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/if_media.h>
  58 #include <net/if_types.h>
  59
  60 #include <net/if_lagg.h>
  61 #include <net/ieee8023ad_lacp.h>
  62
  63 /*
  64  * actor system priority and port priority.
  65  * XXX should be configurable.
  66  */
  67
  68 #define LACP_SYSTEM_PRIO        0x8000
  69 #define LACP_PORT_PRIO          0x8000
  70
  71 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  72     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  73
  74 static const struct tlv_template lacp_info_tlv_template[] = {
  75         { LACP_TYPE_ACTORINFO,
  76             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  77         { LACP_TYPE_PARTNERINFO,
  78             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  79         { LACP_TYPE_COLLECTORINFO,
  80             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  81         { 0, 0 },
  82 };
  83
  84 static const struct tlv_template marker_info_tlv_template[] = {
  85         { MARKER_TYPE_INFO,
  86             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  87         { 0, 0 },
  88 };
  89
  90 static const struct tlv_template marker_response_tlv_template[] = {
  91         { MARKER_TYPE_RESPONSE,
  92             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  93         { 0, 0 },
  94 };
  95
  96 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  97
  98 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
  99 static void     lacp_fill_markerinfo(struct lacp_port *,
 100                     struct lacp_markerinfo *);
 101
 102 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
 103 static void     lacp_suppress_distributing(struct lacp_softc *,
 104                     struct lacp_aggregator *);
 105 static void     lacp_transit_expire(void *);
 106 static void     lacp_update_portmap(struct lacp_softc *);
 107 static void     lacp_select_active_aggregator(struct lacp_softc *);
 108 static uint16_t lacp_compose_key(struct lacp_port *);
 109 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 110                     const struct tlv_template *, boolean_t);
 111 static void     lacp_tick(void *);
 112
 113 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 114                     const struct lacp_port *);
 115 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 116                     const struct lacp_peerinfo *);
 117 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 118                     const struct lacp_port *);
 119 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 120                     const struct lacp_peerinfo *);
 121
 122 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 123                     struct lacp_port *);
 124 static void     lacp_aggregator_addref(struct lacp_softc *,
 125                     struct lacp_aggregator *);
 126 static void     lacp_aggregator_delref(struct lacp_softc *,
 127                     struct lacp_aggregator *);
 128
 129 /* receive machine */
 130
 131 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 132 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 133 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 134 static void     lacp_sm_rx_timer(struct lacp_port *);
 135 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 136 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 137                     const struct lacpdu *);
 138 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 139                     const struct lacpdu *);
 140 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 141                     const struct lacpdu *);
 142 static void     lacp_sm_rx_record_default(struct lacp_port *);
 143 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 144 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 145                     const struct lacp_peerinfo *);
 146
 147 /* mux machine */
 148
 149 static void     lacp_sm_mux(struct lacp_port *);
 150 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 151 static void     lacp_sm_mux_timer(struct lacp_port *);
 152
 153 /* periodic transmit machine */
 154
 155 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 156 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 157 static void     lacp_sm_ptx_timer(struct lacp_port *);
 158
 159 /* transmit machine */
 160
 161 static void     lacp_sm_tx(struct lacp_port *);
 162 static void     lacp_sm_assert_ntt(struct lacp_port *);
 163
 164 static void     lacp_run_timers(struct lacp_port *);
 165 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 166                     const struct lacp_peerinfo *);
 167 static int      lacp_compare_systemid(const struct lacp_systemid *,
 168                     const struct lacp_systemid *);
 169 static void     lacp_port_enable(struct lacp_port *);
 170 static void     lacp_port_disable(struct lacp_port *);
 171 static void     lacp_select(struct lacp_port *);
 172 static void     lacp_unselect(struct lacp_port *);
 173 static void     lacp_disable_collecting(struct lacp_port *);
 174 static void     lacp_enable_collecting(struct lacp_port *);
 175 static void     lacp_disable_distributing(struct lacp_port *);
 176 static void     lacp_enable_distributing(struct lacp_port *);
 177 static int      lacp_xmit_lacpdu(struct lacp_port *);
 178 static int      lacp_xmit_marker(struct lacp_port *);
 179
 180 /* Debugging */
 181
 182 static void     lacp_dump_lacpdu(const struct lacpdu *);
 183 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 184                     size_t);
 185 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 186                     const struct lacp_peerinfo *, char *, size_t);
 187 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 188                     char *, size_t);
 189 static const char *lacp_format_state(uint8_t, char *, size_t);
 190 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 191 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 192                     size_t);
 193 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 194                     size_t);
 195 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 196                     __attribute__((__format__(__printf__, 2, 3)));
 197
 198 VNET_DEFINE_STATIC(int, lacp_debug);
 199 #define V_lacp_debug    VNET(lacp_debug)
 200 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD, 0, "ieee802.3ad");
 201 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 202     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 203
 204 VNET_DEFINE_STATIC(int, lacp_default_strict_mode) = 1;
 205 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode,
 206     CTLFLAG_RWTUN | CTLFLAG_VNET, &VNET_NAME(lacp_default_strict_mode), 0,
 207     "LACP strict protocol compliance default");
 208
 209 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 210 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 211 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 212
 213 /*
 214  * partner administration variables.
 215  * XXX should be configurable.
 216  */
 217
 218 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 219         .lip_systemid = { .lsi_prio = 0xffff },
 220         .lip_portid = { .lpi_prio = 0xffff },
 221         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 222             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 223 };
 224
 225 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 226         .lip_systemid = { .lsi_prio = 0xffff },
 227         .lip_portid = { .lpi_prio = 0xffff },
 228         .lip_state = 0,
 229 };
 230
 231 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 232         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 233         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 234         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 235 };
 236
 237 struct mbuf *
 238 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 239 {
 240         struct lacp_port *lp = LACP_PORT(lgp);
 241         uint8_t subtype;
 242
 243         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 244                 m_freem(m);
 245                 return (NULL);
 246         }
 247
 248         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 249         switch (subtype) {
 250                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 251                         lacp_pdu_input(lp, m);
 252                         return (NULL);
 253
 254                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 255                         lacp_marker_input(lp, m);
 256                         return (NULL);
 257         }
 258
 259         /* Not a subtype we are interested in */
 260         return (m);
 261 }
 262
 263 /*
 264  * lacp_pdu_input: process lacpdu
 265  */
 266 static int
 267 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 268 {
 269         struct lacp_softc *lsc = lp->lp_lsc;
 270         struct lacpdu *du;
 271         int error = 0;
 272
 273         if (m->m_pkthdr.len != sizeof(*du)) {
 274                 goto bad;
 275         }
 276
 277         if ((m->m_flags & M_MCAST) == 0) {
 278                 goto bad;
 279         }
 280
 281         if (m->m_len < sizeof(*du)) {
 282                 m = m_pullup(m, sizeof(*du));
 283                 if (m == NULL) {
 284                         return (ENOMEM);
 285                 }
 286         }
 287
 288         du = mtod(m, struct lacpdu *);
 289
 290         if (memcmp(&du->ldu_eh.ether_dhost,
 291             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 292                 goto bad;
 293         }
 294
 295         /*
 296          * ignore the version for compatibility with
 297          * the future protocol revisions.
 298          */
 299 #if 0
 300         if (du->ldu_sph.sph_version != 1) {
 301                 goto bad;
 302         }
 303 #endif
 304
 305         /*
 306          * ignore tlv types for compatibility with
 307          * the future protocol revisions.
 308          */
 309         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 310             lacp_info_tlv_template, FALSE)) {
 311                 goto bad;
 312         }
 313
 314         if (V_lacp_debug > 0) {
 315                 lacp_dprintf(lp, "lacpdu receive\n");
 316                 lacp_dump_lacpdu(du);
 317         }
 318
 319         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 320                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 321                 goto bad;
 322         }
 323
 324         LACP_LOCK(lsc);
 325         lacp_sm_rx(lp, du);
 326         LACP_UNLOCK(lsc);
 327
 328         m_freem(m);
 329         return (error);
 330
 331 bad:
 332         m_freem(m);
 333         return (EINVAL);
 334 }
 335
 336 static void
 337 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 338 {
 339         struct lagg_port *lgp = lp->lp_lagg;
 340         struct lagg_softc *sc = lgp->lp_softc;
 341
 342         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 343         memcpy(&info->lip_systemid.lsi_mac,
 344             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 345         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 346         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 347         info->lip_state = lp->lp_state;
 348 }
 349
 350 static void
 351 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 352 {
 353         struct ifnet *ifp = lp->lp_ifp;
 354
 355         /* Fill in the port index and system id (encoded as the MAC) */
 356         info->mi_rq_port = htons(ifp->if_index);
 357         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 358         info->mi_rq_xid = htonl(0);
 359 }
 360
 361 static int
 362 lacp_xmit_lacpdu(struct lacp_port *lp)
 363 {
 364         struct lagg_port *lgp = lp->lp_lagg;
 365         struct mbuf *m;
 366         struct lacpdu *du;
 367         int error;
 368
 369         LACP_LOCK_ASSERT(lp->lp_lsc);
 370
 371         m = m_gethdr(M_NOWAIT, MT_DATA);
 372         if (m == NULL) {
 373                 return (ENOMEM);
 374         }
 375         m->m_len = m->m_pkthdr.len = sizeof(*du);
 376
 377         du = mtod(m, struct lacpdu *);
 378         memset(du, 0, sizeof(*du));
 379
 380         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 381             ETHER_ADDR_LEN);
 382         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 383         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 384
 385         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 386         du->ldu_sph.sph_version = 1;
 387
 388         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 389         du->ldu_actor = lp->lp_actor;
 390
 391         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 392             sizeof(du->ldu_partner));
 393         du->ldu_partner = lp->lp_partner;
 394
 395         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 396             sizeof(du->ldu_collector));
 397         du->ldu_collector.lci_maxdelay = 0;
 398
 399         if (V_lacp_debug > 0) {
 400                 lacp_dprintf(lp, "lacpdu transmit\n");
 401                 lacp_dump_lacpdu(du);
 402         }
 403
 404         m->m_flags |= M_MCAST;
 405
 406         /*
 407          * XXX should use higher priority queue.
 408          * otherwise network congestion can break aggregation.
 409          */
 410
 411         error = lagg_enqueue(lp->lp_ifp, m);
 412         return (error);
 413 }
 414
 415 static int
 416 lacp_xmit_marker(struct lacp_port *lp)
 417 {
 418         struct lagg_port *lgp = lp->lp_lagg;
 419         struct mbuf *m;
 420         struct markerdu *mdu;
 421         int error;
 422
 423         LACP_LOCK_ASSERT(lp->lp_lsc);
 424
 425         m = m_gethdr(M_NOWAIT, MT_DATA);
 426         if (m == NULL) {
 427                 return (ENOMEM);
 428         }
 429         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 430
 431         mdu = mtod(m, struct markerdu *);
 432         memset(mdu, 0, sizeof(*mdu));
 433
 434         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 435             ETHER_ADDR_LEN);
 436         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 437         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 438
 439         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 440         mdu->mdu_sph.sph_version = 1;
 441
 442         /* Bump the transaction id and copy over the marker info */
 443         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 444         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 445         mdu->mdu_info = lp->lp_marker;
 446
 447         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 448             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 449             ntohl(mdu->mdu_info.mi_rq_xid)));
 450
 451         m->m_flags |= M_MCAST;
 452         error = lagg_enqueue(lp->lp_ifp, m);
 453         return (error);
 454 }
 455
 456 void
 457 lacp_linkstate(struct lagg_port *lgp)
 458 {
 459         struct lacp_port *lp = LACP_PORT(lgp);
 460         struct lacp_softc *lsc = lp->lp_lsc;
 461         struct ifnet *ifp = lgp->lp_ifp;
 462         struct ifmediareq ifmr;
 463         int error = 0;
 464         u_int media;
 465         uint8_t old_state;
 466         uint16_t old_key;
 467
 468         bzero((char *)&ifmr, sizeof(ifmr));
 469         error = (*ifp->if_ioctl)(ifp, SIOCGIFXMEDIA, (caddr_t)&ifmr);
 470         if (error != 0) {
 471                 bzero((char *)&ifmr, sizeof(ifmr));
 472                 error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 473         }
 474         if (error != 0)
 475                 return;
 476
 477         LACP_LOCK(lsc);
 478         media = ifmr.ifm_active;
 479         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 480             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 481             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 482         old_state = lp->lp_state;
 483         old_key = lp->lp_key;
 484
 485         lp->lp_media = media;
 486         /*
 487          * If the port is not an active full duplex Ethernet link then it can
 488          * not be aggregated.
 489          */
 490         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 491             ifp->if_link_state != LINK_STATE_UP) {
 492                 lacp_port_disable(lp);
 493         } else {
 494                 lacp_port_enable(lp);
 495         }
 496         lp->lp_key = lacp_compose_key(lp);
 497
 498         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 499                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 500                 lp->lp_selected = LACP_UNSELECTED;
 501         }
 502         LACP_UNLOCK(lsc);
 503 }
 504
 505 static void
 506 lacp_tick(void *arg)
 507 {
 508         struct lacp_softc *lsc = arg;
 509         struct lacp_port *lp;
 510
 511         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 512                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 513                         continue;
 514
 515                 CURVNET_SET(lp->lp_ifp->if_vnet);
 516                 lacp_run_timers(lp);
 517
 518                 lacp_select(lp);
 519                 lacp_sm_mux(lp);
 520                 lacp_sm_tx(lp);
 521                 lacp_sm_ptx_tx_schedule(lp);
 522                 CURVNET_RESTORE();
 523         }
 524         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 525 }
 526
 527 int
 528 lacp_port_create(struct lagg_port *lgp)
 529 {
 530         struct lagg_softc *sc = lgp->lp_softc;
 531         struct lacp_softc *lsc = LACP_SOFTC(sc);
 532         struct lacp_port *lp;
 533         struct ifnet *ifp = lgp->lp_ifp;
 534         struct sockaddr_dl sdl;
 535         struct ifmultiaddr *rifma = NULL;
 536         int error;
 537
 538         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 539         sdl.sdl_alen = ETHER_ADDR_LEN;
 540
 541         bcopy(&ethermulticastaddr_slowprotocols,
 542             LLADDR(&sdl), ETHER_ADDR_LEN);
 543         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 544         if (error) {
 545                 printf("%s: ADDMULTI failed on %s\n", __func__,
 546                     lgp->lp_ifp->if_xname);
 547                 return (error);
 548         }
 549
 550         lp = malloc(sizeof(struct lacp_port),
 551             M_DEVBUF, M_NOWAIT|M_ZERO);
 552         if (lp == NULL)
 553                 return (ENOMEM);
 554
 555         LACP_LOCK(lsc);
 556         lgp->lp_psc = lp;
 557         lp->lp_ifp = ifp;
 558         lp->lp_lagg = lgp;
 559         lp->lp_lsc = lsc;
 560         lp->lp_ifma = rifma;
 561
 562         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 563
 564         lacp_fill_actorinfo(lp, &lp->lp_actor);
 565         lacp_fill_markerinfo(lp, &lp->lp_marker);
 566         lp->lp_state = LACP_STATE_ACTIVITY;
 567         lp->lp_aggregator = NULL;
 568         lacp_sm_rx_set_expired(lp);
 569         LACP_UNLOCK(lsc);
 570         lacp_linkstate(lgp);
 571
 572         return (0);
 573 }
 574
 575 void
 576 lacp_port_destroy(struct lagg_port *lgp)
 577 {
 578         struct lacp_port *lp = LACP_PORT(lgp);
 579         struct lacp_softc *lsc = lp->lp_lsc;
 580         int i;
 581
 582         LACP_LOCK(lsc);
 583         for (i = 0; i < LACP_NTIMER; i++) {
 584                 LACP_TIMER_DISARM(lp, i);
 585         }
 586
 587         lacp_disable_collecting(lp);
 588         lacp_disable_distributing(lp);
 589         lacp_unselect(lp);
 590
 591         LIST_REMOVE(lp, lp_next);
 592         LACP_UNLOCK(lsc);
 593
 594         /* The address may have already been removed by if_purgemaddrs() */
 595         if (!lgp->lp_detaching)
 596                 if_delmulti_ifma(lp->lp_ifma);
 597
 598         free(lp, M_DEVBUF);
 599 }
 600
 601 void
 602 lacp_req(struct lagg_softc *sc, void *data)
 603 {
 604         struct lacp_opreq *req = (struct lacp_opreq *)data;
 605         struct lacp_softc *lsc = LACP_SOFTC(sc);
 606         struct lacp_aggregator *la;
 607
 608         bzero(req, sizeof(struct lacp_opreq));
 609
 610         /*
 611          * If the LACP softc is NULL, return with the opreq structure full of
 612          * zeros.  It is normal for the softc to be NULL while the lagg is
 613          * being destroyed.
 614          */
 615         if (NULL == lsc)
 616                 return;
 617
 618         la = lsc->lsc_active_aggregator;
 619         LACP_LOCK(lsc);
 620         if (la != NULL) {
 621                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 622                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 623                     ETHER_ADDR_LEN);
 624                 req->actor_key = ntohs(la->la_actor.lip_key);
 625                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 626                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 627                 req->actor_state = la->la_actor.lip_state;
 628
 629                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 630                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 631                     ETHER_ADDR_LEN);
 632                 req->partner_key = ntohs(la->la_partner.lip_key);
 633                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 634                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 635                 req->partner_state = la->la_partner.lip_state;
 636         }
 637         LACP_UNLOCK(lsc);
 638 }
 639
 640 void
 641 lacp_portreq(struct lagg_port *lgp, void *data)
 642 {
 643         struct lacp_opreq *req = (struct lacp_opreq *)data;
 644         struct lacp_port *lp = LACP_PORT(lgp);
 645         struct lacp_softc *lsc = lp->lp_lsc;
 646
 647         LACP_LOCK(lsc);
 648         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 649         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 650             ETHER_ADDR_LEN);
 651         req->actor_key = ntohs(lp->lp_actor.lip_key);
 652         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 653         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 654         req->actor_state = lp->lp_actor.lip_state;
 655
 656         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 657         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 658             ETHER_ADDR_LEN);
 659         req->partner_key = ntohs(lp->lp_partner.lip_key);
 660         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 661         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 662         req->partner_state = lp->lp_partner.lip_state;
 663         LACP_UNLOCK(lsc);
 664 }
 665
 666 static void
 667 lacp_disable_collecting(struct lacp_port *lp)
 668 {
 669         LACP_DPRINTF((lp, "collecting disabled\n"));
 670         lp->lp_state &= ~LACP_STATE_COLLECTING;
 671 }
 672
 673 static void
 674 lacp_enable_collecting(struct lacp_port *lp)
 675 {
 676         LACP_DPRINTF((lp, "collecting enabled\n"));
 677         lp->lp_state |= LACP_STATE_COLLECTING;
 678 }
 679
 680 static void
 681 lacp_disable_distributing(struct lacp_port *lp)
 682 {
 683         struct lacp_aggregator *la = lp->lp_aggregator;
 684         struct lacp_softc *lsc = lp->lp_lsc;
 685         struct lagg_softc *sc = lsc->lsc_softc;
 686         char buf[LACP_LAGIDSTR_MAX+1];
 687
 688         LACP_LOCK_ASSERT(lsc);
 689
 690         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 691                 return;
 692         }
 693
 694         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 695         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 696         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 697
 698         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 699             "nports %d -> %d\n",
 700             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 701             la->la_nports, la->la_nports - 1));
 702
 703         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 704         la->la_nports--;
 705         sc->sc_active = la->la_nports;
 706
 707         if (lsc->lsc_active_aggregator == la) {
 708                 lacp_suppress_distributing(lsc, la);
 709                 lacp_select_active_aggregator(lsc);
 710                 /* regenerate the port map, the active aggregator has changed */
 711                 lacp_update_portmap(lsc);
 712         }
 713
 714         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 715         if_link_state_change(sc->sc_ifp,
 716             sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
 717 }
 718
 719 static void
 720 lacp_enable_distributing(struct lacp_port *lp)
 721 {
 722         struct lacp_aggregator *la = lp->lp_aggregator;
 723         struct lacp_softc *lsc = lp->lp_lsc;
 724         struct lagg_softc *sc = lsc->lsc_softc;
 725         char buf[LACP_LAGIDSTR_MAX+1];
 726
 727         LACP_LOCK_ASSERT(lsc);
 728
 729         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 730                 return;
 731         }
 732
 733         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 734             "nports %d -> %d\n",
 735             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 736             la->la_nports, la->la_nports + 1));
 737
 738         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 739         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 740         la->la_nports++;
 741         sc->sc_active = la->la_nports;
 742
 743         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 744
 745         if (lsc->lsc_active_aggregator == la) {
 746                 lacp_suppress_distributing(lsc, la);
 747                 lacp_update_portmap(lsc);
 748         } else
 749                 /* try to become the active aggregator */
 750                 lacp_select_active_aggregator(lsc);
 751
 752         if_link_state_change(sc->sc_ifp,
 753             sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
 754 }
 755
 756 static void
 757 lacp_transit_expire(void *vp)
 758 {
 759         struct lacp_softc *lsc = vp;
 760
 761         LACP_LOCK_ASSERT(lsc);
 762
 763         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 764         LACP_TRACE(NULL);
 765         CURVNET_RESTORE();
 766
 767         lsc->lsc_suppress_distributing = FALSE;
 768 }
 769
 770 void
 771 lacp_attach(struct lagg_softc *sc)
 772 {
 773         struct lacp_softc *lsc;
 774
 775         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 776
 777         sc->sc_psc = lsc;
 778         lsc->lsc_softc = sc;
 779
 780         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 781         lsc->lsc_active_aggregator = NULL;
 782         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 783         LACP_LOCK_INIT(lsc);
 784         TAILQ_INIT(&lsc->lsc_aggregators);
 785         LIST_INIT(&lsc->lsc_ports);
 786
 787         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 788         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 789
 790         /* if the lagg is already up then do the same */
 791         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 792                 lacp_init(sc);
 793 }
 794
 795 void
 796 lacp_detach(void *psc)
 797 {
 798         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 799
 800         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 801             ("aggregators still active"));
 802         KASSERT(lsc->lsc_active_aggregator == NULL,
 803             ("aggregator still attached"));
 804
 805         callout_drain(&lsc->lsc_transit_callout);
 806         callout_drain(&lsc->lsc_callout);
 807
 808         LACP_LOCK_DESTROY(lsc);
 809         free(lsc, M_DEVBUF);
 810 }
 811
 812 void
 813 lacp_init(struct lagg_softc *sc)
 814 {
 815         struct lacp_softc *lsc = LACP_SOFTC(sc);
 816
 817         LACP_LOCK(lsc);
 818         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 819         LACP_UNLOCK(lsc);
 820 }
 821
 822 void
 823 lacp_stop(struct lagg_softc *sc)
 824 {
 825         struct lacp_softc *lsc = LACP_SOFTC(sc);
 826
 827         LACP_LOCK(lsc);
 828         callout_stop(&lsc->lsc_transit_callout);
 829         callout_stop(&lsc->lsc_callout);
 830         LACP_UNLOCK(lsc);
 831 }
 832
 833 struct lagg_port *
 834 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m)
 835 {
 836         struct lacp_softc *lsc = LACP_SOFTC(sc);
 837         struct lacp_portmap *pm;
 838         struct lacp_port *lp;
 839         struct lacp_port **map;
 840         uint32_t hash;
 841         int count;
 842
 843         if (__predict_false(lsc->lsc_suppress_distributing)) {
 844                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 845                 return (NULL);
 846         }
 847
 848         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 849         if (pm->pm_count == 0) {
 850                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 851                 return (NULL);
 852         }
 853
 854 #ifdef NUMA
 855         if ((sc->sc_opts & LAGG_OPT_USE_NUMA) &&
 856             pm->pm_num_dom > 1 && m->m_pkthdr.numa_domain < MAXMEMDOM) {
 857                 count = pm->pm_numa[m->m_pkthdr.numa_domain].count;
 858                 if (count > 0) {
 859                         map = pm->pm_numa[m->m_pkthdr.numa_domain].map;
 860                 } else {
 861                         /* No ports on this domain; use global hash. */
 862                         map = pm->pm_map;
 863                         count = pm->pm_count;
 864                 }
 865         } else
 866 #endif
 867         {
 868                 map = pm->pm_map;
 869                 count = pm->pm_count;
 870         }
 871         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 872             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 873                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 874         else
 875                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 876
 877         hash %= count;
 878         lp = map[hash];
 879
 880         KASSERT((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0,
 881             ("aggregated port is not distributing"));
 882
 883         return (lp->lp_lagg);
 884 }
 885
 886 #if defined(RATELIMIT) || defined(KERN_TLS)
 887 struct lagg_port *
 888 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t flowid)
 889 {
 890         struct lacp_softc *lsc = LACP_SOFTC(sc);
 891         struct lacp_portmap *pm;
 892         struct lacp_port *lp;
 893         uint32_t hash;
 894
 895         if (__predict_false(lsc->lsc_suppress_distributing)) {
 896                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 897                 return (NULL);
 898         }
 899
 900         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 901         if (pm->pm_count == 0) {
 902                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 903                 return (NULL);
 904         }
 905
 906         hash = flowid >> sc->flowid_shift;
 907         hash %= pm->pm_count;
 908         lp = pm->pm_map[hash];
 909
 910         return (lp->lp_lagg);
 911 }
 912 #endif
 913
 914 /*
 915  * lacp_suppress_distributing: drop transmit packets for a while
 916  * to preserve packet ordering.
 917  */
 918
 919 static void
 920 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 921 {
 922         struct lacp_port *lp;
 923
 924         if (lsc->lsc_active_aggregator != la) {
 925                 return;
 926         }
 927
 928         LACP_TRACE(NULL);
 929
 930         lsc->lsc_suppress_distributing = TRUE;
 931
 932         /* send a marker frame down each port to verify the queues are empty */
 933         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 934                 lp->lp_flags |= LACP_PORT_MARK;
 935                 lacp_xmit_marker(lp);
 936         }
 937
 938         /* set a timeout for the marker frames */
 939         callout_reset(&lsc->lsc_transit_callout,
 940             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 941 }
 942
 943 static int
 944 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 945     const struct lacp_peerinfo *b)
 946 {
 947         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 948 }
 949
 950 static int
 951 lacp_compare_systemid(const struct lacp_systemid *a,
 952     const struct lacp_systemid *b)
 953 {
 954         return (memcmp(a, b, sizeof(*a)));
 955 }
 956
 957 #if 0   /* unused */
 958 static int
 959 lacp_compare_portid(const struct lacp_portid *a,
 960     const struct lacp_portid *b)
 961 {
 962         return (memcmp(a, b, sizeof(*a)));
 963 }
 964 #endif
 965
 966 static uint64_t
 967 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 968 {
 969         struct lacp_port *lp;
 970         uint64_t speed;
 971
 972         lp = TAILQ_FIRST(&la->la_ports);
 973         if (lp == NULL) {
 974                 return (0);
 975         }
 976
 977         speed = ifmedia_baudrate(lp->lp_media);
 978         speed *= la->la_nports;
 979         if (speed == 0) {
 980                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 981                     lp->lp_media, la->la_nports));
 982         }
 983
 984         return (speed);
 985 }
 986
 987 /*
 988  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 989  * packets from lagg(4) interface.
 990  */
 991
 992 static void
 993 lacp_select_active_aggregator(struct lacp_softc *lsc)
 994 {
 995         struct lacp_aggregator *la;
 996         struct lacp_aggregator *best_la = NULL;
 997         uint64_t best_speed = 0;
 998         char buf[LACP_LAGIDSTR_MAX+1];
 999
1000         LACP_TRACE(NULL);
1001
1002         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1003                 uint64_t speed;
1004
1005                 if (la->la_nports == 0) {
1006                         continue;
1007                 }
1008
1009                 speed = lacp_aggregator_bandwidth(la);
1010                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
1011                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
1012                     speed, la->la_nports));
1013
1014                 /*
1015                  * This aggregator is chosen if the partner has a better
1016                  * system priority or, the total aggregated speed is higher
1017                  * or, it is already the chosen aggregator
1018                  */
1019                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
1020                     LACP_SYS_PRI(best_la->la_partner)) ||
1021                     speed > best_speed ||
1022                     (speed == best_speed &&
1023                     la == lsc->lsc_active_aggregator)) {
1024                         best_la = la;
1025                         best_speed = speed;
1026                 }
1027         }
1028
1029         KASSERT(best_la == NULL || best_la->la_nports > 0,
1030             ("invalid aggregator refcnt"));
1031         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1032             ("invalid aggregator list"));
1033
1034         if (lsc->lsc_active_aggregator != best_la) {
1035                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1036                 LACP_DPRINTF((NULL, "old %s\n",
1037                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1038                     buf, sizeof(buf))));
1039         } else {
1040                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1041         }
1042         LACP_DPRINTF((NULL, "new %s\n",
1043             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1044
1045         if (lsc->lsc_active_aggregator != best_la) {
1046                 lsc->lsc_active_aggregator = best_la;
1047                 lacp_update_portmap(lsc);
1048                 if (best_la) {
1049                         lacp_suppress_distributing(lsc, best_la);
1050                 }
1051         }
1052 }
1053
1054 /*
1055  * Updated the inactive portmap array with the new list of ports and
1056  * make it live.
1057  */
1058 static void
1059 lacp_update_portmap(struct lacp_softc *lsc)
1060 {
1061         struct lagg_softc *sc = lsc->lsc_softc;
1062         struct lacp_aggregator *la;
1063         struct lacp_portmap *p;
1064         struct lacp_port *lp;
1065         uint64_t speed;
1066         u_int newmap;
1067         int i;
1068 #ifdef NUMA
1069         int count;
1070         uint8_t domain;
1071 #endif
1072
1073         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1074         p = &lsc->lsc_pmap[newmap];
1075         la = lsc->lsc_active_aggregator;
1076         speed = 0;
1077         bzero(p, sizeof(struct lacp_portmap));
1078
1079         if (la != NULL && la->la_nports > 0) {
1080                 p->pm_count = la->la_nports;
1081                 i = 0;
1082                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q) {
1083                         p->pm_map[i++] = lp;
1084 #ifdef NUMA
1085                         domain = lp->lp_ifp->if_numa_domain;
1086                         if (domain >= MAXMEMDOM)
1087                                 continue;
1088                         count = p->pm_numa[domain].count;
1089                         p->pm_numa[domain].map[count] = lp;
1090                         p->pm_numa[domain].count++;
1091 #endif
1092                 }
1093                 KASSERT(i == p->pm_count, ("Invalid port count"));
1094
1095 #ifdef NUMA
1096                 for (i = 0; i < MAXMEMDOM; i++) {
1097                         if (p->pm_numa[i].count != 0)
1098                                 p->pm_num_dom++;
1099                 }
1100 #endif
1101                 speed = lacp_aggregator_bandwidth(la);
1102         }
1103         sc->sc_ifp->if_baudrate = speed;
1104
1105         /* switch the active portmap over */
1106         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1107         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1108                     lsc->lsc_activemap,
1109                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1110 }
1111
1112 static uint16_t
1113 lacp_compose_key(struct lacp_port *lp)
1114 {
1115         struct lagg_port *lgp = lp->lp_lagg;
1116         struct lagg_softc *sc = lgp->lp_softc;
1117         u_int media = lp->lp_media;
1118         uint16_t key;
1119
1120         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1121
1122                 /*
1123                  * non-aggregatable links should have unique keys.
1124                  *
1125                  * XXX this isn't really unique as if_index is 16 bit.
1126                  */
1127
1128                 /* bit 0..14:   (some bits of) if_index of this port */
1129                 key = lp->lp_ifp->if_index;
1130                 /* bit 15:      1 */
1131                 key |= 0x8000;
1132         } else {
1133                 u_int subtype = IFM_SUBTYPE(media);
1134
1135                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1136                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1137
1138                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1139                 switch (subtype) {
1140                 case IFM_10_T:
1141                 case IFM_10_2:
1142                 case IFM_10_5:
1143                 case IFM_10_STP:
1144                 case IFM_10_FL:
1145                         key = IFM_10_T;
1146                         break;
1147                 case IFM_100_TX:
1148                 case IFM_100_FX:
1149                 case IFM_100_T4:
1150                 case IFM_100_VG:
1151                 case IFM_100_T2:
1152                 case IFM_100_T:
1153                 case IFM_100_SGMII:
1154                         key = IFM_100_TX;
1155                         break;
1156                 case IFM_1000_SX:
1157                 case IFM_1000_LX:
1158                 case IFM_1000_CX:
1159                 case IFM_1000_T:
1160                 case IFM_1000_KX:
1161                 case IFM_1000_SGMII:
1162                 case IFM_1000_CX_SGMII:
1163                         key = IFM_1000_SX;
1164                         break;
1165                 case IFM_10G_LR:
1166                 case IFM_10G_SR:
1167                 case IFM_10G_CX4:
1168                 case IFM_10G_TWINAX:
1169                 case IFM_10G_TWINAX_LONG:
1170                 case IFM_10G_LRM:
1171                 case IFM_10G_T:
1172                 case IFM_10G_KX4:
1173                 case IFM_10G_KR:
1174                 case IFM_10G_CR1:
1175                 case IFM_10G_ER:
1176                 case IFM_10G_SFI:
1177                 case IFM_10G_AOC:
1178                         key = IFM_10G_LR;
1179                         break;
1180                 case IFM_20G_KR2:
1181                         key = IFM_20G_KR2;
1182                         break;
1183                 case IFM_2500_KX:
1184                 case IFM_2500_T:
1185                 case IFM_2500_X:
1186                         key = IFM_2500_KX;
1187                         break;
1188                 case IFM_5000_T:
1189                 case IFM_5000_KR:
1190                 case IFM_5000_KR_S:
1191                 case IFM_5000_KR1:
1192                         key = IFM_5000_T;
1193                         break;
1194                 case IFM_50G_PCIE:
1195                 case IFM_50G_CR2:
1196                 case IFM_50G_KR2:
1197                 case IFM_50G_SR2:
1198                 case IFM_50G_LR2:
1199                 case IFM_50G_LAUI2_AC:
1200                 case IFM_50G_LAUI2:
1201                 case IFM_50G_AUI2_AC:
1202                 case IFM_50G_AUI2:
1203                 case IFM_50G_CP:
1204                 case IFM_50G_SR:
1205                 case IFM_50G_LR:
1206                 case IFM_50G_FR:
1207                 case IFM_50G_KR_PAM4:
1208                 case IFM_50G_AUI1_AC:
1209                 case IFM_50G_AUI1:
1210                         key = IFM_50G_PCIE;
1211                         break;
1212                 case IFM_56G_R4:
1213                         key = IFM_56G_R4;
1214                         break;
1215                 case IFM_25G_PCIE:
1216                 case IFM_25G_CR:
1217                 case IFM_25G_KR:
1218                 case IFM_25G_SR:
1219                 case IFM_25G_LR:
1220                 case IFM_25G_ACC:
1221                 case IFM_25G_AOC:
1222                 case IFM_25G_T:
1223                 case IFM_25G_CR_S:
1224                 case IFM_25G_CR1:
1225                 case IFM_25G_KR_S:
1226                 case IFM_25G_AUI:
1227                 case IFM_25G_KR1:
1228                         key = IFM_25G_PCIE;
1229                         break;
1230                 case IFM_40G_CR4:
1231                 case IFM_40G_SR4:
1232                 case IFM_40G_LR4:
1233                 case IFM_40G_XLPPI:
1234                 case IFM_40G_KR4:
1235                 case IFM_40G_XLAUI:
1236                 case IFM_40G_XLAUI_AC:
1237                 case IFM_40G_ER4:
1238                         key = IFM_40G_CR4;
1239                         break;
1240                 case IFM_100G_CR4:
1241                 case IFM_100G_SR4:
1242                 case IFM_100G_KR4:
1243                 case IFM_100G_LR4:
1244                 case IFM_100G_CAUI4_AC:
1245                 case IFM_100G_CAUI4:
1246                 case IFM_100G_AUI4_AC:
1247                 case IFM_100G_AUI4:
1248                 case IFM_100G_CR_PAM4:
1249                 case IFM_100G_KR_PAM4:
1250                 case IFM_100G_CP2:
1251                 case IFM_100G_SR2:
1252                 case IFM_100G_DR:
1253                 case IFM_100G_KR2_PAM4:
1254                 case IFM_100G_CAUI2_AC:
1255                 case IFM_100G_CAUI2:
1256                 case IFM_100G_AUI2_AC:
1257                 case IFM_100G_AUI2:
1258                         key = IFM_100G_CR4;
1259                         break;
1260                 case IFM_200G_CR4_PAM4:
1261                 case IFM_200G_SR4:
1262                 case IFM_200G_FR4:
1263                 case IFM_200G_LR4:
1264                 case IFM_200G_DR4:
1265                 case IFM_200G_KR4_PAM4:
1266                 case IFM_200G_AUI4_AC:
1267                 case IFM_200G_AUI4:
1268                 case IFM_200G_AUI8_AC:
1269                 case IFM_200G_AUI8:
1270                         key = IFM_200G_CR4_PAM4;
1271                         break;
1272                 case IFM_400G_FR8:
1273                 case IFM_400G_LR8:
1274                 case IFM_400G_DR4:
1275                 case IFM_400G_AUI8_AC:
1276                 case IFM_400G_AUI8:
1277                         key = IFM_400G_FR8;
1278                         break;
1279                 default:
1280                         key = subtype;
1281                         break;
1282                 }
1283                 /* bit 5..14:   (some bits of) if_index of lagg device */
1284                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1285                 /* bit 15:      0 */
1286         }
1287         return (htons(key));
1288 }
1289
1290 static void
1291 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1292 {
1293         char buf[LACP_LAGIDSTR_MAX+1];
1294
1295         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1296             __func__,
1297             lacp_format_lagid(&la->la_actor, &la->la_partner,
1298             buf, sizeof(buf)),
1299             la->la_refcnt, la->la_refcnt + 1));
1300
1301         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1302         la->la_refcnt++;
1303         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1304 }
1305
1306 static void
1307 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1308 {
1309         char buf[LACP_LAGIDSTR_MAX+1];
1310
1311         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1312             __func__,
1313             lacp_format_lagid(&la->la_actor, &la->la_partner,
1314             buf, sizeof(buf)),
1315             la->la_refcnt, la->la_refcnt - 1));
1316
1317         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1318         la->la_refcnt--;
1319         if (la->la_refcnt > 0) {
1320                 return;
1321         }
1322
1323         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1324         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1325
1326         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1327
1328         free(la, M_DEVBUF);
1329 }
1330
1331 /*
1332  * lacp_aggregator_get: allocate an aggregator.
1333  */
1334
1335 static struct lacp_aggregator *
1336 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1337 {
1338         struct lacp_aggregator *la;
1339
1340         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1341         if (la) {
1342                 la->la_refcnt = 1;
1343                 la->la_nports = 0;
1344                 TAILQ_INIT(&la->la_ports);
1345                 la->la_pending = 0;
1346                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1347         }
1348
1349         return (la);
1350 }
1351
1352 /*
1353  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1354  */
1355
1356 static void
1357 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1358 {
1359         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1360         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1361
1362         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1363 }
1364
1365 static void
1366 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1367     const struct lacp_peerinfo *lpi_port)
1368 {
1369         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1370         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1371         lpi_aggr->lip_key = lpi_port->lip_key;
1372 }
1373
1374 /*
1375  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1376  */
1377
1378 static int
1379 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1380     const struct lacp_port *lp)
1381 {
1382         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1383             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1384                 return (0);
1385         }
1386
1387         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1388                 return (0);
1389         }
1390
1391         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1392                 return (0);
1393         }
1394
1395         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1396                 return (0);
1397         }
1398
1399         return (1);
1400 }
1401
1402 static int
1403 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1404     const struct lacp_peerinfo *b)
1405 {
1406         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1407             sizeof(a->lip_systemid))) {
1408                 return (0);
1409         }
1410
1411         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1412                 return (0);
1413         }
1414
1415         return (1);
1416 }
1417
1418 static void
1419 lacp_port_enable(struct lacp_port *lp)
1420 {
1421         lp->lp_state |= LACP_STATE_AGGREGATION;
1422 }
1423
1424 static void
1425 lacp_port_disable(struct lacp_port *lp)
1426 {
1427         lacp_set_mux(lp, LACP_MUX_DETACHED);
1428
1429         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1430         lp->lp_selected = LACP_UNSELECTED;
1431         lacp_sm_rx_record_default(lp);
1432         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1433         lp->lp_state &= ~LACP_STATE_EXPIRED;
1434 }
1435
1436 /*
1437  * lacp_select: select an aggregator.  create one if necessary.
1438  */
1439 static void
1440 lacp_select(struct lacp_port *lp)
1441 {
1442         struct lacp_softc *lsc = lp->lp_lsc;
1443         struct lacp_aggregator *la;
1444         char buf[LACP_LAGIDSTR_MAX+1];
1445
1446         if (lp->lp_aggregator) {
1447                 return;
1448         }
1449
1450         /* If we haven't heard from our peer, skip this step. */
1451         if (lp->lp_state & LACP_STATE_DEFAULTED)
1452                 return;
1453
1454         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1455             ("timer_wait_while still active"));
1456
1457         LACP_DPRINTF((lp, "port lagid=%s\n",
1458             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1459             buf, sizeof(buf))));
1460
1461         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1462                 if (lacp_aggregator_is_compatible(la, lp)) {
1463                         break;
1464                 }
1465         }
1466
1467         if (la == NULL) {
1468                 la = lacp_aggregator_get(lsc, lp);
1469                 if (la == NULL) {
1470                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1471
1472                         /*
1473                          * will retry on the next tick.
1474                          */
1475
1476                         return;
1477                 }
1478                 lacp_fill_aggregator_id(la, lp);
1479                 LACP_DPRINTF((lp, "aggregator created\n"));
1480         } else {
1481                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1482                 if (la->la_refcnt == LACP_MAX_PORTS)
1483                         return;
1484                 lacp_aggregator_addref(lsc, la);
1485         }
1486
1487         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1488             lacp_format_lagid(&la->la_actor, &la->la_partner,
1489             buf, sizeof(buf))));
1490
1491         lp->lp_aggregator = la;
1492         lp->lp_selected = LACP_SELECTED;
1493 }
1494
1495 /*
1496  * lacp_unselect: finish unselect/detach process.
1497  */
1498
1499 static void
1500 lacp_unselect(struct lacp_port *lp)
1501 {
1502         struct lacp_softc *lsc = lp->lp_lsc;
1503         struct lacp_aggregator *la = lp->lp_aggregator;
1504
1505         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1506             ("timer_wait_while still active"));
1507
1508         if (la == NULL) {
1509                 return;
1510         }
1511
1512         lp->lp_aggregator = NULL;
1513         lacp_aggregator_delref(lsc, la);
1514 }
1515
1516 /* mux machine */
1517
1518 static void
1519 lacp_sm_mux(struct lacp_port *lp)
1520 {
1521         struct lagg_port *lgp = lp->lp_lagg;
1522         struct lagg_softc *sc = lgp->lp_softc;
1523         enum lacp_mux_state new_state;
1524         boolean_t p_sync =
1525                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1526         boolean_t p_collecting =
1527             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1528         enum lacp_selected selected = lp->lp_selected;
1529         struct lacp_aggregator *la;
1530
1531         if (V_lacp_debug > 1)
1532                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1533                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1534                     lp->lp_mux_state, selected, p_sync, p_collecting);
1535
1536 re_eval:
1537         la = lp->lp_aggregator;
1538         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1539             ("MUX not detached"));
1540         new_state = lp->lp_mux_state;
1541         switch (lp->lp_mux_state) {
1542         case LACP_MUX_DETACHED:
1543                 if (selected != LACP_UNSELECTED) {
1544                         new_state = LACP_MUX_WAITING;
1545                 }
1546                 break;
1547         case LACP_MUX_WAITING:
1548                 KASSERT(la->la_pending > 0 ||
1549                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1550                     ("timer_wait_while still active"));
1551                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1552                         new_state = LACP_MUX_ATTACHED;
1553                 } else if (selected == LACP_UNSELECTED) {
1554                         new_state = LACP_MUX_DETACHED;
1555                 }
1556                 break;
1557         case LACP_MUX_ATTACHED:
1558                 if (selected == LACP_SELECTED && p_sync) {
1559                         new_state = LACP_MUX_COLLECTING;
1560                 } else if (selected != LACP_SELECTED) {
1561                         new_state = LACP_MUX_DETACHED;
1562                 }
1563                 break;
1564         case LACP_MUX_COLLECTING:
1565                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1566                         new_state = LACP_MUX_DISTRIBUTING;
1567                 } else if (selected != LACP_SELECTED || !p_sync) {
1568                         new_state = LACP_MUX_ATTACHED;
1569                 }
1570                 break;
1571         case LACP_MUX_DISTRIBUTING:
1572                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1573                         new_state = LACP_MUX_COLLECTING;
1574                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1575                         sc->sc_flapping++;
1576                 }
1577                 break;
1578         default:
1579                 panic("%s: unknown state", __func__);
1580         }
1581
1582         if (lp->lp_mux_state == new_state) {
1583                 return;
1584         }
1585
1586         lacp_set_mux(lp, new_state);
1587         goto re_eval;
1588 }
1589
1590 static void
1591 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1592 {
1593         struct lacp_aggregator *la = lp->lp_aggregator;
1594
1595         if (lp->lp_mux_state == new_state) {
1596                 return;
1597         }
1598
1599         switch (new_state) {
1600         case LACP_MUX_DETACHED:
1601                 lp->lp_state &= ~LACP_STATE_SYNC;
1602                 lacp_disable_distributing(lp);
1603                 lacp_disable_collecting(lp);
1604                 lacp_sm_assert_ntt(lp);
1605                 /* cancel timer */
1606                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1607                         KASSERT(la->la_pending > 0,
1608                             ("timer_wait_while not active"));
1609                         la->la_pending--;
1610                 }
1611                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1612                 lacp_unselect(lp);
1613                 break;
1614         case LACP_MUX_WAITING:
1615                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1616                     LACP_AGGREGATE_WAIT_TIME);
1617                 la->la_pending++;
1618                 break;
1619         case LACP_MUX_ATTACHED:
1620                 lp->lp_state |= LACP_STATE_SYNC;
1621                 lacp_disable_collecting(lp);
1622                 lacp_sm_assert_ntt(lp);
1623                 break;
1624         case LACP_MUX_COLLECTING:
1625                 lacp_enable_collecting(lp);
1626                 lacp_disable_distributing(lp);
1627                 lacp_sm_assert_ntt(lp);
1628                 break;
1629         case LACP_MUX_DISTRIBUTING:
1630                 lacp_enable_distributing(lp);
1631                 break;
1632         default:
1633                 panic("%s: unknown state", __func__);
1634         }
1635
1636         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1637
1638         lp->lp_mux_state = new_state;
1639 }
1640
1641 static void
1642 lacp_sm_mux_timer(struct lacp_port *lp)
1643 {
1644         struct lacp_aggregator *la = lp->lp_aggregator;
1645         char buf[LACP_LAGIDSTR_MAX+1];
1646
1647         KASSERT(la->la_pending > 0, ("no pending event"));
1648
1649         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1650             lacp_format_lagid(&la->la_actor, &la->la_partner,
1651             buf, sizeof(buf)),
1652             la->la_pending, la->la_pending - 1));
1653
1654         la->la_pending--;
1655 }
1656
1657 /* periodic transmit machine */
1658
1659 static void
1660 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1661 {
1662         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1663             LACP_STATE_TIMEOUT)) {
1664                 return;
1665         }
1666
1667         LACP_DPRINTF((lp, "partner timeout changed\n"));
1668
1669         /*
1670          * FAST_PERIODIC -> SLOW_PERIODIC
1671          * or
1672          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1673          *
1674          * let lacp_sm_ptx_tx_schedule to update timeout.
1675          */
1676
1677         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1678
1679         /*
1680          * if timeout has been shortened, assert NTT.
1681          */
1682
1683         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1684                 lacp_sm_assert_ntt(lp);
1685         }
1686 }
1687
1688 static void
1689 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1690 {
1691         int timeout;
1692
1693         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1694             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1695
1696                 /*
1697                  * NO_PERIODIC
1698                  */
1699
1700                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1701                 return;
1702         }
1703
1704         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1705                 return;
1706         }
1707
1708         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1709             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1710
1711         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1712 }
1713
1714 static void
1715 lacp_sm_ptx_timer(struct lacp_port *lp)
1716 {
1717         lacp_sm_assert_ntt(lp);
1718 }
1719
1720 static void
1721 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1722 {
1723         int timeout;
1724
1725         /*
1726          * check LACP_DISABLED first
1727          */
1728
1729         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1730                 return;
1731         }
1732
1733         /*
1734          * check loopback condition.
1735          */
1736
1737         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1738             &lp->lp_actor.lip_systemid)) {
1739                 return;
1740         }
1741
1742         /*
1743          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1744          */
1745
1746         lacp_sm_rx_update_selected(lp, du);
1747         lacp_sm_rx_update_ntt(lp, du);
1748         lacp_sm_rx_record_pdu(lp, du);
1749
1750         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1751             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1752         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1753
1754         lp->lp_state &= ~LACP_STATE_EXPIRED;
1755
1756         /*
1757          * kick transmit machine without waiting the next tick.
1758          */
1759
1760         lacp_sm_tx(lp);
1761 }
1762
1763 static void
1764 lacp_sm_rx_set_expired(struct lacp_port *lp)
1765 {
1766         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1767         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1768         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1769         lp->lp_state |= LACP_STATE_EXPIRED;
1770 }
1771
1772 static void
1773 lacp_sm_rx_timer(struct lacp_port *lp)
1774 {
1775         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1776                 /* CURRENT -> EXPIRED */
1777                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1778                 lacp_sm_rx_set_expired(lp);
1779         } else {
1780                 /* EXPIRED -> DEFAULTED */
1781                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1782                 lacp_sm_rx_update_default_selected(lp);
1783                 lacp_sm_rx_record_default(lp);
1784                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1785         }
1786 }
1787
1788 static void
1789 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1790 {
1791         boolean_t active;
1792         uint8_t oldpstate;
1793         char buf[LACP_STATESTR_MAX+1];
1794
1795         LACP_TRACE(lp);
1796
1797         oldpstate = lp->lp_partner.lip_state;
1798
1799         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1800             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1801             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1802
1803         lp->lp_partner = du->ldu_actor;
1804         if (active &&
1805             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1806             LACP_STATE_AGGREGATION) &&
1807             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1808             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1809                 /*
1810                  * XXX Maintain legacy behavior of leaving the
1811                  * LACP_STATE_SYNC bit unchanged from the partner's
1812                  * advertisement if lsc_strict_mode is false.
1813                  * TODO: We should re-examine the concept of the "strict mode"
1814                  * to ensure it makes sense to maintain a non-strict mode.
1815                  */
1816                 if (lp->lp_lsc->lsc_strict_mode)
1817                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1818         } else {
1819                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1820         }
1821
1822         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1823
1824         if (oldpstate != lp->lp_partner.lip_state) {
1825                 LACP_DPRINTF((lp, "old pstate %s\n",
1826                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1827                 LACP_DPRINTF((lp, "new pstate %s\n",
1828                     lacp_format_state(lp->lp_partner.lip_state, buf,
1829                     sizeof(buf))));
1830         }
1831
1832         lacp_sm_ptx_update_timeout(lp, oldpstate);
1833 }
1834
1835 static void
1836 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1837 {
1838
1839         LACP_TRACE(lp);
1840
1841         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1842             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1843             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1844                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1845                 lacp_sm_assert_ntt(lp);
1846         }
1847 }
1848
1849 static void
1850 lacp_sm_rx_record_default(struct lacp_port *lp)
1851 {
1852         uint8_t oldpstate;
1853
1854         LACP_TRACE(lp);
1855
1856         oldpstate = lp->lp_partner.lip_state;
1857         if (lp->lp_lsc->lsc_strict_mode)
1858                 lp->lp_partner = lacp_partner_admin_strict;
1859         else
1860                 lp->lp_partner = lacp_partner_admin_optimistic;
1861         lp->lp_state |= LACP_STATE_DEFAULTED;
1862         lacp_sm_ptx_update_timeout(lp, oldpstate);
1863 }
1864
1865 static void
1866 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1867     const struct lacp_peerinfo *info)
1868 {
1869
1870         LACP_TRACE(lp);
1871
1872         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1873             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1874             LACP_STATE_AGGREGATION)) {
1875                 lp->lp_selected = LACP_UNSELECTED;
1876                 /* mux machine will clean up lp->lp_aggregator */
1877         }
1878 }
1879
1880 static void
1881 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1882 {
1883
1884         LACP_TRACE(lp);
1885
1886         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1887 }
1888
1889 static void
1890 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1891 {
1892
1893         LACP_TRACE(lp);
1894
1895         if (lp->lp_lsc->lsc_strict_mode)
1896                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1897                     &lacp_partner_admin_strict);
1898         else
1899                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1900                     &lacp_partner_admin_optimistic);
1901 }
1902
1903 /* transmit machine */
1904
1905 static void
1906 lacp_sm_tx(struct lacp_port *lp)
1907 {
1908         int error = 0;
1909
1910         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1911 #if 1
1912             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1913             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1914 #endif
1915             ) {
1916                 lp->lp_flags &= ~LACP_PORT_NTT;
1917         }
1918
1919         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1920                 return;
1921         }
1922
1923         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1924         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1925                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1926                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1927                 return;
1928         }
1929
1930         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1931                 error = lacp_xmit_lacpdu(lp);
1932         } else {
1933                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1934         }
1935
1936         if (error == 0) {
1937                 lp->lp_flags &= ~LACP_PORT_NTT;
1938         } else {
1939                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1940                     error));
1941         }
1942 }
1943
1944 static void
1945 lacp_sm_assert_ntt(struct lacp_port *lp)
1946 {
1947
1948         lp->lp_flags |= LACP_PORT_NTT;
1949 }
1950
1951 static void
1952 lacp_run_timers(struct lacp_port *lp)
1953 {
1954         int i;
1955
1956         for (i = 0; i < LACP_NTIMER; i++) {
1957                 KASSERT(lp->lp_timer[i] >= 0,
1958                     ("invalid timer value %d", lp->lp_timer[i]));
1959                 if (lp->lp_timer[i] == 0) {
1960                         continue;
1961                 } else if (--lp->lp_timer[i] <= 0) {
1962                         if (lacp_timer_funcs[i]) {
1963                                 (*lacp_timer_funcs[i])(lp);
1964                         }
1965                 }
1966         }
1967 }
1968
1969 int
1970 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1971 {
1972         struct lacp_softc *lsc = lp->lp_lsc;
1973         struct lagg_port *lgp = lp->lp_lagg;
1974         struct lacp_port *lp2;
1975         struct markerdu *mdu;
1976         int error = 0;
1977         int pending = 0;
1978
1979         if (m->m_pkthdr.len != sizeof(*mdu)) {
1980                 goto bad;
1981         }
1982
1983         if ((m->m_flags & M_MCAST) == 0) {
1984                 goto bad;
1985         }
1986
1987         if (m->m_len < sizeof(*mdu)) {
1988                 m = m_pullup(m, sizeof(*mdu));
1989                 if (m == NULL) {
1990                         return (ENOMEM);
1991                 }
1992         }
1993
1994         mdu = mtod(m, struct markerdu *);
1995
1996         if (memcmp(&mdu->mdu_eh.ether_dhost,
1997             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1998                 goto bad;
1999         }
2000
2001         if (mdu->mdu_sph.sph_version != 1) {
2002                 goto bad;
2003         }
2004
2005         switch (mdu->mdu_tlv.tlv_type) {
2006         case MARKER_TYPE_INFO:
2007                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2008                     marker_info_tlv_template, TRUE)) {
2009                         goto bad;
2010                 }
2011                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
2012                 memcpy(&mdu->mdu_eh.ether_dhost,
2013                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
2014                 memcpy(&mdu->mdu_eh.ether_shost,
2015                     lgp->lp_lladdr, ETHER_ADDR_LEN);
2016                 error = lagg_enqueue(lp->lp_ifp, m);
2017                 break;
2018
2019         case MARKER_TYPE_RESPONSE:
2020                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2021                     marker_response_tlv_template, TRUE)) {
2022                         goto bad;
2023                 }
2024                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
2025                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
2026                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
2027
2028                 /* Verify that it is the last marker we sent out */
2029                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
2030                     sizeof(struct lacp_markerinfo)))
2031                         goto bad;
2032
2033                 LACP_LOCK(lsc);
2034                 lp->lp_flags &= ~LACP_PORT_MARK;
2035
2036                 if (lsc->lsc_suppress_distributing) {
2037                         /* Check if any ports are waiting for a response */
2038                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
2039                                 if (lp2->lp_flags & LACP_PORT_MARK) {
2040                                         pending = 1;
2041                                         break;
2042                                 }
2043                         }
2044
2045                         if (pending == 0) {
2046                                 /* All interface queues are clear */
2047                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
2048                                 lsc->lsc_suppress_distributing = FALSE;
2049                         }
2050                 }
2051                 LACP_UNLOCK(lsc);
2052                 m_freem(m);
2053                 break;
2054
2055         default:
2056                 goto bad;
2057         }
2058
2059         return (error);
2060
2061 bad:
2062         LACP_DPRINTF((lp, "bad marker frame\n"));
2063         m_freem(m);
2064         return (EINVAL);
2065 }
2066
2067 static int
2068 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
2069     const struct tlv_template *tmpl, boolean_t check_type)
2070 {
2071         while (/* CONSTCOND */ 1) {
2072                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
2073                         return (EINVAL);
2074                 }
2075                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
2076                     tlv->tlv_length != tmpl->tmpl_length) {
2077                         return (EINVAL);
2078                 }
2079                 if (tmpl->tmpl_type == 0) {
2080                         break;
2081                 }
2082                 tlv = (const struct tlvhdr *)
2083                     ((const char *)tlv + tlv->tlv_length);
2084                 tmpl++;
2085         }
2086
2087         return (0);
2088 }
2089
2090 /* Debugging */
2091 const char *
2092 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
2093 {
2094         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
2095             (int)mac[0],
2096             (int)mac[1],
2097             (int)mac[2],
2098             (int)mac[3],
2099             (int)mac[4],
2100             (int)mac[5]);
2101
2102         return (buf);
2103 }
2104
2105 const char *
2106 lacp_format_systemid(const struct lacp_systemid *sysid,
2107     char *buf, size_t buflen)
2108 {
2109         char macbuf[LACP_MACSTR_MAX+1];
2110
2111         snprintf(buf, buflen, "%04X,%s",
2112             ntohs(sysid->lsi_prio),
2113             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2114
2115         return (buf);
2116 }
2117
2118 const char *
2119 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2120 {
2121         snprintf(buf, buflen, "%04X,%04X",
2122             ntohs(portid->lpi_prio),
2123             ntohs(portid->lpi_portno));
2124
2125         return (buf);
2126 }
2127
2128 const char *
2129 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2130 {
2131         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2132         char portid[LACP_PORTIDSTR_MAX+1];
2133
2134         snprintf(buf, buflen, "(%s,%04X,%s)",
2135             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2136             ntohs(peer->lip_key),
2137             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2138
2139         return (buf);
2140 }
2141
2142 const char *
2143 lacp_format_lagid(const struct lacp_peerinfo *a,
2144     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2145 {
2146         char astr[LACP_PARTNERSTR_MAX+1];
2147         char bstr[LACP_PARTNERSTR_MAX+1];
2148
2149 #if 0
2150         /*
2151          * there's a convention to display small numbered peer
2152          * in the left.
2153          */
2154
2155         if (lacp_compare_peerinfo(a, b) > 0) {
2156                 const struct lacp_peerinfo *t;
2157
2158                 t = a;
2159                 a = b;
2160                 b = t;
2161         }
2162 #endif
2163
2164         snprintf(buf, buflen, "[%s,%s]",
2165             lacp_format_partner(a, astr, sizeof(astr)),
2166             lacp_format_partner(b, bstr, sizeof(bstr)));
2167
2168         return (buf);
2169 }
2170
2171 const char *
2172 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2173     char *buf, size_t buflen)
2174 {
2175         if (la == NULL) {
2176                 return ("(none)");
2177         }
2178
2179         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2180 }
2181
2182 const char *
2183 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2184 {
2185         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2186         return (buf);
2187 }
2188
2189 static void
2190 lacp_dump_lacpdu(const struct lacpdu *du)
2191 {
2192         char buf[LACP_PARTNERSTR_MAX+1];
2193         char buf2[LACP_STATESTR_MAX+1];
2194
2195         printf("actor=%s\n",
2196             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2197         printf("actor.state=%s\n",
2198             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2199         printf("partner=%s\n",
2200             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2201         printf("partner.state=%s\n",
2202             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2203
2204         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2205 }
2206
2207 static void
2208 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2209 {
2210         va_list va;
2211
2212         if (lp) {
2213                 printf("%s: ", lp->lp_ifp->if_xname);
2214         }
2215
2216         va_start(va, fmt);
2217         vprintf(fmt, va);
2218         va_end(va);
2219 }