sys/net/ieee8023ad_lacp.c

   1 /*      $NetBSD: ieee8023ad_lacp.c,v 1.3 2005/12/11 12:24:54 christos Exp $     */
   2
   3 /*-
   4  * SPDX-License-Identifier: BSD-2-Clause-NetBSD
   5  *
   6  * Copyright (c)2005 YAMAMOTO Takashi,
   7  * Copyright (c)2008 Andrew Thompson <thompsa@FreeBSD.org>
   8  * All rights reserved.
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  13  * 1. Redistributions of source code must retain the above copyright
  14  *    notice, this list of conditions and the following disclaimer.
  15  * 2. Redistributions in binary form must reproduce the above copyright
  16  *    notice, this list of conditions and the following disclaimer in the
  17  *    documentation and/or other materials provided with the distribution.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  29  * SUCH DAMAGE.
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include "opt_ratelimit.h"
  36
  37 #include <sys/param.h>
  38 #include <sys/callout.h>
  39 #include <sys/eventhandler.h>
  40 #include <sys/mbuf.h>
  41 #include <sys/systm.h>
  42 #include <sys/malloc.h>
  43 #include <sys/kernel.h> /* hz */
  44 #include <sys/socket.h> /* for net/if.h */
  45 #include <sys/sockio.h>
  46 #include <sys/sysctl.h>
  47 #include <machine/stdarg.h>
  48 #include <sys/lock.h>
  49 #include <sys/rwlock.h>
  50 #include <sys/taskqueue.h>
  51
  52 #include <net/if.h>
  53 #include <net/if_var.h>
  54 #include <net/if_dl.h>
  55 #include <net/ethernet.h>
  56 #include <net/if_media.h>
  57 #include <net/if_types.h>
  58
  59 #include <net/if_lagg.h>
  60 #include <net/ieee8023ad_lacp.h>
  61
  62 /*
  63  * actor system priority and port priority.
  64  * XXX should be configurable.
  65  */
  66
  67 #define LACP_SYSTEM_PRIO        0x8000
  68 #define LACP_PORT_PRIO          0x8000
  69
  70 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  71     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  72
  73 static const struct tlv_template lacp_info_tlv_template[] = {
  74         { LACP_TYPE_ACTORINFO,
  75             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  76         { LACP_TYPE_PARTNERINFO,
  77             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  78         { LACP_TYPE_COLLECTORINFO,
  79             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  80         { 0, 0 },
  81 };
  82
  83 static const struct tlv_template marker_info_tlv_template[] = {
  84         { MARKER_TYPE_INFO,
  85             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  86         { 0, 0 },
  87 };
  88
  89 static const struct tlv_template marker_response_tlv_template[] = {
  90         { MARKER_TYPE_RESPONSE,
  91             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  92         { 0, 0 },
  93 };
  94
  95 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  96
  97 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
  98 static void     lacp_fill_markerinfo(struct lacp_port *,
  99                     struct lacp_markerinfo *);
 100
 101 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
 102 static void     lacp_suppress_distributing(struct lacp_softc *,
 103                     struct lacp_aggregator *);
 104 static void     lacp_transit_expire(void *);
 105 static void     lacp_update_portmap(struct lacp_softc *);
 106 static void     lacp_select_active_aggregator(struct lacp_softc *);
 107 static uint16_t lacp_compose_key(struct lacp_port *);
 108 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 109                     const struct tlv_template *, boolean_t);
 110 static void     lacp_tick(void *);
 111
 112 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 113                     const struct lacp_port *);
 114 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 115                     const struct lacp_peerinfo *);
 116 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 117                     const struct lacp_port *);
 118 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 119                     const struct lacp_peerinfo *);
 120
 121 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 122                     struct lacp_port *);
 123 static void     lacp_aggregator_addref(struct lacp_softc *,
 124                     struct lacp_aggregator *);
 125 static void     lacp_aggregator_delref(struct lacp_softc *,
 126                     struct lacp_aggregator *);
 127
 128 /* receive machine */
 129
 130 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 131 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 132 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 133 static void     lacp_sm_rx_timer(struct lacp_port *);
 134 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 135 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 136                     const struct lacpdu *);
 137 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 138                     const struct lacpdu *);
 139 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 140                     const struct lacpdu *);
 141 static void     lacp_sm_rx_record_default(struct lacp_port *);
 142 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 143 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 144                     const struct lacp_peerinfo *);
 145
 146 /* mux machine */
 147
 148 static void     lacp_sm_mux(struct lacp_port *);
 149 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 150 static void     lacp_sm_mux_timer(struct lacp_port *);
 151
 152 /* periodic transmit machine */
 153
 154 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 155 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 156 static void     lacp_sm_ptx_timer(struct lacp_port *);
 157
 158 /* transmit machine */
 159
 160 static void     lacp_sm_tx(struct lacp_port *);
 161 static void     lacp_sm_assert_ntt(struct lacp_port *);
 162
 163 static void     lacp_run_timers(struct lacp_port *);
 164 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 165                     const struct lacp_peerinfo *);
 166 static int      lacp_compare_systemid(const struct lacp_systemid *,
 167                     const struct lacp_systemid *);
 168 static void     lacp_port_enable(struct lacp_port *);
 169 static void     lacp_port_disable(struct lacp_port *);
 170 static void     lacp_select(struct lacp_port *);
 171 static void     lacp_unselect(struct lacp_port *);
 172 static void     lacp_disable_collecting(struct lacp_port *);
 173 static void     lacp_enable_collecting(struct lacp_port *);
 174 static void     lacp_disable_distributing(struct lacp_port *);
 175 static void     lacp_enable_distributing(struct lacp_port *);
 176 static int      lacp_xmit_lacpdu(struct lacp_port *);
 177 static int      lacp_xmit_marker(struct lacp_port *);
 178
 179 /* Debugging */
 180
 181 static void     lacp_dump_lacpdu(const struct lacpdu *);
 182 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 183                     size_t);
 184 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 185                     const struct lacp_peerinfo *, char *, size_t);
 186 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 187                     char *, size_t);
 188 static const char *lacp_format_state(uint8_t, char *, size_t);
 189 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 190 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 191                     size_t);
 192 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 193                     size_t);
 194 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 195                     __attribute__((__format__(__printf__, 2, 3)));
 196
 197 VNET_DEFINE_STATIC(int, lacp_debug);
 198 #define V_lacp_debug    VNET(lacp_debug)
 199 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
 200 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 201     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 202
 203 VNET_DEFINE_STATIC(int, lacp_default_strict_mode) = 1;
 204 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode,
 205     CTLFLAG_RWTUN | CTLFLAG_VNET, &VNET_NAME(lacp_default_strict_mode), 0,
 206     "LACP strict protocol compliance default");
 207
 208 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 209 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 210 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 211
 212 /*
 213  * partner administration variables.
 214  * XXX should be configurable.
 215  */
 216
 217 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 218         .lip_systemid = { .lsi_prio = 0xffff },
 219         .lip_portid = { .lpi_prio = 0xffff },
 220         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 221             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 222 };
 223
 224 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 225         .lip_systemid = { .lsi_prio = 0xffff },
 226         .lip_portid = { .lpi_prio = 0xffff },
 227         .lip_state = 0,
 228 };
 229
 230 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 231         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 232         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 233         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 234 };
 235
 236 struct mbuf *
 237 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 238 {
 239         struct lacp_port *lp = LACP_PORT(lgp);
 240         uint8_t subtype;
 241
 242         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 243                 m_freem(m);
 244                 return (NULL);
 245         }
 246
 247         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 248         switch (subtype) {
 249                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 250                         lacp_pdu_input(lp, m);
 251                         return (NULL);
 252
 253                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 254                         lacp_marker_input(lp, m);
 255                         return (NULL);
 256         }
 257
 258         /* Not a subtype we are interested in */
 259         return (m);
 260 }
 261
 262 /*
 263  * lacp_pdu_input: process lacpdu
 264  */
 265 static int
 266 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 267 {
 268         struct lacp_softc *lsc = lp->lp_lsc;
 269         struct lacpdu *du;
 270         int error = 0;
 271
 272         if (m->m_pkthdr.len != sizeof(*du)) {
 273                 goto bad;
 274         }
 275
 276         if ((m->m_flags & M_MCAST) == 0) {
 277                 goto bad;
 278         }
 279
 280         if (m->m_len < sizeof(*du)) {
 281                 m = m_pullup(m, sizeof(*du));
 282                 if (m == NULL) {
 283                         return (ENOMEM);
 284                 }
 285         }
 286
 287         du = mtod(m, struct lacpdu *);
 288
 289         if (memcmp(&du->ldu_eh.ether_dhost,
 290             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 291                 goto bad;
 292         }
 293
 294         /*
 295          * ignore the version for compatibility with
 296          * the future protocol revisions.
 297          */
 298 #if 0
 299         if (du->ldu_sph.sph_version != 1) {
 300                 goto bad;
 301         }
 302 #endif
 303
 304         /*
 305          * ignore tlv types for compatibility with
 306          * the future protocol revisions.
 307          */
 308         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 309             lacp_info_tlv_template, FALSE)) {
 310                 goto bad;
 311         }
 312
 313         if (V_lacp_debug > 0) {
 314                 lacp_dprintf(lp, "lacpdu receive\n");
 315                 lacp_dump_lacpdu(du);
 316         }
 317
 318         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 319                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 320                 goto bad;
 321         }
 322
 323         LACP_LOCK(lsc);
 324         lacp_sm_rx(lp, du);
 325         LACP_UNLOCK(lsc);
 326
 327         m_freem(m);
 328         return (error);
 329
 330 bad:
 331         m_freem(m);
 332         return (EINVAL);
 333 }
 334
 335 static void
 336 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 337 {
 338         struct lagg_port *lgp = lp->lp_lagg;
 339         struct lagg_softc *sc = lgp->lp_softc;
 340
 341         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 342         memcpy(&info->lip_systemid.lsi_mac,
 343             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 344         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 345         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 346         info->lip_state = lp->lp_state;
 347 }
 348
 349 static void
 350 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 351 {
 352         struct ifnet *ifp = lp->lp_ifp;
 353
 354         /* Fill in the port index and system id (encoded as the MAC) */
 355         info->mi_rq_port = htons(ifp->if_index);
 356         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 357         info->mi_rq_xid = htonl(0);
 358 }
 359
 360 static int
 361 lacp_xmit_lacpdu(struct lacp_port *lp)
 362 {
 363         struct lagg_port *lgp = lp->lp_lagg;
 364         struct mbuf *m;
 365         struct lacpdu *du;
 366         int error;
 367
 368         LACP_LOCK_ASSERT(lp->lp_lsc);
 369
 370         m = m_gethdr(M_NOWAIT, MT_DATA);
 371         if (m == NULL) {
 372                 return (ENOMEM);
 373         }
 374         m->m_len = m->m_pkthdr.len = sizeof(*du);
 375
 376         du = mtod(m, struct lacpdu *);
 377         memset(du, 0, sizeof(*du));
 378
 379         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 380             ETHER_ADDR_LEN);
 381         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 382         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 383
 384         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 385         du->ldu_sph.sph_version = 1;
 386
 387         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 388         du->ldu_actor = lp->lp_actor;
 389
 390         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 391             sizeof(du->ldu_partner));
 392         du->ldu_partner = lp->lp_partner;
 393
 394         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 395             sizeof(du->ldu_collector));
 396         du->ldu_collector.lci_maxdelay = 0;
 397
 398         if (V_lacp_debug > 0) {
 399                 lacp_dprintf(lp, "lacpdu transmit\n");
 400                 lacp_dump_lacpdu(du);
 401         }
 402
 403         m->m_flags |= M_MCAST;
 404
 405         /*
 406          * XXX should use higher priority queue.
 407          * otherwise network congestion can break aggregation.
 408          */
 409
 410         error = lagg_enqueue(lp->lp_ifp, m);
 411         return (error);
 412 }
 413
 414 static int
 415 lacp_xmit_marker(struct lacp_port *lp)
 416 {
 417         struct lagg_port *lgp = lp->lp_lagg;
 418         struct mbuf *m;
 419         struct markerdu *mdu;
 420         int error;
 421
 422         LACP_LOCK_ASSERT(lp->lp_lsc);
 423
 424         m = m_gethdr(M_NOWAIT, MT_DATA);
 425         if (m == NULL) {
 426                 return (ENOMEM);
 427         }
 428         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 429
 430         mdu = mtod(m, struct markerdu *);
 431         memset(mdu, 0, sizeof(*mdu));
 432
 433         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 434             ETHER_ADDR_LEN);
 435         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 436         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 437
 438         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 439         mdu->mdu_sph.sph_version = 1;
 440
 441         /* Bump the transaction id and copy over the marker info */
 442         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 443         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 444         mdu->mdu_info = lp->lp_marker;
 445
 446         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 447             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 448             ntohl(mdu->mdu_info.mi_rq_xid)));
 449
 450         m->m_flags |= M_MCAST;
 451         error = lagg_enqueue(lp->lp_ifp, m);
 452         return (error);
 453 }
 454
 455 void
 456 lacp_linkstate(struct lagg_port *lgp)
 457 {
 458         struct lacp_port *lp = LACP_PORT(lgp);
 459         struct lacp_softc *lsc = lp->lp_lsc;
 460         struct ifnet *ifp = lgp->lp_ifp;
 461         struct ifmediareq ifmr;
 462         int error = 0;
 463         u_int media;
 464         uint8_t old_state;
 465         uint16_t old_key;
 466
 467         bzero((char *)&ifmr, sizeof(ifmr));
 468         error = (*ifp->if_ioctl)(ifp, SIOCGIFXMEDIA, (caddr_t)&ifmr);
 469         if (error != 0) {
 470                 bzero((char *)&ifmr, sizeof(ifmr));
 471                 error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 472         }
 473         if (error != 0)
 474                 return;
 475
 476         LACP_LOCK(lsc);
 477         media = ifmr.ifm_active;
 478         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 479             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 480             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 481         old_state = lp->lp_state;
 482         old_key = lp->lp_key;
 483
 484         lp->lp_media = media;
 485         /*
 486          * If the port is not an active full duplex Ethernet link then it can
 487          * not be aggregated.
 488          */
 489         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 490             ifp->if_link_state != LINK_STATE_UP) {
 491                 lacp_port_disable(lp);
 492         } else {
 493                 lacp_port_enable(lp);
 494         }
 495         lp->lp_key = lacp_compose_key(lp);
 496
 497         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 498                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 499                 lp->lp_selected = LACP_UNSELECTED;
 500         }
 501         LACP_UNLOCK(lsc);
 502 }
 503
 504 static void
 505 lacp_tick(void *arg)
 506 {
 507         struct lacp_softc *lsc = arg;
 508         struct lacp_port *lp;
 509
 510         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 511                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 512                         continue;
 513
 514                 CURVNET_SET(lp->lp_ifp->if_vnet);
 515                 lacp_run_timers(lp);
 516
 517                 lacp_select(lp);
 518                 lacp_sm_mux(lp);
 519                 lacp_sm_tx(lp);
 520                 lacp_sm_ptx_tx_schedule(lp);
 521                 CURVNET_RESTORE();
 522         }
 523         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 524 }
 525
 526 int
 527 lacp_port_create(struct lagg_port *lgp)
 528 {
 529         struct lagg_softc *sc = lgp->lp_softc;
 530         struct lacp_softc *lsc = LACP_SOFTC(sc);
 531         struct lacp_port *lp;
 532         struct ifnet *ifp = lgp->lp_ifp;
 533         struct sockaddr_dl sdl;
 534         struct ifmultiaddr *rifma = NULL;
 535         int error;
 536
 537         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 538         sdl.sdl_alen = ETHER_ADDR_LEN;
 539
 540         bcopy(&ethermulticastaddr_slowprotocols,
 541             LLADDR(&sdl), ETHER_ADDR_LEN);
 542         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 543         if (error) {
 544                 printf("%s: ADDMULTI failed on %s\n", __func__,
 545                     lgp->lp_ifp->if_xname);
 546                 return (error);
 547         }
 548
 549         lp = malloc(sizeof(struct lacp_port),
 550             M_DEVBUF, M_NOWAIT|M_ZERO);
 551         if (lp == NULL)
 552                 return (ENOMEM);
 553
 554         LACP_LOCK(lsc);
 555         lgp->lp_psc = lp;
 556         lp->lp_ifp = ifp;
 557         lp->lp_lagg = lgp;
 558         lp->lp_lsc = lsc;
 559         lp->lp_ifma = rifma;
 560
 561         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 562
 563         lacp_fill_actorinfo(lp, &lp->lp_actor);
 564         lacp_fill_markerinfo(lp, &lp->lp_marker);
 565         lp->lp_state = LACP_STATE_ACTIVITY;
 566         lp->lp_aggregator = NULL;
 567         lacp_sm_rx_set_expired(lp);
 568         LACP_UNLOCK(lsc);
 569         lacp_linkstate(lgp);
 570
 571         return (0);
 572 }
 573
 574 void
 575 lacp_port_destroy(struct lagg_port *lgp)
 576 {
 577         struct lacp_port *lp = LACP_PORT(lgp);
 578         struct lacp_softc *lsc = lp->lp_lsc;
 579         int i;
 580
 581         LACP_LOCK(lsc);
 582         for (i = 0; i < LACP_NTIMER; i++) {
 583                 LACP_TIMER_DISARM(lp, i);
 584         }
 585
 586         lacp_disable_collecting(lp);
 587         lacp_disable_distributing(lp);
 588         lacp_unselect(lp);
 589
 590         LIST_REMOVE(lp, lp_next);
 591         LACP_UNLOCK(lsc);
 592
 593         /* The address may have already been removed by if_purgemaddrs() */
 594         if (!lgp->lp_detaching)
 595                 if_delmulti_ifma(lp->lp_ifma);
 596
 597         free(lp, M_DEVBUF);
 598 }
 599
 600 void
 601 lacp_req(struct lagg_softc *sc, void *data)
 602 {
 603         struct lacp_opreq *req = (struct lacp_opreq *)data;
 604         struct lacp_softc *lsc = LACP_SOFTC(sc);
 605         struct lacp_aggregator *la;
 606
 607         bzero(req, sizeof(struct lacp_opreq));
 608
 609         /*
 610          * If the LACP softc is NULL, return with the opreq structure full of
 611          * zeros.  It is normal for the softc to be NULL while the lagg is
 612          * being destroyed.
 613          */
 614         if (NULL == lsc)
 615                 return;
 616
 617         la = lsc->lsc_active_aggregator;
 618         LACP_LOCK(lsc);
 619         if (la != NULL) {
 620                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 621                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 622                     ETHER_ADDR_LEN);
 623                 req->actor_key = ntohs(la->la_actor.lip_key);
 624                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 625                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 626                 req->actor_state = la->la_actor.lip_state;
 627
 628                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 629                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 630                     ETHER_ADDR_LEN);
 631                 req->partner_key = ntohs(la->la_partner.lip_key);
 632                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 633                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 634                 req->partner_state = la->la_partner.lip_state;
 635         }
 636         LACP_UNLOCK(lsc);
 637 }
 638
 639 void
 640 lacp_portreq(struct lagg_port *lgp, void *data)
 641 {
 642         struct lacp_opreq *req = (struct lacp_opreq *)data;
 643         struct lacp_port *lp = LACP_PORT(lgp);
 644         struct lacp_softc *lsc = lp->lp_lsc;
 645
 646         LACP_LOCK(lsc);
 647         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 648         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 649             ETHER_ADDR_LEN);
 650         req->actor_key = ntohs(lp->lp_actor.lip_key);
 651         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 652         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 653         req->actor_state = lp->lp_actor.lip_state;
 654
 655         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 656         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 657             ETHER_ADDR_LEN);
 658         req->partner_key = ntohs(lp->lp_partner.lip_key);
 659         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 660         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 661         req->partner_state = lp->lp_partner.lip_state;
 662         LACP_UNLOCK(lsc);
 663 }
 664
 665 static void
 666 lacp_disable_collecting(struct lacp_port *lp)
 667 {
 668         LACP_DPRINTF((lp, "collecting disabled\n"));
 669         lp->lp_state &= ~LACP_STATE_COLLECTING;
 670 }
 671
 672 static void
 673 lacp_enable_collecting(struct lacp_port *lp)
 674 {
 675         LACP_DPRINTF((lp, "collecting enabled\n"));
 676         lp->lp_state |= LACP_STATE_COLLECTING;
 677 }
 678
 679 static void
 680 lacp_disable_distributing(struct lacp_port *lp)
 681 {
 682         struct lacp_aggregator *la = lp->lp_aggregator;
 683         struct lacp_softc *lsc = lp->lp_lsc;
 684         struct lagg_softc *sc = lsc->lsc_softc;
 685         char buf[LACP_LAGIDSTR_MAX+1];
 686
 687         LACP_LOCK_ASSERT(lsc);
 688
 689         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 690                 return;
 691         }
 692
 693         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 694         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 695         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 696
 697         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 698             "nports %d -> %d\n",
 699             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 700             la->la_nports, la->la_nports - 1));
 701
 702         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 703         la->la_nports--;
 704         sc->sc_active = la->la_nports;
 705
 706         if (lsc->lsc_active_aggregator == la) {
 707                 lacp_suppress_distributing(lsc, la);
 708                 lacp_select_active_aggregator(lsc);
 709                 /* regenerate the port map, the active aggregator has changed */
 710                 lacp_update_portmap(lsc);
 711         }
 712
 713         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 714         if_link_state_change(sc->sc_ifp,
 715             sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
 716 }
 717
 718 static void
 719 lacp_enable_distributing(struct lacp_port *lp)
 720 {
 721         struct lacp_aggregator *la = lp->lp_aggregator;
 722         struct lacp_softc *lsc = lp->lp_lsc;
 723         struct lagg_softc *sc = lsc->lsc_softc;
 724         char buf[LACP_LAGIDSTR_MAX+1];
 725
 726         LACP_LOCK_ASSERT(lsc);
 727
 728         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 729                 return;
 730         }
 731
 732         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 733             "nports %d -> %d\n",
 734             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 735             la->la_nports, la->la_nports + 1));
 736
 737         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 738         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 739         la->la_nports++;
 740         sc->sc_active = la->la_nports;
 741
 742         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 743
 744         if (lsc->lsc_active_aggregator == la) {
 745                 lacp_suppress_distributing(lsc, la);
 746                 lacp_update_portmap(lsc);
 747         } else
 748                 /* try to become the active aggregator */
 749                 lacp_select_active_aggregator(lsc);
 750
 751         if_link_state_change(sc->sc_ifp,
 752             sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
 753 }
 754
 755 static void
 756 lacp_transit_expire(void *vp)
 757 {
 758         struct lacp_softc *lsc = vp;
 759
 760         LACP_LOCK_ASSERT(lsc);
 761
 762         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 763         LACP_TRACE(NULL);
 764         CURVNET_RESTORE();
 765
 766         lsc->lsc_suppress_distributing = FALSE;
 767 }
 768
 769 void
 770 lacp_attach(struct lagg_softc *sc)
 771 {
 772         struct lacp_softc *lsc;
 773
 774         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 775
 776         sc->sc_psc = lsc;
 777         lsc->lsc_softc = sc;
 778
 779         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 780         lsc->lsc_active_aggregator = NULL;
 781         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 782         LACP_LOCK_INIT(lsc);
 783         TAILQ_INIT(&lsc->lsc_aggregators);
 784         LIST_INIT(&lsc->lsc_ports);
 785
 786         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 787         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 788
 789         /* if the lagg is already up then do the same */
 790         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 791                 lacp_init(sc);
 792 }
 793
 794 void
 795 lacp_detach(void *psc)
 796 {
 797         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 798
 799         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 800             ("aggregators still active"));
 801         KASSERT(lsc->lsc_active_aggregator == NULL,
 802             ("aggregator still attached"));
 803
 804         callout_drain(&lsc->lsc_transit_callout);
 805         callout_drain(&lsc->lsc_callout);
 806
 807         LACP_LOCK_DESTROY(lsc);
 808         free(lsc, M_DEVBUF);
 809 }
 810
 811 void
 812 lacp_init(struct lagg_softc *sc)
 813 {
 814         struct lacp_softc *lsc = LACP_SOFTC(sc);
 815
 816         LACP_LOCK(lsc);
 817         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 818         LACP_UNLOCK(lsc);
 819 }
 820
 821 void
 822 lacp_stop(struct lagg_softc *sc)
 823 {
 824         struct lacp_softc *lsc = LACP_SOFTC(sc);
 825
 826         LACP_LOCK(lsc);
 827         callout_stop(&lsc->lsc_transit_callout);
 828         callout_stop(&lsc->lsc_callout);
 829         LACP_UNLOCK(lsc);
 830 }
 831
 832 struct lagg_port *
 833 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 834 {
 835         struct lacp_softc *lsc = LACP_SOFTC(sc);
 836         struct lacp_portmap *pm;
 837         struct lacp_port *lp;
 838         uint32_t hash;
 839
 840         if (__predict_false(lsc->lsc_suppress_distributing)) {
 841                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 842                 return (NULL);
 843         }
 844
 845         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 846         if (pm->pm_count == 0) {
 847                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 848                 return (NULL);
 849         }
 850
 851         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 852             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 853                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 854         else
 855                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 856         hash %= pm->pm_count;
 857         lp = pm->pm_map[hash];
 858
 859         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 860             ("aggregated port is not distributing"));
 861
 862         return (lp->lp_lagg);
 863 }
 864
 865 #ifdef RATELIMIT
 866 struct lagg_port *
 867 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t flowid)
 868 {
 869         struct lacp_softc *lsc = LACP_SOFTC(sc);
 870         struct lacp_portmap *pm;
 871         struct lacp_port *lp;
 872         uint32_t hash;
 873
 874         if (__predict_false(lsc->lsc_suppress_distributing)) {
 875                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 876                 return (NULL);
 877         }
 878
 879         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 880         if (pm->pm_count == 0) {
 881                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 882                 return (NULL);
 883         }
 884
 885         hash = flowid >> sc->flowid_shift;
 886         hash %= pm->pm_count;
 887         lp = pm->pm_map[hash];
 888
 889         return (lp->lp_lagg);
 890 }
 891 #endif
 892
 893 /*
 894  * lacp_suppress_distributing: drop transmit packets for a while
 895  * to preserve packet ordering.
 896  */
 897
 898 static void
 899 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 900 {
 901         struct lacp_port *lp;
 902
 903         if (lsc->lsc_active_aggregator != la) {
 904                 return;
 905         }
 906
 907         LACP_TRACE(NULL);
 908
 909         lsc->lsc_suppress_distributing = TRUE;
 910
 911         /* send a marker frame down each port to verify the queues are empty */
 912         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 913                 lp->lp_flags |= LACP_PORT_MARK;
 914                 lacp_xmit_marker(lp);
 915         }
 916
 917         /* set a timeout for the marker frames */
 918         callout_reset(&lsc->lsc_transit_callout,
 919             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 920 }
 921
 922 static int
 923 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 924     const struct lacp_peerinfo *b)
 925 {
 926         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 927 }
 928
 929 static int
 930 lacp_compare_systemid(const struct lacp_systemid *a,
 931     const struct lacp_systemid *b)
 932 {
 933         return (memcmp(a, b, sizeof(*a)));
 934 }
 935
 936 #if 0   /* unused */
 937 static int
 938 lacp_compare_portid(const struct lacp_portid *a,
 939     const struct lacp_portid *b)
 940 {
 941         return (memcmp(a, b, sizeof(*a)));
 942 }
 943 #endif
 944
 945 static uint64_t
 946 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 947 {
 948         struct lacp_port *lp;
 949         uint64_t speed;
 950
 951         lp = TAILQ_FIRST(&la->la_ports);
 952         if (lp == NULL) {
 953                 return (0);
 954         }
 955
 956         speed = ifmedia_baudrate(lp->lp_media);
 957         speed *= la->la_nports;
 958         if (speed == 0) {
 959                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 960                     lp->lp_media, la->la_nports));
 961         }
 962
 963         return (speed);
 964 }
 965
 966 /*
 967  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 968  * packets from lagg(4) interface.
 969  */
 970
 971 static void
 972 lacp_select_active_aggregator(struct lacp_softc *lsc)
 973 {
 974         struct lacp_aggregator *la;
 975         struct lacp_aggregator *best_la = NULL;
 976         uint64_t best_speed = 0;
 977         char buf[LACP_LAGIDSTR_MAX+1];
 978
 979         LACP_TRACE(NULL);
 980
 981         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 982                 uint64_t speed;
 983
 984                 if (la->la_nports == 0) {
 985                         continue;
 986                 }
 987
 988                 speed = lacp_aggregator_bandwidth(la);
 989                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 990                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 991                     speed, la->la_nports));
 992
 993                 /*
 994                  * This aggregator is chosen if the partner has a better
 995                  * system priority or, the total aggregated speed is higher
 996                  * or, it is already the chosen aggregator
 997                  */
 998                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 999                     LACP_SYS_PRI(best_la->la_partner)) ||
1000                     speed > best_speed ||
1001                     (speed == best_speed &&
1002                     la == lsc->lsc_active_aggregator)) {
1003                         best_la = la;
1004                         best_speed = speed;
1005                 }
1006         }
1007
1008         KASSERT(best_la == NULL || best_la->la_nports > 0,
1009             ("invalid aggregator refcnt"));
1010         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1011             ("invalid aggregator list"));
1012
1013         if (lsc->lsc_active_aggregator != best_la) {
1014                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1015                 LACP_DPRINTF((NULL, "old %s\n",
1016                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1017                     buf, sizeof(buf))));
1018         } else {
1019                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1020         }
1021         LACP_DPRINTF((NULL, "new %s\n",
1022             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1023
1024         if (lsc->lsc_active_aggregator != best_la) {
1025                 lsc->lsc_active_aggregator = best_la;
1026                 lacp_update_portmap(lsc);
1027                 if (best_la) {
1028                         lacp_suppress_distributing(lsc, best_la);
1029                 }
1030         }
1031 }
1032
1033 /*
1034  * Updated the inactive portmap array with the new list of ports and
1035  * make it live.
1036  */
1037 static void
1038 lacp_update_portmap(struct lacp_softc *lsc)
1039 {
1040         struct lagg_softc *sc = lsc->lsc_softc;
1041         struct lacp_aggregator *la;
1042         struct lacp_portmap *p;
1043         struct lacp_port *lp;
1044         uint64_t speed;
1045         u_int newmap;
1046         int i;
1047
1048         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1049         p = &lsc->lsc_pmap[newmap];
1050         la = lsc->lsc_active_aggregator;
1051         speed = 0;
1052         bzero(p, sizeof(struct lacp_portmap));
1053
1054         if (la != NULL && la->la_nports > 0) {
1055                 p->pm_count = la->la_nports;
1056                 i = 0;
1057                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1058                         p->pm_map[i++] = lp;
1059                 KASSERT(i == p->pm_count, ("Invalid port count"));
1060                 speed = lacp_aggregator_bandwidth(la);
1061         }
1062         sc->sc_ifp->if_baudrate = speed;
1063
1064         /* switch the active portmap over */
1065         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1066         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1067                     lsc->lsc_activemap,
1068                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1069 }
1070
1071 static uint16_t
1072 lacp_compose_key(struct lacp_port *lp)
1073 {
1074         struct lagg_port *lgp = lp->lp_lagg;
1075         struct lagg_softc *sc = lgp->lp_softc;
1076         u_int media = lp->lp_media;
1077         uint16_t key;
1078
1079         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1080
1081                 /*
1082                  * non-aggregatable links should have unique keys.
1083                  *
1084                  * XXX this isn't really unique as if_index is 16 bit.
1085                  */
1086
1087                 /* bit 0..14:   (some bits of) if_index of this port */
1088                 key = lp->lp_ifp->if_index;
1089                 /* bit 15:      1 */
1090                 key |= 0x8000;
1091         } else {
1092                 u_int subtype = IFM_SUBTYPE(media);
1093
1094                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1095                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1096
1097                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1098                 switch (subtype) {
1099                 case IFM_10_T:
1100                 case IFM_10_2:
1101                 case IFM_10_5:
1102                 case IFM_10_STP:
1103                 case IFM_10_FL:
1104                         key = IFM_10_T;
1105                         break;
1106                 case IFM_100_TX:
1107                 case IFM_100_FX:
1108                 case IFM_100_T4:
1109                 case IFM_100_VG:
1110                 case IFM_100_T2:
1111                 case IFM_100_T:
1112                 case IFM_100_SGMII:
1113                         key = IFM_100_TX;
1114                         break;
1115                 case IFM_1000_SX:
1116                 case IFM_1000_LX:
1117                 case IFM_1000_CX:
1118                 case IFM_1000_T:
1119                 case IFM_1000_KX:
1120                 case IFM_1000_SGMII:
1121                 case IFM_1000_CX_SGMII:
1122                         key = IFM_1000_SX;
1123                         break;
1124                 case IFM_10G_LR:
1125                 case IFM_10G_SR:
1126                 case IFM_10G_CX4:
1127                 case IFM_10G_TWINAX:
1128                 case IFM_10G_TWINAX_LONG:
1129                 case IFM_10G_LRM:
1130                 case IFM_10G_T:
1131                 case IFM_10G_KX4:
1132                 case IFM_10G_KR:
1133                 case IFM_10G_CR1:
1134                 case IFM_10G_ER:
1135                 case IFM_10G_SFI:
1136                 case IFM_10G_AOC:
1137                         key = IFM_10G_LR;
1138                         break;
1139                 case IFM_20G_KR2:
1140                         key = IFM_20G_KR2;
1141                         break;
1142                 case IFM_2500_KX:
1143                 case IFM_2500_T:
1144                 case IFM_2500_X:
1145                         key = IFM_2500_KX;
1146                         break;
1147                 case IFM_5000_T:
1148                 case IFM_5000_KR:
1149                 case IFM_5000_KR_S:
1150                 case IFM_5000_KR1:
1151                         key = IFM_5000_T;
1152                         break;
1153                 case IFM_50G_PCIE:
1154                 case IFM_50G_CR2:
1155                 case IFM_50G_KR2:
1156                 case IFM_50G_SR2:
1157                 case IFM_50G_LR2:
1158                 case IFM_50G_LAUI2_AC:
1159                 case IFM_50G_LAUI2:
1160                 case IFM_50G_AUI2_AC:
1161                 case IFM_50G_AUI2:
1162                 case IFM_50G_CP:
1163                 case IFM_50G_SR:
1164                 case IFM_50G_LR:
1165                 case IFM_50G_FR:
1166                 case IFM_50G_KR_PAM4:
1167                 case IFM_50G_AUI1_AC:
1168                 case IFM_50G_AUI1:
1169                         key = IFM_50G_PCIE;
1170                         break;
1171                 case IFM_56G_R4:
1172                         key = IFM_56G_R4;
1173                         break;
1174                 case IFM_25G_PCIE:
1175                 case IFM_25G_CR:
1176                 case IFM_25G_KR:
1177                 case IFM_25G_SR:
1178                 case IFM_25G_LR:
1179                 case IFM_25G_ACC:
1180                 case IFM_25G_AOC:
1181                 case IFM_25G_T:
1182                 case IFM_25G_CR_S:
1183                 case IFM_25G_CR1:
1184                 case IFM_25G_KR_S:
1185                 case IFM_25G_AUI:
1186                 case IFM_25G_KR1:
1187                         key = IFM_25G_PCIE;
1188                         break;
1189                 case IFM_40G_CR4:
1190                 case IFM_40G_SR4:
1191                 case IFM_40G_LR4:
1192                 case IFM_40G_XLPPI:
1193                 case IFM_40G_KR4:
1194                 case IFM_40G_XLAUI:
1195                 case IFM_40G_XLAUI_AC:
1196                 case IFM_40G_ER4:
1197                         key = IFM_40G_CR4;
1198                         break;
1199                 case IFM_100G_CR4:
1200                 case IFM_100G_SR4:
1201                 case IFM_100G_KR4:
1202                 case IFM_100G_LR4:
1203                 case IFM_100G_CAUI4_AC:
1204                 case IFM_100G_CAUI4:
1205                 case IFM_100G_AUI4_AC:
1206                 case IFM_100G_AUI4:
1207                 case IFM_100G_CR_PAM4:
1208                 case IFM_100G_KR_PAM4:
1209                 case IFM_100G_CP2:
1210                 case IFM_100G_SR2:
1211                 case IFM_100G_DR:
1212                 case IFM_100G_KR2_PAM4:
1213                 case IFM_100G_CAUI2_AC:
1214                 case IFM_100G_CAUI2:
1215                 case IFM_100G_AUI2_AC:
1216                 case IFM_100G_AUI2:
1217                         key = IFM_100G_CR4;
1218                         break;
1219                 case IFM_200G_CR4_PAM4:
1220                 case IFM_200G_SR4:
1221                 case IFM_200G_FR4:
1222                 case IFM_200G_LR4:
1223                 case IFM_200G_DR4:
1224                 case IFM_200G_KR4_PAM4:
1225                 case IFM_200G_AUI4_AC:
1226                 case IFM_200G_AUI4:
1227                 case IFM_200G_AUI8_AC:
1228                 case IFM_200G_AUI8:
1229                         key = IFM_200G_CR4_PAM4;
1230                         break;
1231                 case IFM_400G_FR8:
1232                 case IFM_400G_LR8:
1233                 case IFM_400G_DR4:
1234                 case IFM_400G_AUI8_AC:
1235                 case IFM_400G_AUI8:
1236                         key = IFM_400G_FR8;
1237                         break;
1238                 default:
1239                         key = subtype;
1240                         break;
1241                 }
1242                 /* bit 5..14:   (some bits of) if_index of lagg device */
1243                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1244                 /* bit 15:      0 */
1245         }
1246         return (htons(key));
1247 }
1248
1249 static void
1250 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1251 {
1252         char buf[LACP_LAGIDSTR_MAX+1];
1253
1254         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1255             __func__,
1256             lacp_format_lagid(&la->la_actor, &la->la_partner,
1257             buf, sizeof(buf)),
1258             la->la_refcnt, la->la_refcnt + 1));
1259
1260         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1261         la->la_refcnt++;
1262         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1263 }
1264
1265 static void
1266 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1267 {
1268         char buf[LACP_LAGIDSTR_MAX+1];
1269
1270         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1271             __func__,
1272             lacp_format_lagid(&la->la_actor, &la->la_partner,
1273             buf, sizeof(buf)),
1274             la->la_refcnt, la->la_refcnt - 1));
1275
1276         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1277         la->la_refcnt--;
1278         if (la->la_refcnt > 0) {
1279                 return;
1280         }
1281
1282         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1283         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1284
1285         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1286
1287         free(la, M_DEVBUF);
1288 }
1289
1290 /*
1291  * lacp_aggregator_get: allocate an aggregator.
1292  */
1293
1294 static struct lacp_aggregator *
1295 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1296 {
1297         struct lacp_aggregator *la;
1298
1299         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1300         if (la) {
1301                 la->la_refcnt = 1;
1302                 la->la_nports = 0;
1303                 TAILQ_INIT(&la->la_ports);
1304                 la->la_pending = 0;
1305                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1306         }
1307
1308         return (la);
1309 }
1310
1311 /*
1312  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1313  */
1314
1315 static void
1316 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1317 {
1318         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1319         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1320
1321         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1322 }
1323
1324 static void
1325 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1326     const struct lacp_peerinfo *lpi_port)
1327 {
1328         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1329         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1330         lpi_aggr->lip_key = lpi_port->lip_key;
1331 }
1332
1333 /*
1334  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1335  */
1336
1337 static int
1338 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1339     const struct lacp_port *lp)
1340 {
1341         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1342             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1343                 return (0);
1344         }
1345
1346         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1347                 return (0);
1348         }
1349
1350         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1351                 return (0);
1352         }
1353
1354         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1355                 return (0);
1356         }
1357
1358         return (1);
1359 }
1360
1361 static int
1362 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1363     const struct lacp_peerinfo *b)
1364 {
1365         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1366             sizeof(a->lip_systemid))) {
1367                 return (0);
1368         }
1369
1370         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1371                 return (0);
1372         }
1373
1374         return (1);
1375 }
1376
1377 static void
1378 lacp_port_enable(struct lacp_port *lp)
1379 {
1380         lp->lp_state |= LACP_STATE_AGGREGATION;
1381 }
1382
1383 static void
1384 lacp_port_disable(struct lacp_port *lp)
1385 {
1386         lacp_set_mux(lp, LACP_MUX_DETACHED);
1387
1388         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1389         lp->lp_selected = LACP_UNSELECTED;
1390         lacp_sm_rx_record_default(lp);
1391         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1392         lp->lp_state &= ~LACP_STATE_EXPIRED;
1393 }
1394
1395 /*
1396  * lacp_select: select an aggregator.  create one if necessary.
1397  */
1398 static void
1399 lacp_select(struct lacp_port *lp)
1400 {
1401         struct lacp_softc *lsc = lp->lp_lsc;
1402         struct lacp_aggregator *la;
1403         char buf[LACP_LAGIDSTR_MAX+1];
1404
1405         if (lp->lp_aggregator) {
1406                 return;
1407         }
1408
1409         /* If we haven't heard from our peer, skip this step. */
1410         if (lp->lp_state & LACP_STATE_DEFAULTED)
1411                 return;
1412
1413         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1414             ("timer_wait_while still active"));
1415
1416         LACP_DPRINTF((lp, "port lagid=%s\n",
1417             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1418             buf, sizeof(buf))));
1419
1420         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1421                 if (lacp_aggregator_is_compatible(la, lp)) {
1422                         break;
1423                 }
1424         }
1425
1426         if (la == NULL) {
1427                 la = lacp_aggregator_get(lsc, lp);
1428                 if (la == NULL) {
1429                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1430
1431                         /*
1432                          * will retry on the next tick.
1433                          */
1434
1435                         return;
1436                 }
1437                 lacp_fill_aggregator_id(la, lp);
1438                 LACP_DPRINTF((lp, "aggregator created\n"));
1439         } else {
1440                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1441                 if (la->la_refcnt == LACP_MAX_PORTS)
1442                         return;
1443                 lacp_aggregator_addref(lsc, la);
1444         }
1445
1446         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1447             lacp_format_lagid(&la->la_actor, &la->la_partner,
1448             buf, sizeof(buf))));
1449
1450         lp->lp_aggregator = la;
1451         lp->lp_selected = LACP_SELECTED;
1452 }
1453
1454 /*
1455  * lacp_unselect: finish unselect/detach process.
1456  */
1457
1458 static void
1459 lacp_unselect(struct lacp_port *lp)
1460 {
1461         struct lacp_softc *lsc = lp->lp_lsc;
1462         struct lacp_aggregator *la = lp->lp_aggregator;
1463
1464         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1465             ("timer_wait_while still active"));
1466
1467         if (la == NULL) {
1468                 return;
1469         }
1470
1471         lp->lp_aggregator = NULL;
1472         lacp_aggregator_delref(lsc, la);
1473 }
1474
1475 /* mux machine */
1476
1477 static void
1478 lacp_sm_mux(struct lacp_port *lp)
1479 {
1480         struct lagg_port *lgp = lp->lp_lagg;
1481         struct lagg_softc *sc = lgp->lp_softc;
1482         enum lacp_mux_state new_state;
1483         boolean_t p_sync =
1484                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1485         boolean_t p_collecting =
1486             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1487         enum lacp_selected selected = lp->lp_selected;
1488         struct lacp_aggregator *la;
1489
1490         if (V_lacp_debug > 1)
1491                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1492                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1493                     lp->lp_mux_state, selected, p_sync, p_collecting);
1494
1495 re_eval:
1496         la = lp->lp_aggregator;
1497         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1498             ("MUX not detached"));
1499         new_state = lp->lp_mux_state;
1500         switch (lp->lp_mux_state) {
1501         case LACP_MUX_DETACHED:
1502                 if (selected != LACP_UNSELECTED) {
1503                         new_state = LACP_MUX_WAITING;
1504                 }
1505                 break;
1506         case LACP_MUX_WAITING:
1507                 KASSERT(la->la_pending > 0 ||
1508                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1509                     ("timer_wait_while still active"));
1510                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1511                         new_state = LACP_MUX_ATTACHED;
1512                 } else if (selected == LACP_UNSELECTED) {
1513                         new_state = LACP_MUX_DETACHED;
1514                 }
1515                 break;
1516         case LACP_MUX_ATTACHED:
1517                 if (selected == LACP_SELECTED && p_sync) {
1518                         new_state = LACP_MUX_COLLECTING;
1519                 } else if (selected != LACP_SELECTED) {
1520                         new_state = LACP_MUX_DETACHED;
1521                 }
1522                 break;
1523         case LACP_MUX_COLLECTING:
1524                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1525                         new_state = LACP_MUX_DISTRIBUTING;
1526                 } else if (selected != LACP_SELECTED || !p_sync) {
1527                         new_state = LACP_MUX_ATTACHED;
1528                 }
1529                 break;
1530         case LACP_MUX_DISTRIBUTING:
1531                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1532                         new_state = LACP_MUX_COLLECTING;
1533                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1534                         sc->sc_flapping++;
1535                 }
1536                 break;
1537         default:
1538                 panic("%s: unknown state", __func__);
1539         }
1540
1541         if (lp->lp_mux_state == new_state) {
1542                 return;
1543         }
1544
1545         lacp_set_mux(lp, new_state);
1546         goto re_eval;
1547 }
1548
1549 static void
1550 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1551 {
1552         struct lacp_aggregator *la = lp->lp_aggregator;
1553
1554         if (lp->lp_mux_state == new_state) {
1555                 return;
1556         }
1557
1558         switch (new_state) {
1559         case LACP_MUX_DETACHED:
1560                 lp->lp_state &= ~LACP_STATE_SYNC;
1561                 lacp_disable_distributing(lp);
1562                 lacp_disable_collecting(lp);
1563                 lacp_sm_assert_ntt(lp);
1564                 /* cancel timer */
1565                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1566                         KASSERT(la->la_pending > 0,
1567                             ("timer_wait_while not active"));
1568                         la->la_pending--;
1569                 }
1570                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1571                 lacp_unselect(lp);
1572                 break;
1573         case LACP_MUX_WAITING:
1574                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1575                     LACP_AGGREGATE_WAIT_TIME);
1576                 la->la_pending++;
1577                 break;
1578         case LACP_MUX_ATTACHED:
1579                 lp->lp_state |= LACP_STATE_SYNC;
1580                 lacp_disable_collecting(lp);
1581                 lacp_sm_assert_ntt(lp);
1582                 break;
1583         case LACP_MUX_COLLECTING:
1584                 lacp_enable_collecting(lp);
1585                 lacp_disable_distributing(lp);
1586                 lacp_sm_assert_ntt(lp);
1587                 break;
1588         case LACP_MUX_DISTRIBUTING:
1589                 lacp_enable_distributing(lp);
1590                 break;
1591         default:
1592                 panic("%s: unknown state", __func__);
1593         }
1594
1595         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1596
1597         lp->lp_mux_state = new_state;
1598 }
1599
1600 static void
1601 lacp_sm_mux_timer(struct lacp_port *lp)
1602 {
1603         struct lacp_aggregator *la = lp->lp_aggregator;
1604         char buf[LACP_LAGIDSTR_MAX+1];
1605
1606         KASSERT(la->la_pending > 0, ("no pending event"));
1607
1608         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1609             lacp_format_lagid(&la->la_actor, &la->la_partner,
1610             buf, sizeof(buf)),
1611             la->la_pending, la->la_pending - 1));
1612
1613         la->la_pending--;
1614 }
1615
1616 /* periodic transmit machine */
1617
1618 static void
1619 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1620 {
1621         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1622             LACP_STATE_TIMEOUT)) {
1623                 return;
1624         }
1625
1626         LACP_DPRINTF((lp, "partner timeout changed\n"));
1627
1628         /*
1629          * FAST_PERIODIC -> SLOW_PERIODIC
1630          * or
1631          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1632          *
1633          * let lacp_sm_ptx_tx_schedule to update timeout.
1634          */
1635
1636         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1637
1638         /*
1639          * if timeout has been shortened, assert NTT.
1640          */
1641
1642         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1643                 lacp_sm_assert_ntt(lp);
1644         }
1645 }
1646
1647 static void
1648 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1649 {
1650         int timeout;
1651
1652         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1653             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1654
1655                 /*
1656                  * NO_PERIODIC
1657                  */
1658
1659                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1660                 return;
1661         }
1662
1663         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1664                 return;
1665         }
1666
1667         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1668             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1669
1670         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1671 }
1672
1673 static void
1674 lacp_sm_ptx_timer(struct lacp_port *lp)
1675 {
1676         lacp_sm_assert_ntt(lp);
1677 }
1678
1679 static void
1680 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1681 {
1682         int timeout;
1683
1684         /*
1685          * check LACP_DISABLED first
1686          */
1687
1688         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1689                 return;
1690         }
1691
1692         /*
1693          * check loopback condition.
1694          */
1695
1696         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1697             &lp->lp_actor.lip_systemid)) {
1698                 return;
1699         }
1700
1701         /*
1702          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1703          */
1704
1705         lacp_sm_rx_update_selected(lp, du);
1706         lacp_sm_rx_update_ntt(lp, du);
1707         lacp_sm_rx_record_pdu(lp, du);
1708
1709         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1710             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1711         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1712
1713         lp->lp_state &= ~LACP_STATE_EXPIRED;
1714
1715         /*
1716          * kick transmit machine without waiting the next tick.
1717          */
1718
1719         lacp_sm_tx(lp);
1720 }
1721
1722 static void
1723 lacp_sm_rx_set_expired(struct lacp_port *lp)
1724 {
1725         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1726         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1727         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1728         lp->lp_state |= LACP_STATE_EXPIRED;
1729 }
1730
1731 static void
1732 lacp_sm_rx_timer(struct lacp_port *lp)
1733 {
1734         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1735                 /* CURRENT -> EXPIRED */
1736                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1737                 lacp_sm_rx_set_expired(lp);
1738         } else {
1739                 /* EXPIRED -> DEFAULTED */
1740                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1741                 lacp_sm_rx_update_default_selected(lp);
1742                 lacp_sm_rx_record_default(lp);
1743                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1744         }
1745 }
1746
1747 static void
1748 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1749 {
1750         boolean_t active;
1751         uint8_t oldpstate;
1752         char buf[LACP_STATESTR_MAX+1];
1753
1754         LACP_TRACE(lp);
1755
1756         oldpstate = lp->lp_partner.lip_state;
1757
1758         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1759             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1760             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1761
1762         lp->lp_partner = du->ldu_actor;
1763         if (active &&
1764             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1765             LACP_STATE_AGGREGATION) &&
1766             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1767             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1768                 /*
1769                  * XXX Maintain legacy behavior of leaving the
1770                  * LACP_STATE_SYNC bit unchanged from the partner's
1771                  * advertisement if lsc_strict_mode is false.
1772                  * TODO: We should re-examine the concept of the "strict mode"
1773                  * to ensure it makes sense to maintain a non-strict mode.
1774                  */
1775                 if (lp->lp_lsc->lsc_strict_mode)
1776                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1777         } else {
1778                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1779         }
1780
1781         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1782
1783         if (oldpstate != lp->lp_partner.lip_state) {
1784                 LACP_DPRINTF((lp, "old pstate %s\n",
1785                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1786                 LACP_DPRINTF((lp, "new pstate %s\n",
1787                     lacp_format_state(lp->lp_partner.lip_state, buf,
1788                     sizeof(buf))));
1789         }
1790
1791         lacp_sm_ptx_update_timeout(lp, oldpstate);
1792 }
1793
1794 static void
1795 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1796 {
1797
1798         LACP_TRACE(lp);
1799
1800         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1801             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1802             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1803                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1804                 lacp_sm_assert_ntt(lp);
1805         }
1806 }
1807
1808 static void
1809 lacp_sm_rx_record_default(struct lacp_port *lp)
1810 {
1811         uint8_t oldpstate;
1812
1813         LACP_TRACE(lp);
1814
1815         oldpstate = lp->lp_partner.lip_state;
1816         if (lp->lp_lsc->lsc_strict_mode)
1817                 lp->lp_partner = lacp_partner_admin_strict;
1818         else
1819                 lp->lp_partner = lacp_partner_admin_optimistic;
1820         lp->lp_state |= LACP_STATE_DEFAULTED;
1821         lacp_sm_ptx_update_timeout(lp, oldpstate);
1822 }
1823
1824 static void
1825 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1826     const struct lacp_peerinfo *info)
1827 {
1828
1829         LACP_TRACE(lp);
1830
1831         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1832             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1833             LACP_STATE_AGGREGATION)) {
1834                 lp->lp_selected = LACP_UNSELECTED;
1835                 /* mux machine will clean up lp->lp_aggregator */
1836         }
1837 }
1838
1839 static void
1840 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1841 {
1842
1843         LACP_TRACE(lp);
1844
1845         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1846 }
1847
1848 static void
1849 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1850 {
1851
1852         LACP_TRACE(lp);
1853
1854         if (lp->lp_lsc->lsc_strict_mode)
1855                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1856                     &lacp_partner_admin_strict);
1857         else
1858                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1859                     &lacp_partner_admin_optimistic);
1860 }
1861
1862 /* transmit machine */
1863
1864 static void
1865 lacp_sm_tx(struct lacp_port *lp)
1866 {
1867         int error = 0;
1868
1869         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1870 #if 1
1871             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1872             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1873 #endif
1874             ) {
1875                 lp->lp_flags &= ~LACP_PORT_NTT;
1876         }
1877
1878         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1879                 return;
1880         }
1881
1882         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1883         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1884                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1885                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1886                 return;
1887         }
1888
1889         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1890                 error = lacp_xmit_lacpdu(lp);
1891         } else {
1892                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1893         }
1894
1895         if (error == 0) {
1896                 lp->lp_flags &= ~LACP_PORT_NTT;
1897         } else {
1898                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1899                     error));
1900         }
1901 }
1902
1903 static void
1904 lacp_sm_assert_ntt(struct lacp_port *lp)
1905 {
1906
1907         lp->lp_flags |= LACP_PORT_NTT;
1908 }
1909
1910 static void
1911 lacp_run_timers(struct lacp_port *lp)
1912 {
1913         int i;
1914
1915         for (i = 0; i < LACP_NTIMER; i++) {
1916                 KASSERT(lp->lp_timer[i] >= 0,
1917                     ("invalid timer value %d", lp->lp_timer[i]));
1918                 if (lp->lp_timer[i] == 0) {
1919                         continue;
1920                 } else if (--lp->lp_timer[i] <= 0) {
1921                         if (lacp_timer_funcs[i]) {
1922                                 (*lacp_timer_funcs[i])(lp);
1923                         }
1924                 }
1925         }
1926 }
1927
1928 int
1929 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1930 {
1931         struct lacp_softc *lsc = lp->lp_lsc;
1932         struct lagg_port *lgp = lp->lp_lagg;
1933         struct lacp_port *lp2;
1934         struct markerdu *mdu;
1935         int error = 0;
1936         int pending = 0;
1937
1938         if (m->m_pkthdr.len != sizeof(*mdu)) {
1939                 goto bad;
1940         }
1941
1942         if ((m->m_flags & M_MCAST) == 0) {
1943                 goto bad;
1944         }
1945
1946         if (m->m_len < sizeof(*mdu)) {
1947                 m = m_pullup(m, sizeof(*mdu));
1948                 if (m == NULL) {
1949                         return (ENOMEM);
1950                 }
1951         }
1952
1953         mdu = mtod(m, struct markerdu *);
1954
1955         if (memcmp(&mdu->mdu_eh.ether_dhost,
1956             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1957                 goto bad;
1958         }
1959
1960         if (mdu->mdu_sph.sph_version != 1) {
1961                 goto bad;
1962         }
1963
1964         switch (mdu->mdu_tlv.tlv_type) {
1965         case MARKER_TYPE_INFO:
1966                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1967                     marker_info_tlv_template, TRUE)) {
1968                         goto bad;
1969                 }
1970                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1971                 memcpy(&mdu->mdu_eh.ether_dhost,
1972                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1973                 memcpy(&mdu->mdu_eh.ether_shost,
1974                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1975                 error = lagg_enqueue(lp->lp_ifp, m);
1976                 break;
1977
1978         case MARKER_TYPE_RESPONSE:
1979                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1980                     marker_response_tlv_template, TRUE)) {
1981                         goto bad;
1982                 }
1983                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1984                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1985                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1986
1987                 /* Verify that it is the last marker we sent out */
1988                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1989                     sizeof(struct lacp_markerinfo)))
1990                         goto bad;
1991
1992                 LACP_LOCK(lsc);
1993                 lp->lp_flags &= ~LACP_PORT_MARK;
1994
1995                 if (lsc->lsc_suppress_distributing) {
1996                         /* Check if any ports are waiting for a response */
1997                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1998                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1999                                         pending = 1;
2000                                         break;
2001                                 }
2002                         }
2003
2004                         if (pending == 0) {
2005                                 /* All interface queues are clear */
2006                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
2007                                 lsc->lsc_suppress_distributing = FALSE;
2008                         }
2009                 }
2010                 LACP_UNLOCK(lsc);
2011                 m_freem(m);
2012                 break;
2013
2014         default:
2015                 goto bad;
2016         }
2017
2018         return (error);
2019
2020 bad:
2021         LACP_DPRINTF((lp, "bad marker frame\n"));
2022         m_freem(m);
2023         return (EINVAL);
2024 }
2025
2026 static int
2027 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
2028     const struct tlv_template *tmpl, boolean_t check_type)
2029 {
2030         while (/* CONSTCOND */ 1) {
2031                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
2032                         return (EINVAL);
2033                 }
2034                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
2035                     tlv->tlv_length != tmpl->tmpl_length) {
2036                         return (EINVAL);
2037                 }
2038                 if (tmpl->tmpl_type == 0) {
2039                         break;
2040                 }
2041                 tlv = (const struct tlvhdr *)
2042                     ((const char *)tlv + tlv->tlv_length);
2043                 tmpl++;
2044         }
2045
2046         return (0);
2047 }
2048
2049 /* Debugging */
2050 const char *
2051 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
2052 {
2053         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
2054             (int)mac[0],
2055             (int)mac[1],
2056             (int)mac[2],
2057             (int)mac[3],
2058             (int)mac[4],
2059             (int)mac[5]);
2060
2061         return (buf);
2062 }
2063
2064 const char *
2065 lacp_format_systemid(const struct lacp_systemid *sysid,
2066     char *buf, size_t buflen)
2067 {
2068         char macbuf[LACP_MACSTR_MAX+1];
2069
2070         snprintf(buf, buflen, "%04X,%s",
2071             ntohs(sysid->lsi_prio),
2072             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2073
2074         return (buf);
2075 }
2076
2077 const char *
2078 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2079 {
2080         snprintf(buf, buflen, "%04X,%04X",
2081             ntohs(portid->lpi_prio),
2082             ntohs(portid->lpi_portno));
2083
2084         return (buf);
2085 }
2086
2087 const char *
2088 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2089 {
2090         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2091         char portid[LACP_PORTIDSTR_MAX+1];
2092
2093         snprintf(buf, buflen, "(%s,%04X,%s)",
2094             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2095             ntohs(peer->lip_key),
2096             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2097
2098         return (buf);
2099 }
2100
2101 const char *
2102 lacp_format_lagid(const struct lacp_peerinfo *a,
2103     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2104 {
2105         char astr[LACP_PARTNERSTR_MAX+1];
2106         char bstr[LACP_PARTNERSTR_MAX+1];
2107
2108 #if 0
2109         /*
2110          * there's a convention to display small numbered peer
2111          * in the left.
2112          */
2113
2114         if (lacp_compare_peerinfo(a, b) > 0) {
2115                 const struct lacp_peerinfo *t;
2116
2117                 t = a;
2118                 a = b;
2119                 b = t;
2120         }
2121 #endif
2122
2123         snprintf(buf, buflen, "[%s,%s]",
2124             lacp_format_partner(a, astr, sizeof(astr)),
2125             lacp_format_partner(b, bstr, sizeof(bstr)));
2126
2127         return (buf);
2128 }
2129
2130 const char *
2131 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2132     char *buf, size_t buflen)
2133 {
2134         if (la == NULL) {
2135                 return ("(none)");
2136         }
2137
2138         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2139 }
2140
2141 const char *
2142 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2143 {
2144         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2145         return (buf);
2146 }
2147
2148 static void
2149 lacp_dump_lacpdu(const struct lacpdu *du)
2150 {
2151         char buf[LACP_PARTNERSTR_MAX+1];
2152         char buf2[LACP_STATESTR_MAX+1];
2153
2154         printf("actor=%s\n",
2155             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2156         printf("actor.state=%s\n",
2157             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2158         printf("partner=%s\n",
2159             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2160         printf("partner.state=%s\n",
2161             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2162
2163         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2164 }
2165
2166 static void
2167 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2168 {
2169         va_list va;
2170
2171         if (lp) {
2172                 printf("%s: ", lp->lp_ifp->if_xname);
2173         }
2174
2175         va_start(va, fmt);
2176         vprintf(fmt, va);
2177         va_end(va);
2178 }