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_kern_tls.h"
  36 #include "opt_ratelimit.h"
  37
  38 #include <sys/param.h>
  39 #include <sys/callout.h>
  40 #include <sys/eventhandler.h>
  41 #include <sys/mbuf.h>
  42 #include <sys/systm.h>
  43 #include <sys/malloc.h>
  44 #include <sys/kernel.h> /* hz */
  45 #include <sys/socket.h> /* for net/if.h */
  46 #include <sys/sockio.h>
  47 #include <sys/sysctl.h>
  48 #include <machine/stdarg.h>
  49 #include <sys/lock.h>
  50 #include <sys/rwlock.h>
  51 #include <sys/taskqueue.h>
  52
  53 #include <net/if.h>
  54 #include <net/if_var.h>
  55 #include <net/if_dl.h>
  56 #include <net/ethernet.h>
  57 #include <net/infiniband.h>
  58 #include <net/if_media.h>
  59 #include <net/if_types.h>
  60
  61 #include <net/if_lagg.h>
  62 #include <net/ieee8023ad_lacp.h>
  63
  64 /*
  65  * actor system priority and port priority.
  66  * XXX should be configurable.
  67  */
  68
  69 #define LACP_SYSTEM_PRIO        0x8000
  70 #define LACP_PORT_PRIO          0x8000
  71
  72 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  73     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  74
  75 static const struct tlv_template lacp_info_tlv_template[] = {
  76         { LACP_TYPE_ACTORINFO,
  77             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  78         { LACP_TYPE_PARTNERINFO,
  79             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  80         { LACP_TYPE_COLLECTORINFO,
  81             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  82         { 0, 0 },
  83 };
  84
  85 static const struct tlv_template marker_info_tlv_template[] = {
  86         { MARKER_TYPE_INFO,
  87             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  88         { 0, 0 },
  89 };
  90
  91 static const struct tlv_template marker_response_tlv_template[] = {
  92         { MARKER_TYPE_RESPONSE,
  93             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  94         { 0, 0 },
  95 };
  96
  97 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  98
  99 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
 100 static void     lacp_fill_markerinfo(struct lacp_port *,
 101                     struct lacp_markerinfo *);
 102
 103 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
 104 static void     lacp_suppress_distributing(struct lacp_softc *,
 105                     struct lacp_aggregator *);
 106 static void     lacp_transit_expire(void *);
 107 static void     lacp_update_portmap(struct lacp_softc *);
 108 static void     lacp_select_active_aggregator(struct lacp_softc *);
 109 static uint16_t lacp_compose_key(struct lacp_port *);
 110 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 111                     const struct tlv_template *, boolean_t);
 112 static void     lacp_tick(void *);
 113
 114 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 115                     const struct lacp_port *);
 116 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 117                     const struct lacp_peerinfo *);
 118 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 119                     const struct lacp_port *);
 120 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 121                     const struct lacp_peerinfo *);
 122
 123 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 124                     struct lacp_port *);
 125 static void     lacp_aggregator_addref(struct lacp_softc *,
 126                     struct lacp_aggregator *);
 127 static void     lacp_aggregator_delref(struct lacp_softc *,
 128                     struct lacp_aggregator *);
 129
 130 /* receive machine */
 131
 132 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 133 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 134 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 135 static void     lacp_sm_rx_timer(struct lacp_port *);
 136 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 137 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 138                     const struct lacpdu *);
 139 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 140                     const struct lacpdu *);
 141 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 142                     const struct lacpdu *);
 143 static void     lacp_sm_rx_record_default(struct lacp_port *);
 144 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 145 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 146                     const struct lacp_peerinfo *);
 147
 148 /* mux machine */
 149
 150 static void     lacp_sm_mux(struct lacp_port *);
 151 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 152 static void     lacp_sm_mux_timer(struct lacp_port *);
 153
 154 /* periodic transmit machine */
 155
 156 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 157 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 158 static void     lacp_sm_ptx_timer(struct lacp_port *);
 159
 160 /* transmit machine */
 161
 162 static void     lacp_sm_tx(struct lacp_port *);
 163 static void     lacp_sm_assert_ntt(struct lacp_port *);
 164
 165 static void     lacp_run_timers(struct lacp_port *);
 166 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 167                     const struct lacp_peerinfo *);
 168 static int      lacp_compare_systemid(const struct lacp_systemid *,
 169                     const struct lacp_systemid *);
 170 static void     lacp_port_enable(struct lacp_port *);
 171 static void     lacp_port_disable(struct lacp_port *);
 172 static void     lacp_select(struct lacp_port *);
 173 static void     lacp_unselect(struct lacp_port *);
 174 static void     lacp_disable_collecting(struct lacp_port *);
 175 static void     lacp_enable_collecting(struct lacp_port *);
 176 static void     lacp_disable_distributing(struct lacp_port *);
 177 static void     lacp_enable_distributing(struct lacp_port *);
 178 static int      lacp_xmit_lacpdu(struct lacp_port *);
 179 static int      lacp_xmit_marker(struct lacp_port *);
 180
 181 /* Debugging */
 182
 183 static void     lacp_dump_lacpdu(const struct lacpdu *);
 184 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 185                     size_t);
 186 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 187                     const struct lacp_peerinfo *, char *, size_t);
 188 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 189                     char *, size_t);
 190 static const char *lacp_format_state(uint8_t, char *, size_t);
 191 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 192 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 193                     size_t);
 194 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 195                     size_t);
 196 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 197                     __attribute__((__format__(__printf__, 2, 3)));
 198
 199 VNET_DEFINE_STATIC(int, lacp_debug);
 200 #define V_lacp_debug    VNET(lacp_debug)
 201 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
 202     "ieee802.3ad");
 203 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 204     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 205
 206 VNET_DEFINE_STATIC(int, lacp_default_strict_mode) = 1;
 207 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode,
 208     CTLFLAG_RWTUN | CTLFLAG_VNET, &VNET_NAME(lacp_default_strict_mode), 0,
 209     "LACP strict protocol compliance default");
 210 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 211 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 212 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 213
 214 /*
 215  * partner administration variables.
 216  * XXX should be configurable.
 217  */
 218
 219 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 220         .lip_systemid = { .lsi_prio = 0xffff },
 221         .lip_portid = { .lpi_prio = 0xffff },
 222         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 223             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 224 };
 225
 226 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 227         .lip_systemid = { .lsi_prio = 0xffff },
 228         .lip_portid = { .lpi_prio = 0xffff },
 229         .lip_state = 0,
 230 };
 231
 232 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 233         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 234         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 235         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 236 };
 237
 238 struct mbuf *
 239 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 240 {
 241         struct lacp_port *lp = LACP_PORT(lgp);
 242         uint8_t subtype;
 243
 244         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 245                 m_freem(m);
 246                 return (NULL);
 247         }
 248
 249         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 250         switch (subtype) {
 251                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 252                         lacp_pdu_input(lp, m);
 253                         return (NULL);
 254
 255                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 256                         lacp_marker_input(lp, m);
 257                         return (NULL);
 258         }
 259
 260         /* Not a subtype we are interested in */
 261         return (m);
 262 }
 263
 264 /*
 265  * lacp_pdu_input: process lacpdu
 266  */
 267 static int
 268 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 269 {
 270         struct lacp_softc *lsc = lp->lp_lsc;
 271         struct lacpdu *du;
 272         int error = 0;
 273
 274         if (m->m_pkthdr.len != sizeof(*du)) {
 275                 goto bad;
 276         }
 277
 278         if ((m->m_flags & M_MCAST) == 0) {
 279                 goto bad;
 280         }
 281
 282         if (m->m_len < sizeof(*du)) {
 283                 m = m_pullup(m, sizeof(*du));
 284                 if (m == NULL) {
 285                         return (ENOMEM);
 286                 }
 287         }
 288
 289         du = mtod(m, struct lacpdu *);
 290
 291         if (memcmp(&du->ldu_eh.ether_dhost,
 292             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 293                 goto bad;
 294         }
 295
 296         /*
 297          * ignore the version for compatibility with
 298          * the future protocol revisions.
 299          */
 300 #if 0
 301         if (du->ldu_sph.sph_version != 1) {
 302                 goto bad;
 303         }
 304 #endif
 305
 306         /*
 307          * ignore tlv types for compatibility with
 308          * the future protocol revisions.
 309          */
 310         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 311             lacp_info_tlv_template, FALSE)) {
 312                 goto bad;
 313         }
 314
 315         if (V_lacp_debug > 0) {
 316                 lacp_dprintf(lp, "lacpdu receive\n");
 317                 lacp_dump_lacpdu(du);
 318         }
 319
 320         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 321                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 322                 goto bad;
 323         }
 324
 325         LACP_LOCK(lsc);
 326         lacp_sm_rx(lp, du);
 327         LACP_UNLOCK(lsc);
 328
 329         m_freem(m);
 330         return (error);
 331
 332 bad:
 333         m_freem(m);
 334         return (EINVAL);
 335 }
 336
 337 static void
 338 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 339 {
 340         struct lagg_port *lgp = lp->lp_lagg;
 341         struct lagg_softc *sc = lgp->lp_softc;
 342
 343         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 344         memcpy(&info->lip_systemid.lsi_mac,
 345             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 346         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 347         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 348         info->lip_state = lp->lp_state;
 349 }
 350
 351 static void
 352 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 353 {
 354         struct ifnet *ifp = lp->lp_ifp;
 355
 356         /* Fill in the port index and system id (encoded as the MAC) */
 357         info->mi_rq_port = htons(ifp->if_index);
 358         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 359         info->mi_rq_xid = htonl(0);
 360 }
 361
 362 static int
 363 lacp_xmit_lacpdu(struct lacp_port *lp)
 364 {
 365         struct lagg_port *lgp = lp->lp_lagg;
 366         struct mbuf *m;
 367         struct lacpdu *du;
 368         int error;
 369
 370         LACP_LOCK_ASSERT(lp->lp_lsc);
 371
 372         m = m_gethdr(M_NOWAIT, MT_DATA);
 373         if (m == NULL) {
 374                 return (ENOMEM);
 375         }
 376         m->m_len = m->m_pkthdr.len = sizeof(*du);
 377
 378         du = mtod(m, struct lacpdu *);
 379         memset(du, 0, sizeof(*du));
 380
 381         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 382             ETHER_ADDR_LEN);
 383         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 384         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 385
 386         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 387         du->ldu_sph.sph_version = 1;
 388
 389         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 390         du->ldu_actor = lp->lp_actor;
 391
 392         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 393             sizeof(du->ldu_partner));
 394         du->ldu_partner = lp->lp_partner;
 395
 396         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 397             sizeof(du->ldu_collector));
 398         du->ldu_collector.lci_maxdelay = 0;
 399
 400         if (V_lacp_debug > 0) {
 401                 lacp_dprintf(lp, "lacpdu transmit\n");
 402                 lacp_dump_lacpdu(du);
 403         }
 404
 405         m->m_flags |= M_MCAST;
 406
 407         /*
 408          * XXX should use higher priority queue.
 409          * otherwise network congestion can break aggregation.
 410          */
 411
 412         error = lagg_enqueue(lp->lp_ifp, m);
 413         return (error);
 414 }
 415
 416 static int
 417 lacp_xmit_marker(struct lacp_port *lp)
 418 {
 419         struct lagg_port *lgp = lp->lp_lagg;
 420         struct mbuf *m;
 421         struct markerdu *mdu;
 422         int error;
 423
 424         LACP_LOCK_ASSERT(lp->lp_lsc);
 425
 426         m = m_gethdr(M_NOWAIT, MT_DATA);
 427         if (m == NULL) {
 428                 return (ENOMEM);
 429         }
 430         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 431
 432         mdu = mtod(m, struct markerdu *);
 433         memset(mdu, 0, sizeof(*mdu));
 434
 435         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 436             ETHER_ADDR_LEN);
 437         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 438         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 439
 440         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 441         mdu->mdu_sph.sph_version = 1;
 442
 443         /* Bump the transaction id and copy over the marker info */
 444         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 445         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 446         mdu->mdu_info = lp->lp_marker;
 447
 448         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 449             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 450             ntohl(mdu->mdu_info.mi_rq_xid)));
 451
 452         m->m_flags |= M_MCAST;
 453         error = lagg_enqueue(lp->lp_ifp, m);
 454         return (error);
 455 }
 456
 457 void
 458 lacp_linkstate(struct lagg_port *lgp)
 459 {
 460         struct lacp_port *lp = LACP_PORT(lgp);
 461         struct lacp_softc *lsc = lp->lp_lsc;
 462         struct ifnet *ifp = lgp->lp_ifp;
 463         struct ifmediareq ifmr;
 464         int error = 0;
 465         u_int media;
 466         uint8_t old_state;
 467         uint16_t old_key;
 468
 469         bzero((char *)&ifmr, sizeof(ifmr));
 470         error = (*ifp->if_ioctl)(ifp, SIOCGIFXMEDIA, (caddr_t)&ifmr);
 471         if (error != 0) {
 472                 bzero((char *)&ifmr, sizeof(ifmr));
 473                 error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 474         }
 475         if (error != 0)
 476                 return;
 477
 478         LACP_LOCK(lsc);
 479         media = ifmr.ifm_active;
 480         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 481             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 482             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 483         old_state = lp->lp_state;
 484         old_key = lp->lp_key;
 485
 486         lp->lp_media = media;
 487         /*
 488          * If the port is not an active full duplex Ethernet link then it can
 489          * not be aggregated.
 490          */
 491         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 492             ifp->if_link_state != LINK_STATE_UP) {
 493                 lacp_port_disable(lp);
 494         } else {
 495                 lacp_port_enable(lp);
 496         }
 497         lp->lp_key = lacp_compose_key(lp);
 498
 499         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 500                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 501                 lp->lp_selected = LACP_UNSELECTED;
 502         }
 503         LACP_UNLOCK(lsc);
 504 }
 505
 506 static void
 507 lacp_tick(void *arg)
 508 {
 509         struct lacp_softc *lsc = arg;
 510         struct lacp_port *lp;
 511
 512         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 513                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 514                         continue;
 515
 516                 CURVNET_SET(lp->lp_ifp->if_vnet);
 517                 lacp_run_timers(lp);
 518
 519                 lacp_select(lp);
 520                 lacp_sm_mux(lp);
 521                 lacp_sm_tx(lp);
 522                 lacp_sm_ptx_tx_schedule(lp);
 523                 CURVNET_RESTORE();
 524         }
 525         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 526 }
 527
 528 int
 529 lacp_port_create(struct lagg_port *lgp)
 530 {
 531         struct lagg_softc *sc = lgp->lp_softc;
 532         struct lacp_softc *lsc = LACP_SOFTC(sc);
 533         struct lacp_port *lp;
 534         struct ifnet *ifp = lgp->lp_ifp;
 535         struct sockaddr_dl sdl;
 536         struct ifmultiaddr *rifma = NULL;
 537         int error;
 538
 539         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 540         sdl.sdl_alen = ETHER_ADDR_LEN;
 541
 542         bcopy(&ethermulticastaddr_slowprotocols,
 543             LLADDR(&sdl), ETHER_ADDR_LEN);
 544         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 545         if (error) {
 546                 printf("%s: ADDMULTI failed on %s\n", __func__,
 547                     lgp->lp_ifp->if_xname);
 548                 return (error);
 549         }
 550
 551         lp = malloc(sizeof(struct lacp_port),
 552             M_DEVBUF, M_NOWAIT|M_ZERO);
 553         if (lp == NULL)
 554                 return (ENOMEM);
 555
 556         LACP_LOCK(lsc);
 557         lgp->lp_psc = lp;
 558         lp->lp_ifp = ifp;
 559         lp->lp_lagg = lgp;
 560         lp->lp_lsc = lsc;
 561         lp->lp_ifma = rifma;
 562
 563         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 564
 565         lacp_fill_actorinfo(lp, &lp->lp_actor);
 566         lacp_fill_markerinfo(lp, &lp->lp_marker);
 567         lp->lp_state = LACP_STATE_ACTIVITY;
 568         lp->lp_aggregator = NULL;
 569         lacp_sm_rx_set_expired(lp);
 570         LACP_UNLOCK(lsc);
 571         lacp_linkstate(lgp);
 572
 573         return (0);
 574 }
 575
 576 void
 577 lacp_port_destroy(struct lagg_port *lgp)
 578 {
 579         struct lacp_port *lp = LACP_PORT(lgp);
 580         struct lacp_softc *lsc = lp->lp_lsc;
 581         int i;
 582
 583         LACP_LOCK(lsc);
 584         for (i = 0; i < LACP_NTIMER; i++) {
 585                 LACP_TIMER_DISARM(lp, i);
 586         }
 587
 588         lacp_disable_collecting(lp);
 589         lacp_disable_distributing(lp);
 590         lacp_unselect(lp);
 591
 592         LIST_REMOVE(lp, lp_next);
 593         LACP_UNLOCK(lsc);
 594
 595         /* The address may have already been removed by if_purgemaddrs() */
 596         if (!lgp->lp_detaching)
 597                 if_delmulti_ifma(lp->lp_ifma);
 598
 599         free(lp, M_DEVBUF);
 600 }
 601
 602 void
 603 lacp_req(struct lagg_softc *sc, void *data)
 604 {
 605         struct lacp_opreq *req = (struct lacp_opreq *)data;
 606         struct lacp_softc *lsc = LACP_SOFTC(sc);
 607         struct lacp_aggregator *la;
 608
 609         bzero(req, sizeof(struct lacp_opreq));
 610
 611         /*
 612          * If the LACP softc is NULL, return with the opreq structure full of
 613          * zeros.  It is normal for the softc to be NULL while the lagg is
 614          * being destroyed.
 615          */
 616         if (NULL == lsc)
 617                 return;
 618
 619         la = lsc->lsc_active_aggregator;
 620         LACP_LOCK(lsc);
 621         if (la != NULL) {
 622                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 623                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 624                     ETHER_ADDR_LEN);
 625                 req->actor_key = ntohs(la->la_actor.lip_key);
 626                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 627                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 628                 req->actor_state = la->la_actor.lip_state;
 629
 630                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 631                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 632                     ETHER_ADDR_LEN);
 633                 req->partner_key = ntohs(la->la_partner.lip_key);
 634                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 635                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 636                 req->partner_state = la->la_partner.lip_state;
 637         }
 638         LACP_UNLOCK(lsc);
 639 }
 640
 641 void
 642 lacp_portreq(struct lagg_port *lgp, void *data)
 643 {
 644         struct lacp_opreq *req = (struct lacp_opreq *)data;
 645         struct lacp_port *lp = LACP_PORT(lgp);
 646         struct lacp_softc *lsc = lp->lp_lsc;
 647
 648         LACP_LOCK(lsc);
 649         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 650         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 651             ETHER_ADDR_LEN);
 652         req->actor_key = ntohs(lp->lp_actor.lip_key);
 653         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 654         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 655         req->actor_state = lp->lp_actor.lip_state;
 656
 657         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 658         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 659             ETHER_ADDR_LEN);
 660         req->partner_key = ntohs(lp->lp_partner.lip_key);
 661         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 662         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 663         req->partner_state = lp->lp_partner.lip_state;
 664         LACP_UNLOCK(lsc);
 665 }
 666
 667 static void
 668 lacp_disable_collecting(struct lacp_port *lp)
 669 {
 670         LACP_DPRINTF((lp, "collecting disabled\n"));
 671         lp->lp_state &= ~LACP_STATE_COLLECTING;
 672 }
 673
 674 static void
 675 lacp_enable_collecting(struct lacp_port *lp)
 676 {
 677         LACP_DPRINTF((lp, "collecting enabled\n"));
 678         lp->lp_state |= LACP_STATE_COLLECTING;
 679 }
 680
 681 static void
 682 lacp_disable_distributing(struct lacp_port *lp)
 683 {
 684         struct lacp_aggregator *la = lp->lp_aggregator;
 685         struct lacp_softc *lsc = lp->lp_lsc;
 686         struct lagg_softc *sc = lsc->lsc_softc;
 687         char buf[LACP_LAGIDSTR_MAX+1];
 688
 689         LACP_LOCK_ASSERT(lsc);
 690
 691         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 692                 return;
 693         }
 694
 695         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 696         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 697         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 698
 699         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 700             "nports %d -> %d\n",
 701             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 702             la->la_nports, la->la_nports - 1));
 703
 704         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 705         la->la_nports--;
 706         sc->sc_active = la->la_nports;
 707
 708         if (lsc->lsc_active_aggregator == la) {
 709                 lacp_suppress_distributing(lsc, la);
 710                 lacp_select_active_aggregator(lsc);
 711                 /* regenerate the port map, the active aggregator has changed */
 712                 lacp_update_portmap(lsc);
 713         }
 714
 715         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 716         if_link_state_change(sc->sc_ifp,
 717             sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
 718 }
 719
 720 static void
 721 lacp_enable_distributing(struct lacp_port *lp)
 722 {
 723         struct lacp_aggregator *la = lp->lp_aggregator;
 724         struct lacp_softc *lsc = lp->lp_lsc;
 725         struct lagg_softc *sc = lsc->lsc_softc;
 726         char buf[LACP_LAGIDSTR_MAX+1];
 727
 728         LACP_LOCK_ASSERT(lsc);
 729
 730         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 731                 return;
 732         }
 733
 734         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 735             "nports %d -> %d\n",
 736             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 737             la->la_nports, la->la_nports + 1));
 738
 739         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 740         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 741         la->la_nports++;
 742         sc->sc_active = la->la_nports;
 743
 744         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 745
 746         if (lsc->lsc_active_aggregator == la) {
 747                 lacp_suppress_distributing(lsc, la);
 748                 lacp_update_portmap(lsc);
 749         } else
 750                 /* try to become the active aggregator */
 751                 lacp_select_active_aggregator(lsc);
 752
 753         if_link_state_change(sc->sc_ifp,
 754             sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
 755 }
 756
 757 static void
 758 lacp_transit_expire(void *vp)
 759 {
 760         struct lacp_softc *lsc = vp;
 761
 762         LACP_LOCK_ASSERT(lsc);
 763
 764         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 765         LACP_TRACE(NULL);
 766         CURVNET_RESTORE();
 767
 768         lsc->lsc_suppress_distributing = FALSE;
 769 }
 770
 771 void
 772 lacp_attach(struct lagg_softc *sc)
 773 {
 774         struct lacp_softc *lsc;
 775
 776         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 777
 778         sc->sc_psc = lsc;
 779         lsc->lsc_softc = sc;
 780
 781         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 782         lsc->lsc_active_aggregator = NULL;
 783         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 784         LACP_LOCK_INIT(lsc);
 785         TAILQ_INIT(&lsc->lsc_aggregators);
 786         LIST_INIT(&lsc->lsc_ports);
 787
 788         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 789         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 790
 791         /* if the lagg is already up then do the same */
 792         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 793                 lacp_init(sc);
 794 }
 795
 796 void
 797 lacp_detach(void *psc)
 798 {
 799         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 800
 801         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 802             ("aggregators still active"));
 803         KASSERT(lsc->lsc_active_aggregator == NULL,
 804             ("aggregator still attached"));
 805
 806         callout_drain(&lsc->lsc_transit_callout);
 807         callout_drain(&lsc->lsc_callout);
 808
 809         LACP_LOCK_DESTROY(lsc);
 810         free(lsc, M_DEVBUF);
 811 }
 812
 813 void
 814 lacp_init(struct lagg_softc *sc)
 815 {
 816         struct lacp_softc *lsc = LACP_SOFTC(sc);
 817
 818         LACP_LOCK(lsc);
 819         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 820         LACP_UNLOCK(lsc);
 821 }
 822
 823 void
 824 lacp_stop(struct lagg_softc *sc)
 825 {
 826         struct lacp_softc *lsc = LACP_SOFTC(sc);
 827
 828         LACP_LOCK(lsc);
 829         callout_stop(&lsc->lsc_transit_callout);
 830         callout_stop(&lsc->lsc_callout);
 831         LACP_UNLOCK(lsc);
 832 }
 833
 834 struct lagg_port *
 835 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t hash,
 836     uint8_t numa_domain, int *err)
 837 {
 838         struct lacp_softc *lsc = LACP_SOFTC(sc);
 839         struct lacp_portmap *pm;
 840         struct lacp_port *lp;
 841         struct lacp_port **map;
 842         int count;
 843
 844         if (__predict_false(lsc->lsc_suppress_distributing)) {
 845                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 846                 *err = ENOBUFS;
 847                 return (NULL);
 848         }
 849
 850         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 851         if (pm->pm_count == 0) {
 852                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 853                 *err = ENETDOWN;
 854                 return (NULL);
 855         }
 856
 857 #ifdef NUMA
 858         if ((sc->sc_opts & LAGG_OPT_USE_NUMA) &&
 859             pm->pm_num_dom > 1 && numa_domain < MAXMEMDOM) {
 860                 count = pm->pm_numa[numa_domain].count;
 861                 if (count > 0) {
 862                         map = pm->pm_numa[numa_domain].map;
 863                 } else {
 864                         /* No ports on this domain; use global hash. */
 865                         map = pm->pm_map;
 866                         count = pm->pm_count;
 867                 }
 868         } else
 869 #endif
 870         {
 871                 map = pm->pm_map;
 872                 count = pm->pm_count;
 873         }
 874
 875         hash %= count;
 876         lp = map[hash];
 877
 878         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 879             ("aggregated port is not distributing"));
 880
 881         return (lp->lp_lagg);
 882 }
 883
 884 struct lagg_port *
 885 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m, int *err)
 886 {
 887         struct lacp_softc *lsc = LACP_SOFTC(sc);
 888         uint32_t hash;
 889         uint8_t numa_domain;
 890
 891         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 892             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 893                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 894         else
 895                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 896
 897         numa_domain = m->m_pkthdr.numa_domain;
 898         return (lacp_select_tx_port_by_hash(sc, hash, numa_domain, err));
 899 }
 900
 901 /*
 902  * lacp_suppress_distributing: drop transmit packets for a while
 903  * to preserve packet ordering.
 904  */
 905
 906 static void
 907 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 908 {
 909         struct lacp_port *lp;
 910
 911         if (lsc->lsc_active_aggregator != la) {
 912                 return;
 913         }
 914
 915         LACP_TRACE(NULL);
 916
 917         lsc->lsc_suppress_distributing = TRUE;
 918
 919         /* send a marker frame down each port to verify the queues are empty */
 920         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 921                 lp->lp_flags |= LACP_PORT_MARK;
 922                 lacp_xmit_marker(lp);
 923         }
 924
 925         /* set a timeout for the marker frames */
 926         callout_reset(&lsc->lsc_transit_callout,
 927             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 928 }
 929
 930 static int
 931 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 932     const struct lacp_peerinfo *b)
 933 {
 934         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 935 }
 936
 937 static int
 938 lacp_compare_systemid(const struct lacp_systemid *a,
 939     const struct lacp_systemid *b)
 940 {
 941         return (memcmp(a, b, sizeof(*a)));
 942 }
 943
 944 #if 0   /* unused */
 945 static int
 946 lacp_compare_portid(const struct lacp_portid *a,
 947     const struct lacp_portid *b)
 948 {
 949         return (memcmp(a, b, sizeof(*a)));
 950 }
 951 #endif
 952
 953 static uint64_t
 954 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 955 {
 956         struct lacp_port *lp;
 957         uint64_t speed;
 958
 959         lp = TAILQ_FIRST(&la->la_ports);
 960         if (lp == NULL) {
 961                 return (0);
 962         }
 963
 964         speed = ifmedia_baudrate(lp->lp_media);
 965         speed *= la->la_nports;
 966         if (speed == 0) {
 967                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 968                     lp->lp_media, la->la_nports));
 969         }
 970
 971         return (speed);
 972 }
 973
 974 /*
 975  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 976  * packets from lagg(4) interface.
 977  */
 978
 979 static void
 980 lacp_select_active_aggregator(struct lacp_softc *lsc)
 981 {
 982         struct lacp_aggregator *la;
 983         struct lacp_aggregator *best_la = NULL;
 984         uint64_t best_speed = 0;
 985         char buf[LACP_LAGIDSTR_MAX+1];
 986
 987         LACP_TRACE(NULL);
 988
 989         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 990                 uint64_t speed;
 991
 992                 if (la->la_nports == 0) {
 993                         continue;
 994                 }
 995
 996                 speed = lacp_aggregator_bandwidth(la);
 997                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 998                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 999                     speed, la->la_nports));
1000
1001                 /*
1002                  * This aggregator is chosen if the partner has a better
1003                  * system priority or, the total aggregated speed is higher
1004                  * or, it is already the chosen aggregator
1005                  */
1006                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
1007                     LACP_SYS_PRI(best_la->la_partner)) ||
1008                     speed > best_speed ||
1009                     (speed == best_speed &&
1010                     la == lsc->lsc_active_aggregator)) {
1011                         best_la = la;
1012                         best_speed = speed;
1013                 }
1014         }
1015
1016         KASSERT(best_la == NULL || best_la->la_nports > 0,
1017             ("invalid aggregator refcnt"));
1018         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1019             ("invalid aggregator list"));
1020
1021         if (lsc->lsc_active_aggregator != best_la) {
1022                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1023                 LACP_DPRINTF((NULL, "old %s\n",
1024                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1025                     buf, sizeof(buf))));
1026         } else {
1027                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1028         }
1029         LACP_DPRINTF((NULL, "new %s\n",
1030             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1031
1032         if (lsc->lsc_active_aggregator != best_la) {
1033                 lsc->lsc_active_aggregator = best_la;
1034                 lacp_update_portmap(lsc);
1035                 if (best_la) {
1036                         lacp_suppress_distributing(lsc, best_la);
1037                 }
1038         }
1039 }
1040
1041 /*
1042  * Updated the inactive portmap array with the new list of ports and
1043  * make it live.
1044  */
1045 static void
1046 lacp_update_portmap(struct lacp_softc *lsc)
1047 {
1048         struct lagg_softc *sc = lsc->lsc_softc;
1049         struct lacp_aggregator *la;
1050         struct lacp_portmap *p;
1051         struct lacp_port *lp;
1052         uint64_t speed;
1053         u_int newmap;
1054         int i;
1055 #ifdef NUMA
1056         int count;
1057         uint8_t domain;
1058 #endif
1059
1060         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1061         p = &lsc->lsc_pmap[newmap];
1062         la = lsc->lsc_active_aggregator;
1063         speed = 0;
1064         bzero(p, sizeof(struct lacp_portmap));
1065
1066         if (la != NULL && la->la_nports > 0) {
1067                 p->pm_count = la->la_nports;
1068                 i = 0;
1069                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q) {
1070                         p->pm_map[i++] = lp;
1071 #ifdef NUMA
1072                         domain = lp->lp_ifp->if_numa_domain;
1073                         if (domain >= MAXMEMDOM)
1074                                 continue;
1075                         count = p->pm_numa[domain].count;
1076                         p->pm_numa[domain].map[count] = lp;
1077                         p->pm_numa[domain].count++;
1078 #endif
1079                 }
1080                 KASSERT(i == p->pm_count, ("Invalid port count"));
1081
1082 #ifdef NUMA
1083                 for (i = 0; i < MAXMEMDOM; i++) {
1084                         if (p->pm_numa[i].count != 0)
1085                                 p->pm_num_dom++;
1086                 }
1087 #endif
1088                 speed = lacp_aggregator_bandwidth(la);
1089         }
1090         sc->sc_ifp->if_baudrate = speed;
1091
1092         /* switch the active portmap over */
1093         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1094         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1095                     lsc->lsc_activemap,
1096                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1097 }
1098
1099 static uint16_t
1100 lacp_compose_key(struct lacp_port *lp)
1101 {
1102         struct lagg_port *lgp = lp->lp_lagg;
1103         struct lagg_softc *sc = lgp->lp_softc;
1104         u_int media = lp->lp_media;
1105         uint16_t key;
1106
1107         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1108                 /*
1109                  * non-aggregatable links should have unique keys.
1110                  *
1111                  * XXX this isn't really unique as if_index is 16 bit.
1112                  */
1113
1114                 /* bit 0..14:   (some bits of) if_index of this port */
1115                 key = lp->lp_ifp->if_index;
1116                 /* bit 15:      1 */
1117                 key |= 0x8000;
1118         } else {
1119                 u_int subtype = IFM_SUBTYPE(media);
1120
1121                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1122                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1123
1124                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1125                 switch (subtype) {
1126                 case IFM_10_T:
1127                 case IFM_10_2:
1128                 case IFM_10_5:
1129                 case IFM_10_STP:
1130                 case IFM_10_FL:
1131                         key = IFM_10_T;
1132                         break;
1133                 case IFM_100_TX:
1134                 case IFM_100_FX:
1135                 case IFM_100_T4:
1136                 case IFM_100_VG:
1137                 case IFM_100_T2:
1138                 case IFM_100_T:
1139                 case IFM_100_SGMII:
1140                         key = IFM_100_TX;
1141                         break;
1142                 case IFM_1000_SX:
1143                 case IFM_1000_LX:
1144                 case IFM_1000_CX:
1145                 case IFM_1000_T:
1146                 case IFM_1000_KX:
1147                 case IFM_1000_SGMII:
1148                 case IFM_1000_CX_SGMII:
1149                         key = IFM_1000_SX;
1150                         break;
1151                 case IFM_10G_LR:
1152                 case IFM_10G_SR:
1153                 case IFM_10G_CX4:
1154                 case IFM_10G_TWINAX:
1155                 case IFM_10G_TWINAX_LONG:
1156                 case IFM_10G_LRM:
1157                 case IFM_10G_T:
1158                 case IFM_10G_KX4:
1159                 case IFM_10G_KR:
1160                 case IFM_10G_CR1:
1161                 case IFM_10G_ER:
1162                 case IFM_10G_SFI:
1163                 case IFM_10G_AOC:
1164                         key = IFM_10G_LR;
1165                         break;
1166                 case IFM_20G_KR2:
1167                         key = IFM_20G_KR2;
1168                         break;
1169                 case IFM_2500_KX:
1170                 case IFM_2500_T:
1171                 case IFM_2500_X:
1172                         key = IFM_2500_KX;
1173                         break;
1174                 case IFM_5000_T:
1175                 case IFM_5000_KR:
1176                 case IFM_5000_KR_S:
1177                 case IFM_5000_KR1:
1178                         key = IFM_5000_T;
1179                         break;
1180                 case IFM_50G_PCIE:
1181                 case IFM_50G_CR2:
1182                 case IFM_50G_KR2:
1183                 case IFM_50G_KR4:
1184                 case IFM_50G_SR2:
1185                 case IFM_50G_LR2:
1186                 case IFM_50G_LAUI2_AC:
1187                 case IFM_50G_LAUI2:
1188                 case IFM_50G_AUI2_AC:
1189                 case IFM_50G_AUI2:
1190                 case IFM_50G_CP:
1191                 case IFM_50G_SR:
1192                 case IFM_50G_LR:
1193                 case IFM_50G_FR:
1194                 case IFM_50G_KR_PAM4:
1195                 case IFM_50G_AUI1_AC:
1196                 case IFM_50G_AUI1:
1197                         key = IFM_50G_PCIE;
1198                         break;
1199                 case IFM_56G_R4:
1200                         key = IFM_56G_R4;
1201                         break;
1202                 case IFM_25G_PCIE:
1203                 case IFM_25G_CR:
1204                 case IFM_25G_KR:
1205                 case IFM_25G_SR:
1206                 case IFM_25G_LR:
1207                 case IFM_25G_ACC:
1208                 case IFM_25G_AOC:
1209                 case IFM_25G_T:
1210                 case IFM_25G_CR_S:
1211                 case IFM_25G_CR1:
1212                 case IFM_25G_KR_S:
1213                 case IFM_25G_AUI:
1214                 case IFM_25G_KR1:
1215                         key = IFM_25G_PCIE;
1216                         break;
1217                 case IFM_40G_CR4:
1218                 case IFM_40G_SR4:
1219                 case IFM_40G_LR4:
1220                 case IFM_40G_LM4:
1221                 case IFM_40G_XLPPI:
1222                 case IFM_40G_KR4:
1223                 case IFM_40G_XLAUI:
1224                 case IFM_40G_XLAUI_AC:
1225                 case IFM_40G_ER4:
1226                         key = IFM_40G_CR4;
1227                         break;
1228                 case IFM_100G_CR4:
1229                 case IFM_100G_SR4:
1230                 case IFM_100G_KR4:
1231                 case IFM_100G_LR4:
1232                 case IFM_100G_CAUI4_AC:
1233                 case IFM_100G_CAUI4:
1234                 case IFM_100G_AUI4_AC:
1235                 case IFM_100G_AUI4:
1236                 case IFM_100G_CR_PAM4:
1237                 case IFM_100G_KR_PAM4:
1238                 case IFM_100G_CP2:
1239                 case IFM_100G_SR2:
1240                 case IFM_100G_DR:
1241                 case IFM_100G_KR2_PAM4:
1242                 case IFM_100G_CAUI2_AC:
1243                 case IFM_100G_CAUI2:
1244                 case IFM_100G_AUI2_AC:
1245                 case IFM_100G_AUI2:
1246                         key = IFM_100G_CR4;
1247                         break;
1248                 case IFM_200G_CR4_PAM4:
1249                 case IFM_200G_SR4:
1250                 case IFM_200G_FR4:
1251                 case IFM_200G_LR4:
1252                 case IFM_200G_DR4:
1253                 case IFM_200G_KR4_PAM4:
1254                 case IFM_200G_AUI4_AC:
1255                 case IFM_200G_AUI4:
1256                 case IFM_200G_AUI8_AC:
1257                 case IFM_200G_AUI8:
1258                         key = IFM_200G_CR4_PAM4;
1259                         break;
1260                 case IFM_400G_FR8:
1261                 case IFM_400G_LR8:
1262                 case IFM_400G_DR4:
1263                 case IFM_400G_AUI8_AC:
1264                 case IFM_400G_AUI8:
1265                         key = IFM_400G_FR8;
1266                         break;
1267                 default:
1268                         key = subtype;
1269                         break;
1270                 }
1271                 /* bit 5..14:   (some bits of) if_index of lagg device */
1272                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1273                 /* bit 15:      0 */
1274         }
1275         return (htons(key));
1276 }
1277
1278 static void
1279 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1280 {
1281         char buf[LACP_LAGIDSTR_MAX+1];
1282
1283         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1284             __func__,
1285             lacp_format_lagid(&la->la_actor, &la->la_partner,
1286             buf, sizeof(buf)),
1287             la->la_refcnt, la->la_refcnt + 1));
1288
1289         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1290         la->la_refcnt++;
1291         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1292 }
1293
1294 static void
1295 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1296 {
1297         char buf[LACP_LAGIDSTR_MAX+1];
1298
1299         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1300             __func__,
1301             lacp_format_lagid(&la->la_actor, &la->la_partner,
1302             buf, sizeof(buf)),
1303             la->la_refcnt, la->la_refcnt - 1));
1304
1305         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1306         la->la_refcnt--;
1307         if (la->la_refcnt > 0) {
1308                 return;
1309         }
1310
1311         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1312         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1313
1314         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1315
1316         free(la, M_DEVBUF);
1317 }
1318
1319 /*
1320  * lacp_aggregator_get: allocate an aggregator.
1321  */
1322
1323 static struct lacp_aggregator *
1324 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1325 {
1326         struct lacp_aggregator *la;
1327
1328         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1329         if (la) {
1330                 la->la_refcnt = 1;
1331                 la->la_nports = 0;
1332                 TAILQ_INIT(&la->la_ports);
1333                 la->la_pending = 0;
1334                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1335         }
1336
1337         return (la);
1338 }
1339
1340 /*
1341  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1342  */
1343
1344 static void
1345 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1346 {
1347         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1348         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1349
1350         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1351 }
1352
1353 static void
1354 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1355     const struct lacp_peerinfo *lpi_port)
1356 {
1357         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1358         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1359         lpi_aggr->lip_key = lpi_port->lip_key;
1360 }
1361
1362 /*
1363  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1364  */
1365
1366 static int
1367 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1368     const struct lacp_port *lp)
1369 {
1370         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1371             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1372                 return (0);
1373         }
1374
1375         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1376                 return (0);
1377         }
1378
1379         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1380                 return (0);
1381         }
1382
1383         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1384                 return (0);
1385         }
1386
1387         return (1);
1388 }
1389
1390 static int
1391 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1392     const struct lacp_peerinfo *b)
1393 {
1394         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1395             sizeof(a->lip_systemid))) {
1396                 return (0);
1397         }
1398
1399         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1400                 return (0);
1401         }
1402
1403         return (1);
1404 }
1405
1406 static void
1407 lacp_port_enable(struct lacp_port *lp)
1408 {
1409         lp->lp_state |= LACP_STATE_AGGREGATION;
1410 }
1411
1412 static void
1413 lacp_port_disable(struct lacp_port *lp)
1414 {
1415         lacp_set_mux(lp, LACP_MUX_DETACHED);
1416
1417         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1418         lp->lp_selected = LACP_UNSELECTED;
1419         lacp_sm_rx_record_default(lp);
1420         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1421         lp->lp_state &= ~LACP_STATE_EXPIRED;
1422 }
1423
1424 /*
1425  * lacp_select: select an aggregator.  create one if necessary.
1426  */
1427 static void
1428 lacp_select(struct lacp_port *lp)
1429 {
1430         struct lacp_softc *lsc = lp->lp_lsc;
1431         struct lacp_aggregator *la;
1432         char buf[LACP_LAGIDSTR_MAX+1];
1433
1434         if (lp->lp_aggregator) {
1435                 return;
1436         }
1437
1438         /* If we haven't heard from our peer, skip this step. */
1439         if (lp->lp_state & LACP_STATE_DEFAULTED)
1440                 return;
1441
1442         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1443             ("timer_wait_while still active"));
1444
1445         LACP_DPRINTF((lp, "port lagid=%s\n",
1446             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1447             buf, sizeof(buf))));
1448
1449         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1450                 if (lacp_aggregator_is_compatible(la, lp)) {
1451                         break;
1452                 }
1453         }
1454
1455         if (la == NULL) {
1456                 la = lacp_aggregator_get(lsc, lp);
1457                 if (la == NULL) {
1458                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1459
1460                         /*
1461                          * will retry on the next tick.
1462                          */
1463
1464                         return;
1465                 }
1466                 lacp_fill_aggregator_id(la, lp);
1467                 LACP_DPRINTF((lp, "aggregator created\n"));
1468         } else {
1469                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1470                 if (la->la_refcnt == LACP_MAX_PORTS)
1471                         return;
1472                 lacp_aggregator_addref(lsc, la);
1473         }
1474
1475         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1476             lacp_format_lagid(&la->la_actor, &la->la_partner,
1477             buf, sizeof(buf))));
1478
1479         lp->lp_aggregator = la;
1480         lp->lp_selected = LACP_SELECTED;
1481 }
1482
1483 /*
1484  * lacp_unselect: finish unselect/detach process.
1485  */
1486
1487 static void
1488 lacp_unselect(struct lacp_port *lp)
1489 {
1490         struct lacp_softc *lsc = lp->lp_lsc;
1491         struct lacp_aggregator *la = lp->lp_aggregator;
1492
1493         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1494             ("timer_wait_while still active"));
1495
1496         if (la == NULL) {
1497                 return;
1498         }
1499
1500         lp->lp_aggregator = NULL;
1501         lacp_aggregator_delref(lsc, la);
1502 }
1503
1504 /* mux machine */
1505
1506 static void
1507 lacp_sm_mux(struct lacp_port *lp)
1508 {
1509         struct lagg_port *lgp = lp->lp_lagg;
1510         struct lagg_softc *sc = lgp->lp_softc;
1511         enum lacp_mux_state new_state;
1512         boolean_t p_sync =
1513                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1514         boolean_t p_collecting =
1515             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1516         enum lacp_selected selected = lp->lp_selected;
1517         struct lacp_aggregator *la;
1518
1519         if (V_lacp_debug > 1)
1520                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1521                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1522                     lp->lp_mux_state, selected, p_sync, p_collecting);
1523
1524 re_eval:
1525         la = lp->lp_aggregator;
1526         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1527             ("MUX not detached"));
1528         new_state = lp->lp_mux_state;
1529         switch (lp->lp_mux_state) {
1530         case LACP_MUX_DETACHED:
1531                 if (selected != LACP_UNSELECTED) {
1532                         new_state = LACP_MUX_WAITING;
1533                 }
1534                 break;
1535         case LACP_MUX_WAITING:
1536                 KASSERT(la->la_pending > 0 ||
1537                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1538                     ("timer_wait_while still active"));
1539                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1540                         new_state = LACP_MUX_ATTACHED;
1541                 } else if (selected == LACP_UNSELECTED) {
1542                         new_state = LACP_MUX_DETACHED;
1543                 }
1544                 break;
1545         case LACP_MUX_ATTACHED:
1546                 if (selected == LACP_SELECTED && p_sync) {
1547                         new_state = LACP_MUX_COLLECTING;
1548                 } else if (selected != LACP_SELECTED) {
1549                         new_state = LACP_MUX_DETACHED;
1550                 }
1551                 break;
1552         case LACP_MUX_COLLECTING:
1553                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1554                         new_state = LACP_MUX_DISTRIBUTING;
1555                 } else if (selected != LACP_SELECTED || !p_sync) {
1556                         new_state = LACP_MUX_ATTACHED;
1557                 }
1558                 break;
1559         case LACP_MUX_DISTRIBUTING:
1560                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1561                         new_state = LACP_MUX_COLLECTING;
1562                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1563                         sc->sc_flapping++;
1564                 }
1565                 break;
1566         default:
1567                 panic("%s: unknown state", __func__);
1568         }
1569
1570         if (lp->lp_mux_state == new_state) {
1571                 return;
1572         }
1573
1574         lacp_set_mux(lp, new_state);
1575         goto re_eval;
1576 }
1577
1578 static void
1579 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1580 {
1581         struct lacp_aggregator *la = lp->lp_aggregator;
1582
1583         if (lp->lp_mux_state == new_state) {
1584                 return;
1585         }
1586
1587         switch (new_state) {
1588         case LACP_MUX_DETACHED:
1589                 lp->lp_state &= ~LACP_STATE_SYNC;
1590                 lacp_disable_distributing(lp);
1591                 lacp_disable_collecting(lp);
1592                 lacp_sm_assert_ntt(lp);
1593                 /* cancel timer */
1594                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1595                         KASSERT(la->la_pending > 0,
1596                             ("timer_wait_while not active"));
1597                         la->la_pending--;
1598                 }
1599                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1600                 lacp_unselect(lp);
1601                 break;
1602         case LACP_MUX_WAITING:
1603                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1604                     LACP_AGGREGATE_WAIT_TIME);
1605                 la->la_pending++;
1606                 break;
1607         case LACP_MUX_ATTACHED:
1608                 lp->lp_state |= LACP_STATE_SYNC;
1609                 lacp_disable_collecting(lp);
1610                 lacp_sm_assert_ntt(lp);
1611                 break;
1612         case LACP_MUX_COLLECTING:
1613                 lacp_enable_collecting(lp);
1614                 lacp_disable_distributing(lp);
1615                 lacp_sm_assert_ntt(lp);
1616                 break;
1617         case LACP_MUX_DISTRIBUTING:
1618                 lacp_enable_distributing(lp);
1619                 break;
1620         default:
1621                 panic("%s: unknown state", __func__);
1622         }
1623
1624         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1625
1626         lp->lp_mux_state = new_state;
1627 }
1628
1629 static void
1630 lacp_sm_mux_timer(struct lacp_port *lp)
1631 {
1632         struct lacp_aggregator *la = lp->lp_aggregator;
1633         char buf[LACP_LAGIDSTR_MAX+1];
1634
1635         KASSERT(la->la_pending > 0, ("no pending event"));
1636
1637         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1638             lacp_format_lagid(&la->la_actor, &la->la_partner,
1639             buf, sizeof(buf)),
1640             la->la_pending, la->la_pending - 1));
1641
1642         la->la_pending--;
1643 }
1644
1645 /* periodic transmit machine */
1646
1647 static void
1648 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1649 {
1650         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1651             LACP_STATE_TIMEOUT)) {
1652                 return;
1653         }
1654
1655         LACP_DPRINTF((lp, "partner timeout changed\n"));
1656
1657         /*
1658          * FAST_PERIODIC -> SLOW_PERIODIC
1659          * or
1660          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1661          *
1662          * let lacp_sm_ptx_tx_schedule to update timeout.
1663          */
1664
1665         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1666
1667         /*
1668          * if timeout has been shortened, assert NTT.
1669          */
1670
1671         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1672                 lacp_sm_assert_ntt(lp);
1673         }
1674 }
1675
1676 static void
1677 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1678 {
1679         int timeout;
1680
1681         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1682             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1683                 /*
1684                  * NO_PERIODIC
1685                  */
1686
1687                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1688                 return;
1689         }
1690
1691         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1692                 return;
1693         }
1694
1695         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1696             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1697
1698         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1699 }
1700
1701 static void
1702 lacp_sm_ptx_timer(struct lacp_port *lp)
1703 {
1704         lacp_sm_assert_ntt(lp);
1705 }
1706
1707 static void
1708 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1709 {
1710         int timeout;
1711
1712         /*
1713          * check LACP_DISABLED first
1714          */
1715
1716         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1717                 return;
1718         }
1719
1720         /*
1721          * check loopback condition.
1722          */
1723
1724         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1725             &lp->lp_actor.lip_systemid)) {
1726                 return;
1727         }
1728
1729         /*
1730          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1731          */
1732
1733         lacp_sm_rx_update_selected(lp, du);
1734         lacp_sm_rx_update_ntt(lp, du);
1735         lacp_sm_rx_record_pdu(lp, du);
1736
1737         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1738             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1739         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1740
1741         lp->lp_state &= ~LACP_STATE_EXPIRED;
1742
1743         /*
1744          * kick transmit machine without waiting the next tick.
1745          */
1746
1747         lacp_sm_tx(lp);
1748 }
1749
1750 static void
1751 lacp_sm_rx_set_expired(struct lacp_port *lp)
1752 {
1753         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1754         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1755         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1756         lp->lp_state |= LACP_STATE_EXPIRED;
1757 }
1758
1759 static void
1760 lacp_sm_rx_timer(struct lacp_port *lp)
1761 {
1762         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1763                 /* CURRENT -> EXPIRED */
1764                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1765                 lacp_sm_rx_set_expired(lp);
1766         } else {
1767                 /* EXPIRED -> DEFAULTED */
1768                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1769                 lacp_sm_rx_update_default_selected(lp);
1770                 lacp_sm_rx_record_default(lp);
1771                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1772         }
1773 }
1774
1775 static void
1776 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1777 {
1778         boolean_t active;
1779         uint8_t oldpstate;
1780         char buf[LACP_STATESTR_MAX+1];
1781
1782         LACP_TRACE(lp);
1783
1784         oldpstate = lp->lp_partner.lip_state;
1785
1786         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1787             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1788             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1789
1790         lp->lp_partner = du->ldu_actor;
1791         if (active &&
1792             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1793             LACP_STATE_AGGREGATION) &&
1794             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1795             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1796                 /*
1797                  * XXX Maintain legacy behavior of leaving the
1798                  * LACP_STATE_SYNC bit unchanged from the partner's
1799                  * advertisement if lsc_strict_mode is false.
1800                  * TODO: We should re-examine the concept of the "strict mode"
1801                  * to ensure it makes sense to maintain a non-strict mode.
1802                  */
1803                 if (lp->lp_lsc->lsc_strict_mode)
1804                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1805         } else {
1806                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1807         }
1808
1809         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1810
1811         if (oldpstate != lp->lp_partner.lip_state) {
1812                 LACP_DPRINTF((lp, "old pstate %s\n",
1813                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1814                 LACP_DPRINTF((lp, "new pstate %s\n",
1815                     lacp_format_state(lp->lp_partner.lip_state, buf,
1816                     sizeof(buf))));
1817         }
1818
1819         lacp_sm_ptx_update_timeout(lp, oldpstate);
1820 }
1821
1822 static void
1823 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1824 {
1825
1826         LACP_TRACE(lp);
1827
1828         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1829             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1830             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1831                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1832                 lacp_sm_assert_ntt(lp);
1833         }
1834 }
1835
1836 static void
1837 lacp_sm_rx_record_default(struct lacp_port *lp)
1838 {
1839         uint8_t oldpstate;
1840
1841         LACP_TRACE(lp);
1842
1843         oldpstate = lp->lp_partner.lip_state;
1844         if (lp->lp_lsc->lsc_strict_mode)
1845                 lp->lp_partner = lacp_partner_admin_strict;
1846         else
1847                 lp->lp_partner = lacp_partner_admin_optimistic;
1848         lp->lp_state |= LACP_STATE_DEFAULTED;
1849         lacp_sm_ptx_update_timeout(lp, oldpstate);
1850 }
1851
1852 static void
1853 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1854     const struct lacp_peerinfo *info)
1855 {
1856
1857         LACP_TRACE(lp);
1858
1859         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1860             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1861             LACP_STATE_AGGREGATION)) {
1862                 lp->lp_selected = LACP_UNSELECTED;
1863                 /* mux machine will clean up lp->lp_aggregator */
1864         }
1865 }
1866
1867 static void
1868 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1869 {
1870
1871         LACP_TRACE(lp);
1872
1873         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1874 }
1875
1876 static void
1877 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1878 {
1879
1880         LACP_TRACE(lp);
1881
1882         if (lp->lp_lsc->lsc_strict_mode)
1883                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1884                     &lacp_partner_admin_strict);
1885         else
1886                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1887                     &lacp_partner_admin_optimistic);
1888 }
1889
1890 /* transmit machine */
1891
1892 static void
1893 lacp_sm_tx(struct lacp_port *lp)
1894 {
1895         int error = 0;
1896
1897         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1898 #if 1
1899             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1900             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1901 #endif
1902             ) {
1903                 lp->lp_flags &= ~LACP_PORT_NTT;
1904         }
1905
1906         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1907                 return;
1908         }
1909
1910         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1911         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1912                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1913                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1914                 return;
1915         }
1916
1917         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1918                 error = lacp_xmit_lacpdu(lp);
1919         } else {
1920                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1921         }
1922
1923         if (error == 0) {
1924                 lp->lp_flags &= ~LACP_PORT_NTT;
1925         } else {
1926                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1927                     error));
1928         }
1929 }
1930
1931 static void
1932 lacp_sm_assert_ntt(struct lacp_port *lp)
1933 {
1934
1935         lp->lp_flags |= LACP_PORT_NTT;
1936 }
1937
1938 static void
1939 lacp_run_timers(struct lacp_port *lp)
1940 {
1941         int i;
1942
1943         for (i = 0; i < LACP_NTIMER; i++) {
1944                 KASSERT(lp->lp_timer[i] >= 0,
1945                     ("invalid timer value %d", lp->lp_timer[i]));
1946                 if (lp->lp_timer[i] == 0) {
1947                         continue;
1948                 } else if (--lp->lp_timer[i] <= 0) {
1949                         if (lacp_timer_funcs[i]) {
1950                                 (*lacp_timer_funcs[i])(lp);
1951                         }
1952                 }
1953         }
1954 }
1955
1956 int
1957 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1958 {
1959         struct lacp_softc *lsc = lp->lp_lsc;
1960         struct lagg_port *lgp = lp->lp_lagg;
1961         struct lacp_port *lp2;
1962         struct markerdu *mdu;
1963         int error = 0;
1964         int pending = 0;
1965
1966         if (m->m_pkthdr.len != sizeof(*mdu)) {
1967                 goto bad;
1968         }
1969
1970         if ((m->m_flags & M_MCAST) == 0) {
1971                 goto bad;
1972         }
1973
1974         if (m->m_len < sizeof(*mdu)) {
1975                 m = m_pullup(m, sizeof(*mdu));
1976                 if (m == NULL) {
1977                         return (ENOMEM);
1978                 }
1979         }
1980
1981         mdu = mtod(m, struct markerdu *);
1982
1983         if (memcmp(&mdu->mdu_eh.ether_dhost,
1984             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1985                 goto bad;
1986         }
1987
1988         if (mdu->mdu_sph.sph_version != 1) {
1989                 goto bad;
1990         }
1991
1992         switch (mdu->mdu_tlv.tlv_type) {
1993         case MARKER_TYPE_INFO:
1994                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
1995                     marker_info_tlv_template, TRUE)) {
1996                         goto bad;
1997                 }
1998                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
1999                 memcpy(&mdu->mdu_eh.ether_dhost,
2000                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
2001                 memcpy(&mdu->mdu_eh.ether_shost,
2002                     lgp->lp_lladdr, ETHER_ADDR_LEN);
2003                 error = lagg_enqueue(lp->lp_ifp, m);
2004                 break;
2005
2006         case MARKER_TYPE_RESPONSE:
2007                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2008                     marker_response_tlv_template, TRUE)) {
2009                         goto bad;
2010                 }
2011                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
2012                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
2013                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
2014
2015                 /* Verify that it is the last marker we sent out */
2016                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
2017                     sizeof(struct lacp_markerinfo)))
2018                         goto bad;
2019
2020                 LACP_LOCK(lsc);
2021                 lp->lp_flags &= ~LACP_PORT_MARK;
2022
2023                 if (lsc->lsc_suppress_distributing) {
2024                         /* Check if any ports are waiting for a response */
2025                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
2026                                 if (lp2->lp_flags & LACP_PORT_MARK) {
2027                                         pending = 1;
2028                                         break;
2029                                 }
2030                         }
2031
2032                         if (pending == 0) {
2033                                 /* All interface queues are clear */
2034                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
2035                                 lsc->lsc_suppress_distributing = FALSE;
2036                         }
2037                 }
2038                 LACP_UNLOCK(lsc);
2039                 m_freem(m);
2040                 break;
2041
2042         default:
2043                 goto bad;
2044         }
2045
2046         return (error);
2047
2048 bad:
2049         LACP_DPRINTF((lp, "bad marker frame\n"));
2050         m_freem(m);
2051         return (EINVAL);
2052 }
2053
2054 static int
2055 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
2056     const struct tlv_template *tmpl, boolean_t check_type)
2057 {
2058         while (/* CONSTCOND */ 1) {
2059                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
2060                         return (EINVAL);
2061                 }
2062                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
2063                     tlv->tlv_length != tmpl->tmpl_length) {
2064                         return (EINVAL);
2065                 }
2066                 if (tmpl->tmpl_type == 0) {
2067                         break;
2068                 }
2069                 tlv = (const struct tlvhdr *)
2070                     ((const char *)tlv + tlv->tlv_length);
2071                 tmpl++;
2072         }
2073
2074         return (0);
2075 }
2076
2077 /* Debugging */
2078 const char *
2079 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
2080 {
2081         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
2082             (int)mac[0],
2083             (int)mac[1],
2084             (int)mac[2],
2085             (int)mac[3],
2086             (int)mac[4],
2087             (int)mac[5]);
2088
2089         return (buf);
2090 }
2091
2092 const char *
2093 lacp_format_systemid(const struct lacp_systemid *sysid,
2094     char *buf, size_t buflen)
2095 {
2096         char macbuf[LACP_MACSTR_MAX+1];
2097
2098         snprintf(buf, buflen, "%04X,%s",
2099             ntohs(sysid->lsi_prio),
2100             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2101
2102         return (buf);
2103 }
2104
2105 const char *
2106 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2107 {
2108         snprintf(buf, buflen, "%04X,%04X",
2109             ntohs(portid->lpi_prio),
2110             ntohs(portid->lpi_portno));
2111
2112         return (buf);
2113 }
2114
2115 const char *
2116 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2117 {
2118         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2119         char portid[LACP_PORTIDSTR_MAX+1];
2120
2121         snprintf(buf, buflen, "(%s,%04X,%s)",
2122             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2123             ntohs(peer->lip_key),
2124             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2125
2126         return (buf);
2127 }
2128
2129 const char *
2130 lacp_format_lagid(const struct lacp_peerinfo *a,
2131     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2132 {
2133         char astr[LACP_PARTNERSTR_MAX+1];
2134         char bstr[LACP_PARTNERSTR_MAX+1];
2135
2136 #if 0
2137         /*
2138          * there's a convention to display small numbered peer
2139          * in the left.
2140          */
2141
2142         if (lacp_compare_peerinfo(a, b) > 0) {
2143                 const struct lacp_peerinfo *t;
2144
2145                 t = a;
2146                 a = b;
2147                 b = t;
2148         }
2149 #endif
2150
2151         snprintf(buf, buflen, "[%s,%s]",
2152             lacp_format_partner(a, astr, sizeof(astr)),
2153             lacp_format_partner(b, bstr, sizeof(bstr)));
2154
2155         return (buf);
2156 }
2157
2158 const char *
2159 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2160     char *buf, size_t buflen)
2161 {
2162         if (la == NULL) {
2163                 return ("(none)");
2164         }
2165
2166         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2167 }
2168
2169 const char *
2170 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2171 {
2172         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2173         return (buf);
2174 }
2175
2176 static void
2177 lacp_dump_lacpdu(const struct lacpdu *du)
2178 {
2179         char buf[LACP_PARTNERSTR_MAX+1];
2180         char buf2[LACP_STATESTR_MAX+1];
2181
2182         printf("actor=%s\n",
2183             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2184         printf("actor.state=%s\n",
2185             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2186         printf("partner=%s\n",
2187             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2188         printf("partner.state=%s\n",
2189             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2190
2191         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2192 }
2193
2194 static void
2195 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2196 {
2197         va_list va;
2198
2199         if (lp) {
2200                 printf("%s: ", lp->lp_ifp->if_xname);
2201         }
2202
2203         va_start(va, fmt);
2204         vprintf(fmt, va);
2205         va_end(va);
2206 }