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 }
 715
 716 static void
 717 lacp_enable_distributing(struct lacp_port *lp)
 718 {
 719         struct lacp_aggregator *la = lp->lp_aggregator;
 720         struct lacp_softc *lsc = lp->lp_lsc;
 721         struct lagg_softc *sc = lsc->lsc_softc;
 722         char buf[LACP_LAGIDSTR_MAX+1];
 723
 724         LACP_LOCK_ASSERT(lsc);
 725
 726         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 727                 return;
 728         }
 729
 730         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 731             "nports %d -> %d\n",
 732             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 733             la->la_nports, la->la_nports + 1));
 734
 735         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 736         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 737         la->la_nports++;
 738         sc->sc_active = la->la_nports;
 739
 740         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 741
 742         if (lsc->lsc_active_aggregator == la) {
 743                 lacp_suppress_distributing(lsc, la);
 744                 lacp_update_portmap(lsc);
 745         } else
 746                 /* try to become the active aggregator */
 747                 lacp_select_active_aggregator(lsc);
 748 }
 749
 750 static void
 751 lacp_transit_expire(void *vp)
 752 {
 753         struct lacp_softc *lsc = vp;
 754
 755         LACP_LOCK_ASSERT(lsc);
 756
 757         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 758         LACP_TRACE(NULL);
 759         CURVNET_RESTORE();
 760
 761         lsc->lsc_suppress_distributing = FALSE;
 762 }
 763
 764 void
 765 lacp_attach(struct lagg_softc *sc)
 766 {
 767         struct lacp_softc *lsc;
 768
 769         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 770
 771         sc->sc_psc = lsc;
 772         lsc->lsc_softc = sc;
 773
 774         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 775         lsc->lsc_active_aggregator = NULL;
 776         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 777         LACP_LOCK_INIT(lsc);
 778         TAILQ_INIT(&lsc->lsc_aggregators);
 779         LIST_INIT(&lsc->lsc_ports);
 780
 781         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 782         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 783
 784         /* if the lagg is already up then do the same */
 785         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 786                 lacp_init(sc);
 787 }
 788
 789 void
 790 lacp_detach(void *psc)
 791 {
 792         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 793
 794         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 795             ("aggregators still active"));
 796         KASSERT(lsc->lsc_active_aggregator == NULL,
 797             ("aggregator still attached"));
 798
 799         callout_drain(&lsc->lsc_transit_callout);
 800         callout_drain(&lsc->lsc_callout);
 801
 802         LACP_LOCK_DESTROY(lsc);
 803         free(lsc, M_DEVBUF);
 804 }
 805
 806 void
 807 lacp_init(struct lagg_softc *sc)
 808 {
 809         struct lacp_softc *lsc = LACP_SOFTC(sc);
 810
 811         LACP_LOCK(lsc);
 812         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 813         LACP_UNLOCK(lsc);
 814 }
 815
 816 void
 817 lacp_stop(struct lagg_softc *sc)
 818 {
 819         struct lacp_softc *lsc = LACP_SOFTC(sc);
 820
 821         LACP_LOCK(lsc);
 822         callout_stop(&lsc->lsc_transit_callout);
 823         callout_stop(&lsc->lsc_callout);
 824         LACP_UNLOCK(lsc);
 825 }
 826
 827 struct lagg_port *
 828 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 829 {
 830         struct lacp_softc *lsc = LACP_SOFTC(sc);
 831         struct lacp_portmap *pm;
 832         struct lacp_port *lp;
 833         uint32_t hash;
 834
 835         if (__predict_false(lsc->lsc_suppress_distributing)) {
 836                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 837                 return (NULL);
 838         }
 839
 840         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 841         if (pm->pm_count == 0) {
 842                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 843                 return (NULL);
 844         }
 845
 846         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 847             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 848                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 849         else
 850                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 851         hash %= pm->pm_count;
 852         lp = pm->pm_map[hash];
 853
 854         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 855             ("aggregated port is not distributing"));
 856
 857         return (lp->lp_lagg);
 858 }
 859
 860 #ifdef RATELIMIT
 861 struct lagg_port *
 862 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t flowid)
 863 {
 864         struct lacp_softc *lsc = LACP_SOFTC(sc);
 865         struct lacp_portmap *pm;
 866         struct lacp_port *lp;
 867         uint32_t hash;
 868
 869         if (__predict_false(lsc->lsc_suppress_distributing)) {
 870                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 871                 return (NULL);
 872         }
 873
 874         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 875         if (pm->pm_count == 0) {
 876                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 877                 return (NULL);
 878         }
 879
 880         hash = flowid >> sc->flowid_shift;
 881         hash %= pm->pm_count;
 882         lp = pm->pm_map[hash];
 883
 884         return (lp->lp_lagg);
 885 }
 886 #endif
 887
 888 /*
 889  * lacp_suppress_distributing: drop transmit packets for a while
 890  * to preserve packet ordering.
 891  */
 892
 893 static void
 894 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 895 {
 896         struct lacp_port *lp;
 897
 898         if (lsc->lsc_active_aggregator != la) {
 899                 return;
 900         }
 901
 902         LACP_TRACE(NULL);
 903
 904         lsc->lsc_suppress_distributing = TRUE;
 905
 906         /* send a marker frame down each port to verify the queues are empty */
 907         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 908                 lp->lp_flags |= LACP_PORT_MARK;
 909                 lacp_xmit_marker(lp);
 910         }
 911
 912         /* set a timeout for the marker frames */
 913         callout_reset(&lsc->lsc_transit_callout,
 914             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 915 }
 916
 917 static int
 918 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 919     const struct lacp_peerinfo *b)
 920 {
 921         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 922 }
 923
 924 static int
 925 lacp_compare_systemid(const struct lacp_systemid *a,
 926     const struct lacp_systemid *b)
 927 {
 928         return (memcmp(a, b, sizeof(*a)));
 929 }
 930
 931 #if 0   /* unused */
 932 static int
 933 lacp_compare_portid(const struct lacp_portid *a,
 934     const struct lacp_portid *b)
 935 {
 936         return (memcmp(a, b, sizeof(*a)));
 937 }
 938 #endif
 939
 940 static uint64_t
 941 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 942 {
 943         struct lacp_port *lp;
 944         uint64_t speed;
 945
 946         lp = TAILQ_FIRST(&la->la_ports);
 947         if (lp == NULL) {
 948                 return (0);
 949         }
 950
 951         speed = ifmedia_baudrate(lp->lp_media);
 952         speed *= la->la_nports;
 953         if (speed == 0) {
 954                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 955                     lp->lp_media, la->la_nports));
 956         }
 957
 958         return (speed);
 959 }
 960
 961 /*
 962  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 963  * packets from lagg(4) interface.
 964  */
 965
 966 static void
 967 lacp_select_active_aggregator(struct lacp_softc *lsc)
 968 {
 969         struct lacp_aggregator *la;
 970         struct lacp_aggregator *best_la = NULL;
 971         uint64_t best_speed = 0;
 972         char buf[LACP_LAGIDSTR_MAX+1];
 973
 974         LACP_TRACE(NULL);
 975
 976         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 977                 uint64_t speed;
 978
 979                 if (la->la_nports == 0) {
 980                         continue;
 981                 }
 982
 983                 speed = lacp_aggregator_bandwidth(la);
 984                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 985                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 986                     speed, la->la_nports));
 987
 988                 /*
 989                  * This aggregator is chosen if the partner has a better
 990                  * system priority or, the total aggregated speed is higher
 991                  * or, it is already the chosen aggregator
 992                  */
 993                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
 994                     LACP_SYS_PRI(best_la->la_partner)) ||
 995                     speed > best_speed ||
 996                     (speed == best_speed &&
 997                     la == lsc->lsc_active_aggregator)) {
 998                         best_la = la;
 999                         best_speed = speed;
1000                 }
1001         }
1002
1003         KASSERT(best_la == NULL || best_la->la_nports > 0,
1004             ("invalid aggregator refcnt"));
1005         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1006             ("invalid aggregator list"));
1007
1008         if (lsc->lsc_active_aggregator != best_la) {
1009                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1010                 LACP_DPRINTF((NULL, "old %s\n",
1011                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1012                     buf, sizeof(buf))));
1013         } else {
1014                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1015         }
1016         LACP_DPRINTF((NULL, "new %s\n",
1017             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1018
1019         if (lsc->lsc_active_aggregator != best_la) {
1020                 lsc->lsc_active_aggregator = best_la;
1021                 lacp_update_portmap(lsc);
1022                 if (best_la) {
1023                         lacp_suppress_distributing(lsc, best_la);
1024                 }
1025         }
1026 }
1027
1028 /*
1029  * Updated the inactive portmap array with the new list of ports and
1030  * make it live.
1031  */
1032 static void
1033 lacp_update_portmap(struct lacp_softc *lsc)
1034 {
1035         struct lagg_softc *sc = lsc->lsc_softc;
1036         struct lacp_aggregator *la;
1037         struct lacp_portmap *p;
1038         struct lacp_port *lp;
1039         uint64_t speed;
1040         u_int newmap;
1041         int i;
1042
1043         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1044         p = &lsc->lsc_pmap[newmap];
1045         la = lsc->lsc_active_aggregator;
1046         speed = 0;
1047         bzero(p, sizeof(struct lacp_portmap));
1048
1049         if (la != NULL && la->la_nports > 0) {
1050                 p->pm_count = la->la_nports;
1051                 i = 0;
1052                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q)
1053                         p->pm_map[i++] = lp;
1054                 KASSERT(i == p->pm_count, ("Invalid port count"));
1055                 speed = lacp_aggregator_bandwidth(la);
1056         }
1057         sc->sc_ifp->if_baudrate = speed;
1058
1059         /* switch the active portmap over */
1060         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1061         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1062                     lsc->lsc_activemap,
1063                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1064 }
1065
1066 static uint16_t
1067 lacp_compose_key(struct lacp_port *lp)
1068 {
1069         struct lagg_port *lgp = lp->lp_lagg;
1070         struct lagg_softc *sc = lgp->lp_softc;
1071         u_int media = lp->lp_media;
1072         uint16_t key;
1073
1074         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1075
1076                 /*
1077                  * non-aggregatable links should have unique keys.
1078                  *
1079                  * XXX this isn't really unique as if_index is 16 bit.
1080                  */
1081
1082                 /* bit 0..14:   (some bits of) if_index of this port */
1083                 key = lp->lp_ifp->if_index;
1084                 /* bit 15:      1 */
1085                 key |= 0x8000;
1086         } else {
1087                 u_int subtype = IFM_SUBTYPE(media);
1088
1089                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1090                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1091
1092                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1093                 switch (subtype) {
1094                 case IFM_10_T:
1095                 case IFM_10_2:
1096                 case IFM_10_5:
1097                 case IFM_10_STP:
1098                 case IFM_10_FL:
1099                         key = IFM_10_T;
1100                         break;
1101                 case IFM_100_TX:
1102                 case IFM_100_FX:
1103                 case IFM_100_T4:
1104                 case IFM_100_VG:
1105                 case IFM_100_T2:
1106                 case IFM_100_T:
1107                         key = IFM_100_TX;
1108                         break;
1109                 case IFM_1000_SX:
1110                 case IFM_1000_LX:
1111                 case IFM_1000_CX:
1112                 case IFM_1000_T:
1113                 case IFM_1000_KX:
1114                 case IFM_1000_SGMII:
1115                 case IFM_1000_CX_SGMII:
1116                         key = IFM_1000_SX;
1117                         break;
1118                 case IFM_10G_LR:
1119                 case IFM_10G_SR:
1120                 case IFM_10G_CX4:
1121                 case IFM_10G_TWINAX:
1122                 case IFM_10G_TWINAX_LONG:
1123                 case IFM_10G_LRM:
1124                 case IFM_10G_T:
1125                 case IFM_10G_KX4:
1126                 case IFM_10G_KR:
1127                 case IFM_10G_CR1:
1128                 case IFM_10G_ER:
1129                 case IFM_10G_SFI:
1130                 case IFM_10G_AOC:
1131                         key = IFM_10G_LR;
1132                         break;
1133                 case IFM_20G_KR2:
1134                         key = IFM_20G_KR2;
1135                         break;
1136                 case IFM_2500_KX:
1137                 case IFM_2500_T:
1138                         key = IFM_2500_KX;
1139                         break;
1140                 case IFM_5000_T:
1141                         key = IFM_5000_T;
1142                         break;
1143                 case IFM_50G_PCIE:
1144                 case IFM_50G_CR2:
1145                 case IFM_50G_KR2:
1146                         key = IFM_50G_PCIE;
1147                         break;
1148                 case IFM_56G_R4:
1149                         key = IFM_56G_R4;
1150                         break;
1151                 case IFM_25G_PCIE:
1152                 case IFM_25G_CR:
1153                 case IFM_25G_KR:
1154                 case IFM_25G_SR:
1155                 case IFM_25G_LR:
1156                 case IFM_25G_ACC:
1157                 case IFM_25G_AOC:
1158                         key = IFM_25G_PCIE;
1159                         break;
1160                 case IFM_40G_CR4:
1161                 case IFM_40G_SR4:
1162                 case IFM_40G_LR4:
1163                 case IFM_40G_XLPPI:
1164                 case IFM_40G_KR4:
1165                         key = IFM_40G_CR4;
1166                         break;
1167                 case IFM_100G_CR4:
1168                 case IFM_100G_SR4:
1169                 case IFM_100G_KR4:
1170                 case IFM_100G_LR4:
1171                         key = IFM_100G_CR4;
1172                         break;
1173                 default:
1174                         key = subtype;
1175                         break;
1176                 }
1177                 /* bit 5..14:   (some bits of) if_index of lagg device */
1178                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1179                 /* bit 15:      0 */
1180         }
1181         return (htons(key));
1182 }
1183
1184 static void
1185 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1186 {
1187         char buf[LACP_LAGIDSTR_MAX+1];
1188
1189         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1190             __func__,
1191             lacp_format_lagid(&la->la_actor, &la->la_partner,
1192             buf, sizeof(buf)),
1193             la->la_refcnt, la->la_refcnt + 1));
1194
1195         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1196         la->la_refcnt++;
1197         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1198 }
1199
1200 static void
1201 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1202 {
1203         char buf[LACP_LAGIDSTR_MAX+1];
1204
1205         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1206             __func__,
1207             lacp_format_lagid(&la->la_actor, &la->la_partner,
1208             buf, sizeof(buf)),
1209             la->la_refcnt, la->la_refcnt - 1));
1210
1211         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1212         la->la_refcnt--;
1213         if (la->la_refcnt > 0) {
1214                 return;
1215         }
1216
1217         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1218         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1219
1220         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1221
1222         free(la, M_DEVBUF);
1223 }
1224
1225 /*
1226  * lacp_aggregator_get: allocate an aggregator.
1227  */
1228
1229 static struct lacp_aggregator *
1230 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1231 {
1232         struct lacp_aggregator *la;
1233
1234         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1235         if (la) {
1236                 la->la_refcnt = 1;
1237                 la->la_nports = 0;
1238                 TAILQ_INIT(&la->la_ports);
1239                 la->la_pending = 0;
1240                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1241         }
1242
1243         return (la);
1244 }
1245
1246 /*
1247  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1248  */
1249
1250 static void
1251 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1252 {
1253         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1254         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1255
1256         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1257 }
1258
1259 static void
1260 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1261     const struct lacp_peerinfo *lpi_port)
1262 {
1263         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1264         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1265         lpi_aggr->lip_key = lpi_port->lip_key;
1266 }
1267
1268 /*
1269  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1270  */
1271
1272 static int
1273 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1274     const struct lacp_port *lp)
1275 {
1276         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1277             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1278                 return (0);
1279         }
1280
1281         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1282                 return (0);
1283         }
1284
1285         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1286                 return (0);
1287         }
1288
1289         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1290                 return (0);
1291         }
1292
1293         return (1);
1294 }
1295
1296 static int
1297 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1298     const struct lacp_peerinfo *b)
1299 {
1300         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1301             sizeof(a->lip_systemid))) {
1302                 return (0);
1303         }
1304
1305         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1306                 return (0);
1307         }
1308
1309         return (1);
1310 }
1311
1312 static void
1313 lacp_port_enable(struct lacp_port *lp)
1314 {
1315         lp->lp_state |= LACP_STATE_AGGREGATION;
1316 }
1317
1318 static void
1319 lacp_port_disable(struct lacp_port *lp)
1320 {
1321         lacp_set_mux(lp, LACP_MUX_DETACHED);
1322
1323         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1324         lp->lp_selected = LACP_UNSELECTED;
1325         lacp_sm_rx_record_default(lp);
1326         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1327         lp->lp_state &= ~LACP_STATE_EXPIRED;
1328 }
1329
1330 /*
1331  * lacp_select: select an aggregator.  create one if necessary.
1332  */
1333 static void
1334 lacp_select(struct lacp_port *lp)
1335 {
1336         struct lacp_softc *lsc = lp->lp_lsc;
1337         struct lacp_aggregator *la;
1338         char buf[LACP_LAGIDSTR_MAX+1];
1339
1340         if (lp->lp_aggregator) {
1341                 return;
1342         }
1343
1344         /* If we haven't heard from our peer, skip this step. */
1345         if (lp->lp_state & LACP_STATE_DEFAULTED)
1346                 return;
1347
1348         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1349             ("timer_wait_while still active"));
1350
1351         LACP_DPRINTF((lp, "port lagid=%s\n",
1352             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1353             buf, sizeof(buf))));
1354
1355         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1356                 if (lacp_aggregator_is_compatible(la, lp)) {
1357                         break;
1358                 }
1359         }
1360
1361         if (la == NULL) {
1362                 la = lacp_aggregator_get(lsc, lp);
1363                 if (la == NULL) {
1364                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1365
1366                         /*
1367                          * will retry on the next tick.
1368                          */
1369
1370                         return;
1371                 }
1372                 lacp_fill_aggregator_id(la, lp);
1373                 LACP_DPRINTF((lp, "aggregator created\n"));
1374         } else {
1375                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1376                 if (la->la_refcnt == LACP_MAX_PORTS)
1377                         return;
1378                 lacp_aggregator_addref(lsc, la);
1379         }
1380
1381         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1382             lacp_format_lagid(&la->la_actor, &la->la_partner,
1383             buf, sizeof(buf))));
1384
1385         lp->lp_aggregator = la;
1386         lp->lp_selected = LACP_SELECTED;
1387 }
1388
1389 /*
1390  * lacp_unselect: finish unselect/detach process.
1391  */
1392
1393 static void
1394 lacp_unselect(struct lacp_port *lp)
1395 {
1396         struct lacp_softc *lsc = lp->lp_lsc;
1397         struct lacp_aggregator *la = lp->lp_aggregator;
1398
1399         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1400             ("timer_wait_while still active"));
1401
1402         if (la == NULL) {
1403                 return;
1404         }
1405
1406         lp->lp_aggregator = NULL;
1407         lacp_aggregator_delref(lsc, la);
1408 }
1409
1410 /* mux machine */
1411
1412 static void
1413 lacp_sm_mux(struct lacp_port *lp)
1414 {
1415         struct lagg_port *lgp = lp->lp_lagg;
1416         struct lagg_softc *sc = lgp->lp_softc;
1417         enum lacp_mux_state new_state;
1418         boolean_t p_sync =
1419                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1420         boolean_t p_collecting =
1421             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1422         enum lacp_selected selected = lp->lp_selected;
1423         struct lacp_aggregator *la;
1424
1425         if (V_lacp_debug > 1)
1426                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1427                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1428                     lp->lp_mux_state, selected, p_sync, p_collecting);
1429
1430 re_eval:
1431         la = lp->lp_aggregator;
1432         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1433             ("MUX not detached"));
1434         new_state = lp->lp_mux_state;
1435         switch (lp->lp_mux_state) {
1436         case LACP_MUX_DETACHED:
1437                 if (selected != LACP_UNSELECTED) {
1438                         new_state = LACP_MUX_WAITING;
1439                 }
1440                 break;
1441         case LACP_MUX_WAITING:
1442                 KASSERT(la->la_pending > 0 ||
1443                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1444                     ("timer_wait_while still active"));
1445                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1446                         new_state = LACP_MUX_ATTACHED;
1447                 } else if (selected == LACP_UNSELECTED) {
1448                         new_state = LACP_MUX_DETACHED;
1449                 }
1450                 break;
1451         case LACP_MUX_ATTACHED:
1452                 if (selected == LACP_SELECTED && p_sync) {
1453                         new_state = LACP_MUX_COLLECTING;
1454                 } else if (selected != LACP_SELECTED) {
1455                         new_state = LACP_MUX_DETACHED;
1456                 }
1457                 break;
1458         case LACP_MUX_COLLECTING:
1459                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1460                         new_state = LACP_MUX_DISTRIBUTING;
1461                 } else if (selected != LACP_SELECTED || !p_sync) {
1462                         new_state = LACP_MUX_ATTACHED;
1463                 }
1464                 break;
1465         case LACP_MUX_DISTRIBUTING:
1466                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1467                         new_state = LACP_MUX_COLLECTING;
1468                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1469                         sc->sc_flapping++;
1470                 }
1471                 break;
1472         default:
1473                 panic("%s: unknown state", __func__);
1474         }
1475
1476         if (lp->lp_mux_state == new_state) {
1477                 return;
1478         }
1479
1480         lacp_set_mux(lp, new_state);
1481         goto re_eval;
1482 }
1483
1484 static void
1485 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1486 {
1487         struct lacp_aggregator *la = lp->lp_aggregator;
1488
1489         if (lp->lp_mux_state == new_state) {
1490                 return;
1491         }
1492
1493         switch (new_state) {
1494         case LACP_MUX_DETACHED:
1495                 lp->lp_state &= ~LACP_STATE_SYNC;
1496                 lacp_disable_distributing(lp);
1497                 lacp_disable_collecting(lp);
1498                 lacp_sm_assert_ntt(lp);
1499                 /* cancel timer */
1500                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1501                         KASSERT(la->la_pending > 0,
1502                             ("timer_wait_while not active"));
1503                         la->la_pending--;
1504                 }
1505                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1506                 lacp_unselect(lp);
1507                 break;
1508         case LACP_MUX_WAITING:
1509                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1510                     LACP_AGGREGATE_WAIT_TIME);
1511                 la->la_pending++;
1512                 break;
1513         case LACP_MUX_ATTACHED:
1514                 lp->lp_state |= LACP_STATE_SYNC;
1515                 lacp_disable_collecting(lp);
1516                 lacp_sm_assert_ntt(lp);
1517                 break;
1518         case LACP_MUX_COLLECTING:
1519                 lacp_enable_collecting(lp);
1520                 lacp_disable_distributing(lp);
1521                 lacp_sm_assert_ntt(lp);
1522                 break;
1523         case LACP_MUX_DISTRIBUTING:
1524                 lacp_enable_distributing(lp);
1525                 break;
1526         default:
1527                 panic("%s: unknown state", __func__);
1528         }
1529
1530         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1531
1532         lp->lp_mux_state = new_state;
1533 }
1534
1535 static void
1536 lacp_sm_mux_timer(struct lacp_port *lp)
1537 {
1538         struct lacp_aggregator *la = lp->lp_aggregator;
1539         char buf[LACP_LAGIDSTR_MAX+1];
1540
1541         KASSERT(la->la_pending > 0, ("no pending event"));
1542
1543         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1544             lacp_format_lagid(&la->la_actor, &la->la_partner,
1545             buf, sizeof(buf)),
1546             la->la_pending, la->la_pending - 1));
1547
1548         la->la_pending--;
1549 }
1550
1551 /* periodic transmit machine */
1552
1553 static void
1554 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1555 {
1556         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1557             LACP_STATE_TIMEOUT)) {
1558                 return;
1559         }
1560
1561         LACP_DPRINTF((lp, "partner timeout changed\n"));
1562
1563         /*
1564          * FAST_PERIODIC -> SLOW_PERIODIC
1565          * or
1566          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1567          *
1568          * let lacp_sm_ptx_tx_schedule to update timeout.
1569          */
1570
1571         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1572
1573         /*
1574          * if timeout has been shortened, assert NTT.
1575          */
1576
1577         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1578                 lacp_sm_assert_ntt(lp);
1579         }
1580 }
1581
1582 static void
1583 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1584 {
1585         int timeout;
1586
1587         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1588             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1589
1590                 /*
1591                  * NO_PERIODIC
1592                  */
1593
1594                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1595                 return;
1596         }
1597
1598         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1599                 return;
1600         }
1601
1602         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1603             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1604
1605         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1606 }
1607
1608 static void
1609 lacp_sm_ptx_timer(struct lacp_port *lp)
1610 {
1611         lacp_sm_assert_ntt(lp);
1612 }
1613
1614 static void
1615 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1616 {
1617         int timeout;
1618
1619         /*
1620          * check LACP_DISABLED first
1621          */
1622
1623         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1624                 return;
1625         }
1626
1627         /*
1628          * check loopback condition.
1629          */
1630
1631         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1632             &lp->lp_actor.lip_systemid)) {
1633                 return;
1634         }
1635
1636         /*
1637          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1638          */
1639
1640         lacp_sm_rx_update_selected(lp, du);
1641         lacp_sm_rx_update_ntt(lp, du);
1642         lacp_sm_rx_record_pdu(lp, du);
1643
1644         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1645             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1646         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1647
1648         lp->lp_state &= ~LACP_STATE_EXPIRED;
1649
1650         /*
1651          * kick transmit machine without waiting the next tick.
1652          */
1653
1654         lacp_sm_tx(lp);
1655 }
1656
1657 static void
1658 lacp_sm_rx_set_expired(struct lacp_port *lp)
1659 {
1660         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1661         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1662         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1663         lp->lp_state |= LACP_STATE_EXPIRED;
1664 }
1665
1666 static void
1667 lacp_sm_rx_timer(struct lacp_port *lp)
1668 {
1669         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1670                 /* CURRENT -> EXPIRED */
1671                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1672                 lacp_sm_rx_set_expired(lp);
1673         } else {
1674                 /* EXPIRED -> DEFAULTED */
1675                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1676                 lacp_sm_rx_update_default_selected(lp);
1677                 lacp_sm_rx_record_default(lp);
1678                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1679         }
1680 }
1681
1682 static void
1683 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1684 {
1685         boolean_t active;
1686         uint8_t oldpstate;
1687         char buf[LACP_STATESTR_MAX+1];
1688
1689         LACP_TRACE(lp);
1690
1691         oldpstate = lp->lp_partner.lip_state;
1692
1693         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1694             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1695             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1696
1697         lp->lp_partner = du->ldu_actor;
1698         if (active &&
1699             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1700             LACP_STATE_AGGREGATION) &&
1701             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1702             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1703                 /*
1704                  * XXX Maintain legacy behavior of leaving the
1705                  * LACP_STATE_SYNC bit unchanged from the partner's
1706                  * advertisement if lsc_strict_mode is false.
1707                  * TODO: We should re-examine the concept of the "strict mode"
1708                  * to ensure it makes sense to maintain a non-strict mode.
1709                  */
1710                 if (lp->lp_lsc->lsc_strict_mode)
1711                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1712         } else {
1713                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1714         }
1715
1716         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1717
1718         if (oldpstate != lp->lp_partner.lip_state) {
1719                 LACP_DPRINTF((lp, "old pstate %s\n",
1720                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1721                 LACP_DPRINTF((lp, "new pstate %s\n",
1722                     lacp_format_state(lp->lp_partner.lip_state, buf,
1723                     sizeof(buf))));
1724         }
1725
1726         lacp_sm_ptx_update_timeout(lp, oldpstate);
1727 }
1728
1729 static void
1730 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1731 {
1732
1733         LACP_TRACE(lp);
1734
1735         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1736             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1737             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1738                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1739                 lacp_sm_assert_ntt(lp);
1740         }
1741 }
1742
1743 static void
1744 lacp_sm_rx_record_default(struct lacp_port *lp)
1745 {
1746         uint8_t oldpstate;
1747
1748         LACP_TRACE(lp);
1749
1750         oldpstate = lp->lp_partner.lip_state;
1751         if (lp->lp_lsc->lsc_strict_mode)
1752                 lp->lp_partner = lacp_partner_admin_strict;
1753         else
1754                 lp->lp_partner = lacp_partner_admin_optimistic;
1755         lp->lp_state |= LACP_STATE_DEFAULTED;
1756         lacp_sm_ptx_update_timeout(lp, oldpstate);
1757 }
1758
1759 static void
1760 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1761     const struct lacp_peerinfo *info)
1762 {
1763
1764         LACP_TRACE(lp);
1765
1766         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1767             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1768             LACP_STATE_AGGREGATION)) {
1769                 lp->lp_selected = LACP_UNSELECTED;
1770                 /* mux machine will clean up lp->lp_aggregator */
1771         }
1772 }
1773
1774 static void
1775 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1776 {
1777
1778         LACP_TRACE(lp);
1779
1780         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1781 }
1782
1783 static void
1784 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1785 {
1786
1787         LACP_TRACE(lp);
1788
1789         if (lp->lp_lsc->lsc_strict_mode)
1790                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1791                     &lacp_partner_admin_strict);
1792         else
1793                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1794                     &lacp_partner_admin_optimistic);
1795 }
1796
1797 /* transmit machine */
1798
1799 static void
1800 lacp_sm_tx(struct lacp_port *lp)
1801 {
1802         int error = 0;
1803
1804         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1805 #if 1
1806             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1807             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1808 #endif
1809             ) {
1810                 lp->lp_flags &= ~LACP_PORT_NTT;
1811         }
1812
1813         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1814                 return;
1815         }
1816
1817         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1818         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1819                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1820                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1821                 return;
1822         }
1823
1824         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1825                 error = lacp_xmit_lacpdu(lp);
1826         } else {
1827                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1828         }
1829
1830         if (error == 0) {
1831                 lp->lp_flags &= ~LACP_PORT_NTT;
1832         } else {
1833                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1834                     error));
1835         }
1836 }
1837
1838 static void
1839 lacp_sm_assert_ntt(struct lacp_port *lp)
1840 {
1841
1842         lp->lp_flags |= LACP_PORT_NTT;
1843 }
1844
1845 static void
1846 lacp_run_timers(struct lacp_port *lp)
1847 {
1848         int i;
1849
1850         for (i = 0; i < LACP_NTIMER; i++) {
1851                 KASSERT(lp->lp_timer[i] >= 0,
1852                     ("invalid timer value %d", lp->lp_timer[i]));
1853                 if (lp->lp_timer[i] == 0) {
1854                         continue;
1855                 } else if (--lp->lp_timer[i] <= 0) {
1856                         if (lacp_timer_funcs[i]) {
1857                                 (*lacp_timer_funcs[i])(lp);
1858                         }
1859                 }
1860         }
1861 }
1862
1863 int
1864 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1865 {
1866         struct lacp_softc *lsc = lp->lp_lsc;
1867         struct lagg_port *lgp = lp->lp_lagg;
1868         struct lacp_port *lp2;
1869         struct markerdu *mdu;
1870         int error = 0;
1871         int pending = 0;
1872
1873         if (m->m_pkthdr.len != sizeof(*mdu)) {
1874                 goto bad;
1875         }
1876
1877         if ((m->m_flags & M_MCAST) == 0) {
1878                 goto bad;
1879         }
1880
1881         if (m->m_len < sizeof(*mdu)) {
1882                 m = m_pullup(m, sizeof(*mdu));
1883                 if (m == NULL) {
1884                         return (ENOMEM);
1885                 }
1886         }
1887
1888         mdu = mtod(m, struct markerdu *);
1889
1890         if (memcmp(&mdu->mdu_eh.ether_dhost,
1891             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1892                 goto bad;
1893         }
1894
1895         if (mdu->mdu_sph.sph_version != 1) {
1896                 goto bad;
1897         }
1898
1899         switch (mdu->mdu_tlv.tlv_type) {
1900         case MARKER_TYPE_INFO:
1901                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1902                     marker_info_tlv_template, TRUE)) {
1903                         goto bad;
1904                 }
1905                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1906                 memcpy(&mdu->mdu_eh.ether_dhost,
1907                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
1908                 memcpy(&mdu->mdu_eh.ether_shost,
1909                     lgp->lp_lladdr, ETHER_ADDR_LEN);
1910                 error = lagg_enqueue(lp->lp_ifp, m);
1911                 break;
1912
1913         case MARKER_TYPE_RESPONSE:
1914                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1915                     marker_response_tlv_template, TRUE)) {
1916                         goto bad;
1917                 }
1918                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
1919                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
1920                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
1921
1922                 /* Verify that it is the last marker we sent out */
1923                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
1924                     sizeof(struct lacp_markerinfo)))
1925                         goto bad;
1926
1927                 LACP_LOCK(lsc);
1928                 lp->lp_flags &= ~LACP_PORT_MARK;
1929
1930                 if (lsc->lsc_suppress_distributing) {
1931                         /* Check if any ports are waiting for a response */
1932                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
1933                                 if (lp2->lp_flags & LACP_PORT_MARK) {
1934                                         pending = 1;
1935                                         break;
1936                                 }
1937                         }
1938
1939                         if (pending == 0) {
1940                                 /* All interface queues are clear */
1941                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
1942                                 lsc->lsc_suppress_distributing = FALSE;
1943                         }
1944                 }
1945                 LACP_UNLOCK(lsc);
1946                 m_freem(m);
1947                 break;
1948
1949         default:
1950                 goto bad;
1951         }
1952
1953         return (error);
1954
1955 bad:
1956         LACP_DPRINTF((lp, "bad marker frame\n"));
1957         m_freem(m);
1958         return (EINVAL);
1959 }
1960
1961 static int
1962 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
1963     const struct tlv_template *tmpl, boolean_t check_type)
1964 {
1965         while (/* CONSTCOND */ 1) {
1966                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
1967                         return (EINVAL);
1968                 }
1969                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
1970                     tlv->tlv_length != tmpl->tmpl_length) {
1971                         return (EINVAL);
1972                 }
1973                 if (tmpl->tmpl_type == 0) {
1974                         break;
1975                 }
1976                 tlv = (const struct tlvhdr *)
1977                     ((const char *)tlv + tlv->tlv_length);
1978                 tmpl++;
1979         }
1980
1981         return (0);
1982 }
1983
1984 /* Debugging */
1985 const char *
1986 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
1987 {
1988         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
1989             (int)mac[0],
1990             (int)mac[1],
1991             (int)mac[2],
1992             (int)mac[3],
1993             (int)mac[4],
1994             (int)mac[5]);
1995
1996         return (buf);
1997 }
1998
1999 const char *
2000 lacp_format_systemid(const struct lacp_systemid *sysid,
2001     char *buf, size_t buflen)
2002 {
2003         char macbuf[LACP_MACSTR_MAX+1];
2004
2005         snprintf(buf, buflen, "%04X,%s",
2006             ntohs(sysid->lsi_prio),
2007             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2008
2009         return (buf);
2010 }
2011
2012 const char *
2013 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2014 {
2015         snprintf(buf, buflen, "%04X,%04X",
2016             ntohs(portid->lpi_prio),
2017             ntohs(portid->lpi_portno));
2018
2019         return (buf);
2020 }
2021
2022 const char *
2023 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2024 {
2025         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2026         char portid[LACP_PORTIDSTR_MAX+1];
2027
2028         snprintf(buf, buflen, "(%s,%04X,%s)",
2029             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2030             ntohs(peer->lip_key),
2031             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2032
2033         return (buf);
2034 }
2035
2036 const char *
2037 lacp_format_lagid(const struct lacp_peerinfo *a,
2038     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2039 {
2040         char astr[LACP_PARTNERSTR_MAX+1];
2041         char bstr[LACP_PARTNERSTR_MAX+1];
2042
2043 #if 0
2044         /*
2045          * there's a convention to display small numbered peer
2046          * in the left.
2047          */
2048
2049         if (lacp_compare_peerinfo(a, b) > 0) {
2050                 const struct lacp_peerinfo *t;
2051
2052                 t = a;
2053                 a = b;
2054                 b = t;
2055         }
2056 #endif
2057
2058         snprintf(buf, buflen, "[%s,%s]",
2059             lacp_format_partner(a, astr, sizeof(astr)),
2060             lacp_format_partner(b, bstr, sizeof(bstr)));
2061
2062         return (buf);
2063 }
2064
2065 const char *
2066 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2067     char *buf, size_t buflen)
2068 {
2069         if (la == NULL) {
2070                 return ("(none)");
2071         }
2072
2073         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2074 }
2075
2076 const char *
2077 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2078 {
2079         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2080         return (buf);
2081 }
2082
2083 static void
2084 lacp_dump_lacpdu(const struct lacpdu *du)
2085 {
2086         char buf[LACP_PARTNERSTR_MAX+1];
2087         char buf2[LACP_STATESTR_MAX+1];
2088
2089         printf("actor=%s\n",
2090             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2091         printf("actor.state=%s\n",
2092             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2093         printf("partner=%s\n",
2094             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2095         printf("partner.state=%s\n",
2096             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2097
2098         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2099 }
2100
2101 static void
2102 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2103 {
2104         va_list va;
2105
2106         if (lp) {
2107                 printf("%s: ", lp->lp_ifp->if_xname);
2108         }
2109
2110         va_start(va, fmt);
2111         vprintf(fmt, va);
2112         va_end(va);
2113 }