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