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         struct lacp_port **map;
 839         uint32_t hash;
 840         int count;
 841
 842         if (__predict_false(lsc->lsc_suppress_distributing)) {
 843                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 844                 return (NULL);
 845         }
 846
 847         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 848         if (pm->pm_count == 0) {
 849                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 850                 return (NULL);
 851         }
 852
 853 #ifdef NUMA
 854         if ((sc->sc_opts & LAGG_OPT_USE_NUMA) &&
 855             pm->pm_num_dom > 1 && m->m_pkthdr.numa_domain < MAXMEMDOM) {
 856                 count = pm->pm_numa[m->m_pkthdr.numa_domain].count;
 857                 if (count > 0) {
 858                         map = pm->pm_numa[m->m_pkthdr.numa_domain].map;
 859                 } else {
 860                         /* No ports on this domain; use global hash. */
 861                         map = pm->pm_map;
 862                         count = pm->pm_count;
 863                 }
 864         } else
 865 #endif
 866         {
 867                 map = pm->pm_map;
 868                 count = pm->pm_count;
 869         }
 870         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 871             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 872                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 873         else
 874                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 875
 876         hash %= count;
 877         lp = map[hash];
 878
 879         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 880             ("aggregated port is not distributing"));
 881
 882         return (lp->lp_lagg);
 883 }
 884
 885 #ifdef RATELIMIT
 886 struct lagg_port *
 887 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t flowid)
 888 {
 889         struct lacp_softc *lsc = LACP_SOFTC(sc);
 890         struct lacp_portmap *pm;
 891         struct lacp_port *lp;
 892         uint32_t hash;
 893
 894         if (__predict_false(lsc->lsc_suppress_distributing)) {
 895                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 896                 return (NULL);
 897         }
 898
 899         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 900         if (pm->pm_count == 0) {
 901                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 902                 return (NULL);
 903         }
 904
 905         hash = flowid >> sc->flowid_shift;
 906         hash %= pm->pm_count;
 907         lp = pm->pm_map[hash];
 908
 909         return (lp->lp_lagg);
 910 }
 911 #endif
 912
 913 /*
 914  * lacp_suppress_distributing: drop transmit packets for a while
 915  * to preserve packet ordering.
 916  */
 917
 918 static void
 919 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 920 {
 921         struct lacp_port *lp;
 922
 923         if (lsc->lsc_active_aggregator != la) {
 924                 return;
 925         }
 926
 927         LACP_TRACE(NULL);
 928
 929         lsc->lsc_suppress_distributing = TRUE;
 930
 931         /* send a marker frame down each port to verify the queues are empty */
 932         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 933                 lp->lp_flags |= LACP_PORT_MARK;
 934                 lacp_xmit_marker(lp);
 935         }
 936
 937         /* set a timeout for the marker frames */
 938         callout_reset(&lsc->lsc_transit_callout,
 939             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 940 }
 941
 942 static int
 943 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 944     const struct lacp_peerinfo *b)
 945 {
 946         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 947 }
 948
 949 static int
 950 lacp_compare_systemid(const struct lacp_systemid *a,
 951     const struct lacp_systemid *b)
 952 {
 953         return (memcmp(a, b, sizeof(*a)));
 954 }
 955
 956 #if 0   /* unused */
 957 static int
 958 lacp_compare_portid(const struct lacp_portid *a,
 959     const struct lacp_portid *b)
 960 {
 961         return (memcmp(a, b, sizeof(*a)));
 962 }
 963 #endif
 964
 965 static uint64_t
 966 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 967 {
 968         struct lacp_port *lp;
 969         uint64_t speed;
 970
 971         lp = TAILQ_FIRST(&la->la_ports);
 972         if (lp == NULL) {
 973                 return (0);
 974         }
 975
 976         speed = ifmedia_baudrate(lp->lp_media);
 977         speed *= la->la_nports;
 978         if (speed == 0) {
 979                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 980                     lp->lp_media, la->la_nports));
 981         }
 982
 983         return (speed);
 984 }
 985
 986 /*
 987  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 988  * packets from lagg(4) interface.
 989  */
 990
 991 static void
 992 lacp_select_active_aggregator(struct lacp_softc *lsc)
 993 {
 994         struct lacp_aggregator *la;
 995         struct lacp_aggregator *best_la = NULL;
 996         uint64_t best_speed = 0;
 997         char buf[LACP_LAGIDSTR_MAX+1];
 998
 999         LACP_TRACE(NULL);
1000
1001         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1002                 uint64_t speed;
1003
1004                 if (la->la_nports == 0) {
1005                         continue;
1006                 }
1007
1008                 speed = lacp_aggregator_bandwidth(la);
1009                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
1010                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
1011                     speed, la->la_nports));
1012
1013                 /*
1014                  * This aggregator is chosen if the partner has a better
1015                  * system priority or, the total aggregated speed is higher
1016                  * or, it is already the chosen aggregator
1017                  */
1018                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
1019                     LACP_SYS_PRI(best_la->la_partner)) ||
1020                     speed > best_speed ||
1021                     (speed == best_speed &&
1022                     la == lsc->lsc_active_aggregator)) {
1023                         best_la = la;
1024                         best_speed = speed;
1025                 }
1026         }
1027
1028         KASSERT(best_la == NULL || best_la->la_nports > 0,
1029             ("invalid aggregator refcnt"));
1030         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1031             ("invalid aggregator list"));
1032
1033         if (lsc->lsc_active_aggregator != best_la) {
1034                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1035                 LACP_DPRINTF((NULL, "old %s\n",
1036                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1037                     buf, sizeof(buf))));
1038         } else {
1039                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1040         }
1041         LACP_DPRINTF((NULL, "new %s\n",
1042             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1043
1044         if (lsc->lsc_active_aggregator != best_la) {
1045                 lsc->lsc_active_aggregator = best_la;
1046                 lacp_update_portmap(lsc);
1047                 if (best_la) {
1048                         lacp_suppress_distributing(lsc, best_la);
1049                 }
1050         }
1051 }
1052
1053 /*
1054  * Updated the inactive portmap array with the new list of ports and
1055  * make it live.
1056  */
1057 static void
1058 lacp_update_portmap(struct lacp_softc *lsc)
1059 {
1060         struct lagg_softc *sc = lsc->lsc_softc;
1061         struct lacp_aggregator *la;
1062         struct lacp_portmap *p;
1063         struct lacp_port *lp;
1064         uint64_t speed;
1065         u_int newmap;
1066         int i;
1067 #ifdef NUMA
1068         int count;
1069         uint8_t domain;
1070 #endif
1071
1072         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1073         p = &lsc->lsc_pmap[newmap];
1074         la = lsc->lsc_active_aggregator;
1075         speed = 0;
1076         bzero(p, sizeof(struct lacp_portmap));
1077
1078         if (la != NULL && la->la_nports > 0) {
1079                 p->pm_count = la->la_nports;
1080                 i = 0;
1081                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q) {
1082                         p->pm_map[i++] = lp;
1083 #ifdef NUMA
1084                         domain = lp->lp_ifp->if_numa_domain;
1085                         if (domain >= MAXMEMDOM)
1086                                 continue;
1087                         count = p->pm_numa[domain].count;
1088                         p->pm_numa[domain].map[count] = lp;
1089                         p->pm_numa[domain].count++;
1090 #endif
1091                 }
1092                 KASSERT(i == p->pm_count, ("Invalid port count"));
1093
1094 #ifdef NUMA
1095                 for (i = 0; i < MAXMEMDOM; i++) {
1096                         if (p->pm_numa[i].count != 0)
1097                                 p->pm_num_dom++;
1098                 }
1099 #endif
1100                 speed = lacp_aggregator_bandwidth(la);
1101         }
1102         sc->sc_ifp->if_baudrate = speed;
1103
1104         /* switch the active portmap over */
1105         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1106         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1107                     lsc->lsc_activemap,
1108                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1109 }
1110
1111 static uint16_t
1112 lacp_compose_key(struct lacp_port *lp)
1113 {
1114         struct lagg_port *lgp = lp->lp_lagg;
1115         struct lagg_softc *sc = lgp->lp_softc;
1116         u_int media = lp->lp_media;
1117         uint16_t key;
1118
1119         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1120
1121                 /*
1122                  * non-aggregatable links should have unique keys.
1123                  *
1124                  * XXX this isn't really unique as if_index is 16 bit.
1125                  */
1126
1127                 /* bit 0..14:   (some bits of) if_index of this port */
1128                 key = lp->lp_ifp->if_index;
1129                 /* bit 15:      1 */
1130                 key |= 0x8000;
1131         } else {
1132                 u_int subtype = IFM_SUBTYPE(media);
1133
1134                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1135                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1136
1137                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1138                 switch (subtype) {
1139                 case IFM_10_T:
1140                 case IFM_10_2:
1141                 case IFM_10_5:
1142                 case IFM_10_STP:
1143                 case IFM_10_FL:
1144                         key = IFM_10_T;
1145                         break;
1146                 case IFM_100_TX:
1147                 case IFM_100_FX:
1148                 case IFM_100_T4:
1149                 case IFM_100_VG:
1150                 case IFM_100_T2:
1151                 case IFM_100_T:
1152                 case IFM_100_SGMII:
1153                         key = IFM_100_TX;
1154                         break;
1155                 case IFM_1000_SX:
1156                 case IFM_1000_LX:
1157                 case IFM_1000_CX:
1158                 case IFM_1000_T:
1159                 case IFM_1000_KX:
1160                 case IFM_1000_SGMII:
1161                 case IFM_1000_CX_SGMII:
1162                         key = IFM_1000_SX;
1163                         break;
1164                 case IFM_10G_LR:
1165                 case IFM_10G_SR:
1166                 case IFM_10G_CX4:
1167                 case IFM_10G_TWINAX:
1168                 case IFM_10G_TWINAX_LONG:
1169                 case IFM_10G_LRM:
1170                 case IFM_10G_T:
1171                 case IFM_10G_KX4:
1172                 case IFM_10G_KR:
1173                 case IFM_10G_CR1:
1174                 case IFM_10G_ER:
1175                 case IFM_10G_SFI:
1176                 case IFM_10G_AOC:
1177                         key = IFM_10G_LR;
1178                         break;
1179                 case IFM_20G_KR2:
1180                         key = IFM_20G_KR2;
1181                         break;
1182                 case IFM_2500_KX:
1183                 case IFM_2500_T:
1184                 case IFM_2500_X:
1185                         key = IFM_2500_KX;
1186                         break;
1187                 case IFM_5000_T:
1188                 case IFM_5000_KR:
1189                 case IFM_5000_KR_S:
1190                 case IFM_5000_KR1:
1191                         key = IFM_5000_T;
1192                         break;
1193                 case IFM_50G_PCIE:
1194                 case IFM_50G_CR2:
1195                 case IFM_50G_KR2:
1196                 case IFM_50G_SR2:
1197                 case IFM_50G_LR2:
1198                 case IFM_50G_LAUI2_AC:
1199                 case IFM_50G_LAUI2:
1200                 case IFM_50G_AUI2_AC:
1201                 case IFM_50G_AUI2:
1202                 case IFM_50G_CP:
1203                 case IFM_50G_SR:
1204                 case IFM_50G_LR:
1205                 case IFM_50G_FR:
1206                 case IFM_50G_KR_PAM4:
1207                 case IFM_50G_AUI1_AC:
1208                 case IFM_50G_AUI1:
1209                         key = IFM_50G_PCIE;
1210                         break;
1211                 case IFM_56G_R4:
1212                         key = IFM_56G_R4;
1213                         break;
1214                 case IFM_25G_PCIE:
1215                 case IFM_25G_CR:
1216                 case IFM_25G_KR:
1217                 case IFM_25G_SR:
1218                 case IFM_25G_LR:
1219                 case IFM_25G_ACC:
1220                 case IFM_25G_AOC:
1221                 case IFM_25G_T:
1222                 case IFM_25G_CR_S:
1223                 case IFM_25G_CR1:
1224                 case IFM_25G_KR_S:
1225                 case IFM_25G_AUI:
1226                 case IFM_25G_KR1:
1227                         key = IFM_25G_PCIE;
1228                         break;
1229                 case IFM_40G_CR4:
1230                 case IFM_40G_SR4:
1231                 case IFM_40G_LR4:
1232                 case IFM_40G_XLPPI:
1233                 case IFM_40G_KR4:
1234                 case IFM_40G_XLAUI:
1235                 case IFM_40G_XLAUI_AC:
1236                 case IFM_40G_ER4:
1237                         key = IFM_40G_CR4;
1238                         break;
1239                 case IFM_100G_CR4:
1240                 case IFM_100G_SR4:
1241                 case IFM_100G_KR4:
1242                 case IFM_100G_LR4:
1243                 case IFM_100G_CAUI4_AC:
1244                 case IFM_100G_CAUI4:
1245                 case IFM_100G_AUI4_AC:
1246                 case IFM_100G_AUI4:
1247                 case IFM_100G_CR_PAM4:
1248                 case IFM_100G_KR_PAM4:
1249                 case IFM_100G_CP2:
1250                 case IFM_100G_SR2:
1251                 case IFM_100G_DR:
1252                 case IFM_100G_KR2_PAM4:
1253                 case IFM_100G_CAUI2_AC:
1254                 case IFM_100G_CAUI2:
1255                 case IFM_100G_AUI2_AC:
1256                 case IFM_100G_AUI2:
1257                         key = IFM_100G_CR4;
1258                         break;
1259                 case IFM_200G_CR4_PAM4:
1260                 case IFM_200G_SR4:
1261                 case IFM_200G_FR4:
1262                 case IFM_200G_LR4:
1263                 case IFM_200G_DR4:
1264                 case IFM_200G_KR4_PAM4:
1265                 case IFM_200G_AUI4_AC:
1266                 case IFM_200G_AUI4:
1267                 case IFM_200G_AUI8_AC:
1268                 case IFM_200G_AUI8:
1269                         key = IFM_200G_CR4_PAM4;
1270                         break;
1271                 case IFM_400G_FR8:
1272                 case IFM_400G_LR8:
1273                 case IFM_400G_DR4:
1274                 case IFM_400G_AUI8_AC:
1275                 case IFM_400G_AUI8:
1276                         key = IFM_400G_FR8;
1277                         break;
1278                 default:
1279                         key = subtype;
1280                         break;
1281                 }
1282                 /* bit 5..14:   (some bits of) if_index of lagg device */
1283                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1284                 /* bit 15:      0 */
1285         }
1286         return (htons(key));
1287 }
1288
1289 static void
1290 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1291 {
1292         char buf[LACP_LAGIDSTR_MAX+1];
1293
1294         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1295             __func__,
1296             lacp_format_lagid(&la->la_actor, &la->la_partner,
1297             buf, sizeof(buf)),
1298             la->la_refcnt, la->la_refcnt + 1));
1299
1300         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1301         la->la_refcnt++;
1302         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1303 }
1304
1305 static void
1306 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1307 {
1308         char buf[LACP_LAGIDSTR_MAX+1];
1309
1310         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1311             __func__,
1312             lacp_format_lagid(&la->la_actor, &la->la_partner,
1313             buf, sizeof(buf)),
1314             la->la_refcnt, la->la_refcnt - 1));
1315
1316         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1317         la->la_refcnt--;
1318         if (la->la_refcnt > 0) {
1319                 return;
1320         }
1321
1322         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1323         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1324
1325         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1326
1327         free(la, M_DEVBUF);
1328 }
1329
1330 /*
1331  * lacp_aggregator_get: allocate an aggregator.
1332  */
1333
1334 static struct lacp_aggregator *
1335 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1336 {
1337         struct lacp_aggregator *la;
1338
1339         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1340         if (la) {
1341                 la->la_refcnt = 1;
1342                 la->la_nports = 0;
1343                 TAILQ_INIT(&la->la_ports);
1344                 la->la_pending = 0;
1345                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1346         }
1347
1348         return (la);
1349 }
1350
1351 /*
1352  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1353  */
1354
1355 static void
1356 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1357 {
1358         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1359         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1360
1361         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1362 }
1363
1364 static void
1365 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1366     const struct lacp_peerinfo *lpi_port)
1367 {
1368         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1369         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1370         lpi_aggr->lip_key = lpi_port->lip_key;
1371 }
1372
1373 /*
1374  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1375  */
1376
1377 static int
1378 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1379     const struct lacp_port *lp)
1380 {
1381         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1382             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1383                 return (0);
1384         }
1385
1386         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1387                 return (0);
1388         }
1389
1390         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1391                 return (0);
1392         }
1393
1394         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1395                 return (0);
1396         }
1397
1398         return (1);
1399 }
1400
1401 static int
1402 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1403     const struct lacp_peerinfo *b)
1404 {
1405         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1406             sizeof(a->lip_systemid))) {
1407                 return (0);
1408         }
1409
1410         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1411                 return (0);
1412         }
1413
1414         return (1);
1415 }
1416
1417 static void
1418 lacp_port_enable(struct lacp_port *lp)
1419 {
1420         lp->lp_state |= LACP_STATE_AGGREGATION;
1421 }
1422
1423 static void
1424 lacp_port_disable(struct lacp_port *lp)
1425 {
1426         lacp_set_mux(lp, LACP_MUX_DETACHED);
1427
1428         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1429         lp->lp_selected = LACP_UNSELECTED;
1430         lacp_sm_rx_record_default(lp);
1431         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1432         lp->lp_state &= ~LACP_STATE_EXPIRED;
1433 }
1434
1435 /*
1436  * lacp_select: select an aggregator.  create one if necessary.
1437  */
1438 static void
1439 lacp_select(struct lacp_port *lp)
1440 {
1441         struct lacp_softc *lsc = lp->lp_lsc;
1442         struct lacp_aggregator *la;
1443         char buf[LACP_LAGIDSTR_MAX+1];
1444
1445         if (lp->lp_aggregator) {
1446                 return;
1447         }
1448
1449         /* If we haven't heard from our peer, skip this step. */
1450         if (lp->lp_state & LACP_STATE_DEFAULTED)
1451                 return;
1452
1453         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1454             ("timer_wait_while still active"));
1455
1456         LACP_DPRINTF((lp, "port lagid=%s\n",
1457             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1458             buf, sizeof(buf))));
1459
1460         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1461                 if (lacp_aggregator_is_compatible(la, lp)) {
1462                         break;
1463                 }
1464         }
1465
1466         if (la == NULL) {
1467                 la = lacp_aggregator_get(lsc, lp);
1468                 if (la == NULL) {
1469                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1470
1471                         /*
1472                          * will retry on the next tick.
1473                          */
1474
1475                         return;
1476                 }
1477                 lacp_fill_aggregator_id(la, lp);
1478                 LACP_DPRINTF((lp, "aggregator created\n"));
1479         } else {
1480                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1481                 if (la->la_refcnt == LACP_MAX_PORTS)
1482                         return;
1483                 lacp_aggregator_addref(lsc, la);
1484         }
1485
1486         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1487             lacp_format_lagid(&la->la_actor, &la->la_partner,
1488             buf, sizeof(buf))));
1489
1490         lp->lp_aggregator = la;
1491         lp->lp_selected = LACP_SELECTED;
1492 }
1493
1494 /*
1495  * lacp_unselect: finish unselect/detach process.
1496  */
1497
1498 static void
1499 lacp_unselect(struct lacp_port *lp)
1500 {
1501         struct lacp_softc *lsc = lp->lp_lsc;
1502         struct lacp_aggregator *la = lp->lp_aggregator;
1503
1504         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1505             ("timer_wait_while still active"));
1506
1507         if (la == NULL) {
1508                 return;
1509         }
1510
1511         lp->lp_aggregator = NULL;
1512         lacp_aggregator_delref(lsc, la);
1513 }
1514
1515 /* mux machine */
1516
1517 static void
1518 lacp_sm_mux(struct lacp_port *lp)
1519 {
1520         struct lagg_port *lgp = lp->lp_lagg;
1521         struct lagg_softc *sc = lgp->lp_softc;
1522         enum lacp_mux_state new_state;
1523         boolean_t p_sync =
1524                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1525         boolean_t p_collecting =
1526             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1527         enum lacp_selected selected = lp->lp_selected;
1528         struct lacp_aggregator *la;
1529
1530         if (V_lacp_debug > 1)
1531                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1532                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1533                     lp->lp_mux_state, selected, p_sync, p_collecting);
1534
1535 re_eval:
1536         la = lp->lp_aggregator;
1537         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1538             ("MUX not detached"));
1539         new_state = lp->lp_mux_state;
1540         switch (lp->lp_mux_state) {
1541         case LACP_MUX_DETACHED:
1542                 if (selected != LACP_UNSELECTED) {
1543                         new_state = LACP_MUX_WAITING;
1544                 }
1545                 break;
1546         case LACP_MUX_WAITING:
1547                 KASSERT(la->la_pending > 0 ||
1548                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1549                     ("timer_wait_while still active"));
1550                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1551                         new_state = LACP_MUX_ATTACHED;
1552                 } else if (selected == LACP_UNSELECTED) {
1553                         new_state = LACP_MUX_DETACHED;
1554                 }
1555                 break;
1556         case LACP_MUX_ATTACHED:
1557                 if (selected == LACP_SELECTED && p_sync) {
1558                         new_state = LACP_MUX_COLLECTING;
1559                 } else if (selected != LACP_SELECTED) {
1560                         new_state = LACP_MUX_DETACHED;
1561                 }
1562                 break;
1563         case LACP_MUX_COLLECTING:
1564                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1565                         new_state = LACP_MUX_DISTRIBUTING;
1566                 } else if (selected != LACP_SELECTED || !p_sync) {
1567                         new_state = LACP_MUX_ATTACHED;
1568                 }
1569                 break;
1570         case LACP_MUX_DISTRIBUTING:
1571                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1572                         new_state = LACP_MUX_COLLECTING;
1573                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1574                         sc->sc_flapping++;
1575                 }
1576                 break;
1577         default:
1578                 panic("%s: unknown state", __func__);
1579         }
1580
1581         if (lp->lp_mux_state == new_state) {
1582                 return;
1583         }
1584
1585         lacp_set_mux(lp, new_state);
1586         goto re_eval;
1587 }
1588
1589 static void
1590 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1591 {
1592         struct lacp_aggregator *la = lp->lp_aggregator;
1593
1594         if (lp->lp_mux_state == new_state) {
1595                 return;
1596         }
1597
1598         switch (new_state) {
1599         case LACP_MUX_DETACHED:
1600                 lp->lp_state &= ~LACP_STATE_SYNC;
1601                 lacp_disable_distributing(lp);
1602                 lacp_disable_collecting(lp);
1603                 lacp_sm_assert_ntt(lp);
1604                 /* cancel timer */
1605                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1606                         KASSERT(la->la_pending > 0,
1607                             ("timer_wait_while not active"));
1608                         la->la_pending--;
1609                 }
1610                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1611                 lacp_unselect(lp);
1612                 break;
1613         case LACP_MUX_WAITING:
1614                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1615                     LACP_AGGREGATE_WAIT_TIME);
1616                 la->la_pending++;
1617                 break;
1618         case LACP_MUX_ATTACHED:
1619                 lp->lp_state |= LACP_STATE_SYNC;
1620                 lacp_disable_collecting(lp);
1621                 lacp_sm_assert_ntt(lp);
1622                 break;
1623         case LACP_MUX_COLLECTING:
1624                 lacp_enable_collecting(lp);
1625                 lacp_disable_distributing(lp);
1626                 lacp_sm_assert_ntt(lp);
1627                 break;
1628         case LACP_MUX_DISTRIBUTING:
1629                 lacp_enable_distributing(lp);
1630                 break;
1631         default:
1632                 panic("%s: unknown state", __func__);
1633         }
1634
1635         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1636
1637         lp->lp_mux_state = new_state;
1638 }
1639
1640 static void
1641 lacp_sm_mux_timer(struct lacp_port *lp)
1642 {
1643         struct lacp_aggregator *la = lp->lp_aggregator;
1644         char buf[LACP_LAGIDSTR_MAX+1];
1645
1646         KASSERT(la->la_pending > 0, ("no pending event"));
1647
1648         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1649             lacp_format_lagid(&la->la_actor, &la->la_partner,
1650             buf, sizeof(buf)),
1651             la->la_pending, la->la_pending - 1));
1652
1653         la->la_pending--;
1654 }
1655
1656 /* periodic transmit machine */
1657
1658 static void
1659 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1660 {
1661         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1662             LACP_STATE_TIMEOUT)) {
1663                 return;
1664         }
1665
1666         LACP_DPRINTF((lp, "partner timeout changed\n"));
1667
1668         /*
1669          * FAST_PERIODIC -> SLOW_PERIODIC
1670          * or
1671          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1672          *
1673          * let lacp_sm_ptx_tx_schedule to update timeout.
1674          */
1675
1676         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1677
1678         /*
1679          * if timeout has been shortened, assert NTT.
1680          */
1681
1682         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1683                 lacp_sm_assert_ntt(lp);
1684         }
1685 }
1686
1687 static void
1688 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1689 {
1690         int timeout;
1691
1692         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1693             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1694
1695                 /*
1696                  * NO_PERIODIC
1697                  */
1698
1699                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1700                 return;
1701         }
1702
1703         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1704                 return;
1705         }
1706
1707         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1708             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1709
1710         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1711 }
1712
1713 static void
1714 lacp_sm_ptx_timer(struct lacp_port *lp)
1715 {
1716         lacp_sm_assert_ntt(lp);
1717 }
1718
1719 static void
1720 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1721 {
1722         int timeout;
1723
1724         /*
1725          * check LACP_DISABLED first
1726          */
1727
1728         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1729                 return;
1730         }
1731
1732         /*
1733          * check loopback condition.
1734          */
1735
1736         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1737             &lp->lp_actor.lip_systemid)) {
1738                 return;
1739         }
1740
1741         /*
1742          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1743          */
1744
1745         lacp_sm_rx_update_selected(lp, du);
1746         lacp_sm_rx_update_ntt(lp, du);
1747         lacp_sm_rx_record_pdu(lp, du);
1748
1749         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1750             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1751         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1752
1753         lp->lp_state &= ~LACP_STATE_EXPIRED;
1754
1755         /*
1756          * kick transmit machine without waiting the next tick.
1757          */
1758
1759         lacp_sm_tx(lp);
1760 }
1761
1762 static void
1763 lacp_sm_rx_set_expired(struct lacp_port *lp)
1764 {
1765         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1766         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1767         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1768         lp->lp_state |= LACP_STATE_EXPIRED;
1769 }
1770
1771 static void
1772 lacp_sm_rx_timer(struct lacp_port *lp)
1773 {
1774         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1775                 /* CURRENT -> EXPIRED */
1776                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1777                 lacp_sm_rx_set_expired(lp);
1778         } else {
1779                 /* EXPIRED -> DEFAULTED */
1780                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1781                 lacp_sm_rx_update_default_selected(lp);
1782                 lacp_sm_rx_record_default(lp);
1783                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1784         }
1785 }
1786
1787 static void
1788 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1789 {
1790         boolean_t active;
1791         uint8_t oldpstate;
1792         char buf[LACP_STATESTR_MAX+1];
1793
1794         LACP_TRACE(lp);
1795
1796         oldpstate = lp->lp_partner.lip_state;
1797
1798         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1799             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1800             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1801
1802         lp->lp_partner = du->ldu_actor;
1803         if (active &&
1804             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1805             LACP_STATE_AGGREGATION) &&
1806             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1807             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1808                 /*
1809                  * XXX Maintain legacy behavior of leaving the
1810                  * LACP_STATE_SYNC bit unchanged from the partner's
1811                  * advertisement if lsc_strict_mode is false.
1812                  * TODO: We should re-examine the concept of the "strict mode"
1813                  * to ensure it makes sense to maintain a non-strict mode.
1814                  */
1815                 if (lp->lp_lsc->lsc_strict_mode)
1816                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1817         } else {
1818                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1819         }
1820
1821         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1822
1823         if (oldpstate != lp->lp_partner.lip_state) {
1824                 LACP_DPRINTF((lp, "old pstate %s\n",
1825                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1826                 LACP_DPRINTF((lp, "new pstate %s\n",
1827                     lacp_format_state(lp->lp_partner.lip_state, buf,
1828                     sizeof(buf))));
1829         }
1830
1831         lacp_sm_ptx_update_timeout(lp, oldpstate);
1832 }
1833
1834 static void
1835 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1836 {
1837
1838         LACP_TRACE(lp);
1839
1840         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1841             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1842             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1843                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1844                 lacp_sm_assert_ntt(lp);
1845         }
1846 }
1847
1848 static void
1849 lacp_sm_rx_record_default(struct lacp_port *lp)
1850 {
1851         uint8_t oldpstate;
1852
1853         LACP_TRACE(lp);
1854
1855         oldpstate = lp->lp_partner.lip_state;
1856         if (lp->lp_lsc->lsc_strict_mode)
1857                 lp->lp_partner = lacp_partner_admin_strict;
1858         else
1859                 lp->lp_partner = lacp_partner_admin_optimistic;
1860         lp->lp_state |= LACP_STATE_DEFAULTED;
1861         lacp_sm_ptx_update_timeout(lp, oldpstate);
1862 }
1863
1864 static void
1865 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1866     const struct lacp_peerinfo *info)
1867 {
1868
1869         LACP_TRACE(lp);
1870
1871         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1872             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1873             LACP_STATE_AGGREGATION)) {
1874                 lp->lp_selected = LACP_UNSELECTED;
1875                 /* mux machine will clean up lp->lp_aggregator */
1876         }
1877 }
1878
1879 static void
1880 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1881 {
1882
1883         LACP_TRACE(lp);
1884
1885         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1886 }
1887
1888 static void
1889 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1890 {
1891
1892         LACP_TRACE(lp);
1893
1894         if (lp->lp_lsc->lsc_strict_mode)
1895                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1896                     &lacp_partner_admin_strict);
1897         else
1898                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1899                     &lacp_partner_admin_optimistic);
1900 }
1901
1902 /* transmit machine */
1903
1904 static void
1905 lacp_sm_tx(struct lacp_port *lp)
1906 {
1907         int error = 0;
1908
1909         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1910 #if 1
1911             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1912             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1913 #endif
1914             ) {
1915                 lp->lp_flags &= ~LACP_PORT_NTT;
1916         }
1917
1918         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1919                 return;
1920         }
1921
1922         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1923         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1924                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1925                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1926                 return;
1927         }
1928
1929         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1930                 error = lacp_xmit_lacpdu(lp);
1931         } else {
1932                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1933         }
1934
1935         if (error == 0) {
1936                 lp->lp_flags &= ~LACP_PORT_NTT;
1937         } else {
1938                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1939                     error));
1940         }
1941 }
1942
1943 static void
1944 lacp_sm_assert_ntt(struct lacp_port *lp)
1945 {
1946
1947         lp->lp_flags |= LACP_PORT_NTT;
1948 }
1949
1950 static void
1951 lacp_run_timers(struct lacp_port *lp)
1952 {
1953         int i;
1954
1955         for (i = 0; i < LACP_NTIMER; i++) {
1956                 KASSERT(lp->lp_timer[i] >= 0,
1957                     ("invalid timer value %d", lp->lp_timer[i]));
1958                 if (lp->lp_timer[i] == 0) {
1959                         continue;
1960                 } else if (--lp->lp_timer[i] <= 0) {
1961                         if (lacp_timer_funcs[i]) {
1962                                 (*lacp_timer_funcs[i])(lp);
1963                         }
1964                 }
1965         }
1966 }
1967
1968 int
1969 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1970 {
1971         struct lacp_softc *lsc = lp->lp_lsc;
1972         struct lagg_port *lgp = lp->lp_lagg;
1973         struct lacp_port *lp2;
1974         struct markerdu *mdu;
1975         int error = 0;
1976         int pending = 0;
1977
1978         if (m->m_pkthdr.len != sizeof(*mdu)) {
1979                 goto bad;
1980         }
1981
1982         if ((m->m_flags & M_MCAST) == 0) {
1983                 goto bad;
1984         }
1985
1986         if (m->m_len < sizeof(*mdu)) {
1987                 m = m_pullup(m, sizeof(*mdu));
1988                 if (m == NULL) {
1989                         return (ENOMEM);
1990                 }
1991         }
1992
1993         mdu = mtod(m, struct markerdu *);
1994
1995         if (memcmp(&mdu->mdu_eh.ether_dhost,
1996             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1997                 goto bad;
1998         }
1999
2000         if (mdu->mdu_sph.sph_version != 1) {
2001                 goto bad;
2002         }
2003
2004         switch (mdu->mdu_tlv.tlv_type) {
2005         case MARKER_TYPE_INFO:
2006                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2007                     marker_info_tlv_template, TRUE)) {
2008                         goto bad;
2009                 }
2010                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
2011                 memcpy(&mdu->mdu_eh.ether_dhost,
2012                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
2013                 memcpy(&mdu->mdu_eh.ether_shost,
2014                     lgp->lp_lladdr, ETHER_ADDR_LEN);
2015                 error = lagg_enqueue(lp->lp_ifp, m);
2016                 break;
2017
2018         case MARKER_TYPE_RESPONSE:
2019                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2020                     marker_response_tlv_template, TRUE)) {
2021                         goto bad;
2022                 }
2023                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
2024                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
2025                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
2026
2027                 /* Verify that it is the last marker we sent out */
2028                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
2029                     sizeof(struct lacp_markerinfo)))
2030                         goto bad;
2031
2032                 LACP_LOCK(lsc);
2033                 lp->lp_flags &= ~LACP_PORT_MARK;
2034
2035                 if (lsc->lsc_suppress_distributing) {
2036                         /* Check if any ports are waiting for a response */
2037                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
2038                                 if (lp2->lp_flags & LACP_PORT_MARK) {
2039                                         pending = 1;
2040                                         break;
2041                                 }
2042                         }
2043
2044                         if (pending == 0) {
2045                                 /* All interface queues are clear */
2046                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
2047                                 lsc->lsc_suppress_distributing = FALSE;
2048                         }
2049                 }
2050                 LACP_UNLOCK(lsc);
2051                 m_freem(m);
2052                 break;
2053
2054         default:
2055                 goto bad;
2056         }
2057
2058         return (error);
2059
2060 bad:
2061         LACP_DPRINTF((lp, "bad marker frame\n"));
2062         m_freem(m);
2063         return (EINVAL);
2064 }
2065
2066 static int
2067 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
2068     const struct tlv_template *tmpl, boolean_t check_type)
2069 {
2070         while (/* CONSTCOND */ 1) {
2071                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
2072                         return (EINVAL);
2073                 }
2074                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
2075                     tlv->tlv_length != tmpl->tmpl_length) {
2076                         return (EINVAL);
2077                 }
2078                 if (tmpl->tmpl_type == 0) {
2079                         break;
2080                 }
2081                 tlv = (const struct tlvhdr *)
2082                     ((const char *)tlv + tlv->tlv_length);
2083                 tmpl++;
2084         }
2085
2086         return (0);
2087 }
2088
2089 /* Debugging */
2090 const char *
2091 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
2092 {
2093         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
2094             (int)mac[0],
2095             (int)mac[1],
2096             (int)mac[2],
2097             (int)mac[3],
2098             (int)mac[4],
2099             (int)mac[5]);
2100
2101         return (buf);
2102 }
2103
2104 const char *
2105 lacp_format_systemid(const struct lacp_systemid *sysid,
2106     char *buf, size_t buflen)
2107 {
2108         char macbuf[LACP_MACSTR_MAX+1];
2109
2110         snprintf(buf, buflen, "%04X,%s",
2111             ntohs(sysid->lsi_prio),
2112             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2113
2114         return (buf);
2115 }
2116
2117 const char *
2118 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2119 {
2120         snprintf(buf, buflen, "%04X,%04X",
2121             ntohs(portid->lpi_prio),
2122             ntohs(portid->lpi_portno));
2123
2124         return (buf);
2125 }
2126
2127 const char *
2128 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2129 {
2130         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2131         char portid[LACP_PORTIDSTR_MAX+1];
2132
2133         snprintf(buf, buflen, "(%s,%04X,%s)",
2134             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2135             ntohs(peer->lip_key),
2136             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2137
2138         return (buf);
2139 }
2140
2141 const char *
2142 lacp_format_lagid(const struct lacp_peerinfo *a,
2143     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2144 {
2145         char astr[LACP_PARTNERSTR_MAX+1];
2146         char bstr[LACP_PARTNERSTR_MAX+1];
2147
2148 #if 0
2149         /*
2150          * there's a convention to display small numbered peer
2151          * in the left.
2152          */
2153
2154         if (lacp_compare_peerinfo(a, b) > 0) {
2155                 const struct lacp_peerinfo *t;
2156
2157                 t = a;
2158                 a = b;
2159                 b = t;
2160         }
2161 #endif
2162
2163         snprintf(buf, buflen, "[%s,%s]",
2164             lacp_format_partner(a, astr, sizeof(astr)),
2165             lacp_format_partner(b, bstr, sizeof(bstr)));
2166
2167         return (buf);
2168 }
2169
2170 const char *
2171 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2172     char *buf, size_t buflen)
2173 {
2174         if (la == NULL) {
2175                 return ("(none)");
2176         }
2177
2178         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2179 }
2180
2181 const char *
2182 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2183 {
2184         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2185         return (buf);
2186 }
2187
2188 static void
2189 lacp_dump_lacpdu(const struct lacpdu *du)
2190 {
2191         char buf[LACP_PARTNERSTR_MAX+1];
2192         char buf2[LACP_STATESTR_MAX+1];
2193
2194         printf("actor=%s\n",
2195             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2196         printf("actor.state=%s\n",
2197             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2198         printf("partner=%s\n",
2199             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2200         printf("partner.state=%s\n",
2201             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2202
2203         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2204 }
2205
2206 static void
2207 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2208 {
2209         va_list va;
2210
2211         if (lp) {
2212                 printf("%s: ", lp->lp_ifp->if_xname);
2213         }
2214
2215         va_start(va, fmt);
2216         vprintf(fmt, va);
2217         va_end(va);
2218 }