1450                 lp->lp_state &= ~LACP_STATE_SYNC;
1451                 lacp_disable_distributing(lp);
1452                 lacp_disable_collecting(lp);
1453                 lacp_sm_assert_ntt(lp);
1454                 /* cancel timer */
1455                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1456                         KASSERT(la->la_pending > 0,
1457                             ("timer_wait_while not active"));
1458                         la->la_pending--;
1459                 }
1460                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1461                 lacp_unselect(lp);
1462                 break;
1463         case LACP_MUX_WAITING:
1464                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1465                     LACP_AGGREGATE_WAIT_TIME);
1466                 la->la_pending++;
1467                 break;
1468         case LACP_MUX_ATTACHED:
1469                 lp->lp_state |= LACP_STATE_SYNC;
1470                 lacp_disable_collecting(lp);
1471                 lacp_sm_assert_ntt(lp);
1472                 break;
1473         case LACP_MUX_COLLECTING:
1474                 lacp_enable_collecting(lp);
1475                 lacp_disable_distributing(lp);
1476                 lacp_sm_assert_ntt(lp);
1477                 break;
1478         case LACP_MUX_DISTRIBUTING:
1479                 lacp_enable_distributing(lp);
1480                 break;
1481         default:
1482                 panic("%s: unknown state", __func__);
1483         }
1484
1485         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1486
1487         lp->lp_mux_state = new_state;
1488 }
1489
1490 static void
1491 lacp_sm_mux_timer(struct lacp_port *lp)
1492 {
1493         struct lacp_aggregator *la = lp->lp_aggregator;
1494         char buf[LACP_LAGIDSTR_MAX+1];
1495
1496         KASSERT(la->la_pending > 0, ("no pending event"));
1497
1498         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1499             lacp_format_lagid(&la->la_actor, &la->la_partner,
1500             buf, sizeof(buf)),
1501             la->la_pending, la->la_pending - 1));
1502
1503         la->la_pending--;
1504 }
1505
1506 /* periodic transmit machine */
1507
1508 static void
1509 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1510 {
1511         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1512             LACP_STATE_TIMEOUT)) {
1513                 return;
1514         }
1515
1516         LACP_DPRINTF((lp, "partner timeout changed\n"));
1517
1518         /*
1519          * FAST_PERIODIC -> SLOW_PERIODIC
1520          * or
1521          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1522          *
1523          * let lacp_sm_ptx_tx_schedule to update timeout.
1524          */
1525
1526         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1527
1528         /*
1529          * if timeout has been shortened, assert NTT.
1530          */
1531
1532         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1533                 lacp_sm_assert_ntt(lp);
1534         }
1535 }
1536
1537 static void
1538 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1539 {
1540         int timeout;
1541
1542         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1543             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1544
1545                 /*
1546                  * NO_PERIODIC
1547                  */
1548
1549                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1550                 return;
1551         }
1552
1553         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1554                 return;
1555         }
1556
1557         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1558             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1559
1560         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1561 }
1562
1563 static void
1564 lacp_sm_ptx_timer(struct lacp_port *lp)
1565 {
1566         lacp_sm_assert_ntt(lp);
1567 }
1568
1569 static void
1570 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1571 {
1572         int timeout;
1573
1574         /*
1575          * check LACP_DISABLED first
1576          */
1577
1578         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1579                 return;
1580         }
1581
1582         /*
1583          * check loopback condition.
1584          */
1585
1586         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1587             &lp->lp_actor.lip_systemid)) {
1588                 return;
1589         }
1590
1591         /*
1592          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1593          */
1594
1595         lacp_sm_rx_update_selected(lp, du);
1596         lacp_sm_rx_update_ntt(lp, du);
1597         lacp_sm_rx_record_pdu(lp, du);
1598
1599         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1600             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1601         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1602
1603         lp->lp_state &= ~LACP_STATE_EXPIRED;
1604
1605         /*
1606          * kick transmit machine without waiting the next tick.
1607          */
1608
1609         lacp_sm_tx(lp);
1610 }
1611
1612 static void
1613 lacp_sm_rx_set_expired(struct lacp_port *lp)
1614 {
1615         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1616         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1617         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1618         lp->lp_state |= LACP_STATE_EXPIRED;
1619 }
1620
1621 static void
1622 lacp_sm_rx_timer(struct lacp_port *lp)
1623 {
1624         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1625                 /* CURRENT -> EXPIRED */
1626                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1627                 lacp_sm_rx_set_expired(lp);
1628         } else {
1629                 /* EXPIRED -> DEFAULTED */
1630                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1631                 lacp_sm_rx_update_default_selected(lp);
1632                 lacp_sm_rx_record_default(lp);
1633                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1634         }
1635 }
1636
1637 static void
1638 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1639 {
1640         boolean_t active;
1641         uint8_t oldpstate;
1642         char buf[LACP_STATESTR_MAX+1];
1643
1644         LACP_TRACE(lp);
1645
1646         oldpstate = lp->lp_partner.lip_state;
1647
1648         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1649             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1650             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1651
1652         lp->lp_partner = du->ldu_actor;
1653         if (active &&
1654             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1655             LACP_STATE_AGGREGATION) &&
1656             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1657             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1658                 /* XXX nothing? */
1659         } else {
1660                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1661         }
1662
1663         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1664
1665         if (oldpstate != lp->lp_partner.lip_state) {
1666                 LACP_DPRINTF((lp, "old pstate %s\n",
1667                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1668                 LACP_DPRINTF((lp, "new pstate %s\n",
1669                     lacp_format_state(lp->lp_partner.lip_state, buf,
1670                     sizeof(buf))));
1671         }
1672
1673         /* XXX Hack, still need to implement 5.4.9 para 2,3,4 */
1674         if (lp->lp_lsc->lsc_strict_mode)
1675                 lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1676
1677         lacp_sm_ptx_update_timeout(lp, oldpstate);
1678 }
1679
1680 static void
1681 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1682 {
1683
1684         LACP_TRACE(lp);
1685
1686         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1687             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1688             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1689                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1690                 lacp_sm_assert_ntt(lp);
1691         }
1692 }
1693
1694 static void
1695 lacp_sm_rx_record_default(struct lacp_port *lp)
1696 {
1697         uint8_t oldpstate;
1698
1699         LACP_TRACE(lp);
1700
1701         oldpstate = lp->lp_partner.lip_state;
1702         if (lp->lp_lsc->lsc_strict_mode)
1703                 lp->lp_partner = lacp_partner_admin_strict;
1704         else
1705                 lp->lp_partner = lacp_partner_admin_optimistic;;
1706         lp->lp_state |= LACP_STATE_DEFAULTED;
1707         lacp_sm_ptx_update_timeout(lp, oldpstate);
1708 }
1709
1710 static void
1711 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1712     const struct lacp_peerinfo *info)
1713 {
1714
1715         LACP_TRACE(lp);
1716
1717         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1718             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1719             LACP_STATE_AGGREGATION)) {
1720                 lp->lp_selected = LACP_UNSELECTED;
1721                 /* mux machine will clean up lp->lp_aggregator */
1722         }
1723 }
1724
1725 static void
1726 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1727 {
1728
1729         LACP_TRACE(lp);
1730
1731         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1732 }
1733
1734 static void
1735 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1736 {
1737
1738         LACP_TRACE(lp);
1739
1740         if (lp->lp_lsc->lsc_strict_mode)
1741                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1742                     &lacp_partner_admin_strict);
1743         else
1744                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1745                     &lacp_partner_admin_optimistic);
1746 }
1747
1748 /* transmit machine */
1749
1750 static void
1751 lacp_sm_tx(struct lacp_port *lp)
1752 {
1753         int error = 0;
1754
1755         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1756 #if 1
1757             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1758             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1759 #endif
1760             ) {
1761                 lp->lp_flags &= ~LACP_PORT_NTT;
1762         }
1763
1764         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1765                 return;
1766         }
1767
1768         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1769         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1770                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1771                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1772                 return;
1773         }
1774
1775         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1776                 error = lacp_xmit_lacpdu(lp);
1777         } else {
1778                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1779         }
1780
1781         if (error == 0) {
1782                 lp->lp_flags &= ~LACP_PORT_NTT;
1783         } else {
1784                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1785                     error));
1786         }
1787 }
1788
1789 static void
1790 lacp_sm_assert_ntt(struct lacp_port *lp)
1791 {
1792
1793         lp->lp_flags |= LACP_PORT_NTT;
1794 }
1795
1796 static void
1797 lacp_run_timers(struct lacp_port *lp)
1798 {
1799         int i;
1800
1801         for (i = 0; i < LACP_NTIMER; i++) {
1802                 KASSERT(lp->lp_timer[i] >= 0,
1803                     ("invalid timer value %d", lp->lp_timer[i]));
1804                 if (lp->lp_timer[i] == 0) {
1805                         continue;
1806                 } else if (--lp->lp_timer[i] <= 0) {
1807                         if (lacp_timer_funcs[i]) {
1808                                 (*lacp_timer_funcs[i])(lp);
1809                         }
1810                 }
1811         }
1812 }
1813
1814 int
1815 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1816 {
1817         struct lacp_softc *lsc = lp->lp_lsc;
1818         struct lagg_port *lgp = lp->lp_lagg;
1819         struct lacp_port *lp2;
1820         struct markerdu *mdu;
1821         int error = 0;
1822         int pending = 0;
1823
1824         if (m->m_pkthdr.len != sizeof(*mdu)) {
1825                 goto bad;
1826         }
1827
1828         if ((m->m_flags & M_MCAST) == 0) {
1829                 goto bad;
1830         }
1831
1832         if (m->m_len < sizeof(*mdu)) {
1833                 m = m_pullup(m, sizeof(*mdu));
1834                 if (m == NULL) {
1835                         return (ENOMEM);
1836                 }
1837         }
1838
1839         mdu = mtod(m, struct markerdu *);
1840
1841         if (memcmp(&mdu->mdu_eh.ether_dhost,
1842             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1843                 goto bad;
1844         }
1845
1846         if (mdu->mdu_sph.sph_version != 1) {
1847                 goto bad;
1848         }
1849
1850         switch (mdu->mdu_tlv.tlv_type) {
1851         case MARKER_TYPE_INFO:
1852                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1853                     marker_info_tlv_template, TRUE)) {
1854                         goto bad;
1855                 }
1856                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1857                 memcpy(&mdu->mdu_eh.ether_dhost,
1858                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1859                 memcpy(&mdu->mdu_eh.ether_shost,
1860                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1861                 error = lagg_enqueue(lp->lp_ifp, m);
1862                 break;
1863
1864         case MARKER_TYPE_RESPONSE:
1865                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1866                     marker_response_tlv_template, TRUE)) {
1867                         goto bad;
1868                 }
1869                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1870                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1871                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1872
1873                 /* Verify that it is the last marker we sent out */
1874                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1875                     sizeof(struct lacp_markerinfo)))
1876                         goto bad;
1877
1878                 LACP_LOCK(lsc);
1879                 lp->lp_flags &= ~LACP_PORT_MARK;
1880
1881                 if (lsc->lsc_suppress_distributing) {
1882                         /* Check if any ports are waiting for a response */
1883                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1884                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1885                                         pending = 1;
1886                                         break;
1887                                 }
1888                         }
1889
1890                         if (pending == 0) {
1891                                 /* All interface queues are clear */
1892                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1893                                 lsc->lsc_suppress_distributing = FALSE;
1894                         }
1895                 }
1896                 LACP_UNLOCK(lsc);
1897                 m_freem(m);
1898                 break;
1899
1900         default:
1901                 goto bad;
1902         }
1903
1904         return (error);
1905
1906 bad:
1907         LACP_DPRINTF((lp, "bad marker frame\n"));
1908         m_freem(m);
1909         return (EINVAL);
1910 }
1911
1912 static int
1913 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1914     const struct tlv_template *tmpl, boolean_t check_type)
1915 {
1916         while (/* CONSTCOND */ 1) {
1917                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1918                         return (EINVAL);
1919                 }
1920                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1921                     tlv->tlv_length != tmpl->tmpl_length) {
1922                         return (EINVAL);
1923                 }
1924                 if (tmpl->tmpl_type == 0) {
1925                         break;
1926                 }
1927                 tlv = (const struct tlvhdr *)
1928                     ((const char *)tlv + tlv->tlv_length);
1929                 tmpl++;
1930         }
1931
1932         return (0);
1933 }
1934
1935 /* Debugging */
1936 const char *
1937 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1938 {
1939         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1940             (int)mac[0],
1941             (int)mac[1],
1942             (int)mac[2],
1943             (int)mac[3],
1944             (int)mac[4],
1945             (int)mac[5]);
1946
1947         return (buf);
1948 }
1949
1950 const char *
1951 lacp_format_systemid(const struct lacp_systemid *sysid,
1952     char *buf, size_t buflen)
1953 {
1954         char macbuf[LACP_MACSTR_MAX+1];
1955
1956         snprintf(buf, buflen, "%04X,%s",
1957             ntohs(sysid->lsi_prio),
1958             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
1959
1960         return (buf);
1961 }
1962
1963 const char *
1964 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
1965 {
1966         snprintf(buf, buflen, "%04X,%04X",
1967             ntohs(portid->lpi_prio),
1968             ntohs(portid->lpi_portno));
1969
1970         return (buf);
1971 }
1972
1973 const char *
1974 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
1975 {
1976         char sysid[LACP_SYSTEMIDSTR_MAX+1];
1977         char portid[LACP_PORTIDSTR_MAX+1];
1978
1979         snprintf(buf, buflen, "(%s,%04X,%s)",
1980             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
1981             ntohs(peer->lip_key),
1982             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
1983
1984         return (buf);
1985 }
1986
1987 const char *
1988 lacp_format_lagid(const struct lacp_peerinfo *a,
1989     const struct lacp_peerinfo *b, char *buf, size_t buflen)
1990 {
1991         char astr[LACP_PARTNERSTR_MAX+1];
1992         char bstr[LACP_PARTNERSTR_MAX+1];
1993
1994 #if 0
1995         /*
1996          * there's a convention to display small numbered peer
1997          * in the left.
1998          */
1999
2000         if (lacp_compare_peerinfo(a, b) > 0) {
2001                 const struct lacp_peerinfo *t;
2002
2003                 t = a;
2004                 a = b;
2005                 b = t;
2006         }
2007 #endif
2008
2009         snprintf(buf, buflen, "[%s,%s]",
2010             lacp_format_partner(a, astr, sizeof(astr)),
2011             lacp_format_partner(b, bstr, sizeof(bstr)));
2012
2013         return (buf);
2014 }
2015
2016 const char *
2017 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2018     char *buf, size_t buflen)
2019 {
2020         if (la == NULL) {
2021                 return ("(none)");
2022         }
2023
2024         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2025 }
2026
2027 const char *
2028 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2029 {
2030         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2031         return (buf);
2032 }
2033
2034 static void
2035 lacp_dump_lacpdu(const struct lacpdu *du)
2036 {
2037         char buf[LACP_PARTNERSTR_MAX+1];
2038         char buf2[LACP_STATESTR_MAX+1];
2039
2040         printf("actor=%s\n",
2041             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2042         printf("actor.state=%s\n",
2043             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2044         printf("partner=%s\n",
2045             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2046         printf("partner.state=%s\n",
2047             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2048
2049         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2050 }
2051
2052 static void
2053 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2054 {
2055         va_list va;
2056
2057         if (lp) {
2058                 printf("%s: ", lp->lp_ifp->if_xname);
2059         }
2060
2061         va_start(va, fmt);
2062         vprintf(fmt, va);
2063         va_end(va);
2064 }