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 #include <sys/time.h>
  53
  54 #include <net/if.h>
  55 #include <net/if_var.h>
  56 #include <net/if_dl.h>
  57 #include <net/ethernet.h>
  58 #include <net/infiniband.h>
  59 #include <net/if_media.h>
  60 #include <net/if_types.h>
  61
  62 #include <net/if_lagg.h>
  63 #include <net/ieee8023ad_lacp.h>
  64
  65 /*
  66  * actor system priority and port priority.
  67  * XXX should be configurable.
  68  */
  69
  70 #define LACP_SYSTEM_PRIO        0x8000
  71 #define LACP_PORT_PRIO          0x8000
  72
  73 const uint8_t ethermulticastaddr_slowprotocols[ETHER_ADDR_LEN] =
  74     { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x02 };
  75
  76 static const struct tlv_template lacp_info_tlv_template[] = {
  77         { LACP_TYPE_ACTORINFO,
  78             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  79         { LACP_TYPE_PARTNERINFO,
  80             sizeof(struct tlvhdr) + sizeof(struct lacp_peerinfo) },
  81         { LACP_TYPE_COLLECTORINFO,
  82             sizeof(struct tlvhdr) + sizeof(struct lacp_collectorinfo) },
  83         { 0, 0 },
  84 };
  85
  86 static const struct tlv_template marker_info_tlv_template[] = {
  87         { MARKER_TYPE_INFO,
  88             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  89         { 0, 0 },
  90 };
  91
  92 static const struct tlv_template marker_response_tlv_template[] = {
  93         { MARKER_TYPE_RESPONSE,
  94             sizeof(struct tlvhdr) + sizeof(struct lacp_markerinfo) },
  95         { 0, 0 },
  96 };
  97
  98 typedef void (*lacp_timer_func_t)(struct lacp_port *);
  99
 100 static void     lacp_fill_actorinfo(struct lacp_port *, struct lacp_peerinfo *);
 101 static void     lacp_fill_markerinfo(struct lacp_port *,
 102                     struct lacp_markerinfo *);
 103
 104 static uint64_t lacp_aggregator_bandwidth(struct lacp_aggregator *);
 105 static void     lacp_suppress_distributing(struct lacp_softc *,
 106                     struct lacp_aggregator *);
 107 static void     lacp_transit_expire(void *);
 108 static void     lacp_update_portmap(struct lacp_softc *);
 109 static void     lacp_select_active_aggregator(struct lacp_softc *);
 110 static uint16_t lacp_compose_key(struct lacp_port *);
 111 static int      tlv_check(const void *, size_t, const struct tlvhdr *,
 112                     const struct tlv_template *, boolean_t);
 113 static void     lacp_tick(void *);
 114
 115 static void     lacp_fill_aggregator_id(struct lacp_aggregator *,
 116                     const struct lacp_port *);
 117 static void     lacp_fill_aggregator_id_peer(struct lacp_peerinfo *,
 118                     const struct lacp_peerinfo *);
 119 static int      lacp_aggregator_is_compatible(const struct lacp_aggregator *,
 120                     const struct lacp_port *);
 121 static int      lacp_peerinfo_is_compatible(const struct lacp_peerinfo *,
 122                     const struct lacp_peerinfo *);
 123
 124 static struct lacp_aggregator *lacp_aggregator_get(struct lacp_softc *,
 125                     struct lacp_port *);
 126 static void     lacp_aggregator_addref(struct lacp_softc *,
 127                     struct lacp_aggregator *);
 128 static void     lacp_aggregator_delref(struct lacp_softc *,
 129                     struct lacp_aggregator *);
 130
 131 /* receive machine */
 132
 133 static int      lacp_pdu_input(struct lacp_port *, struct mbuf *);
 134 static int      lacp_marker_input(struct lacp_port *, struct mbuf *);
 135 static void     lacp_sm_rx(struct lacp_port *, const struct lacpdu *);
 136 static void     lacp_sm_rx_timer(struct lacp_port *);
 137 static void     lacp_sm_rx_set_expired(struct lacp_port *);
 138 static void     lacp_sm_rx_update_ntt(struct lacp_port *,
 139                     const struct lacpdu *);
 140 static void     lacp_sm_rx_record_pdu(struct lacp_port *,
 141                     const struct lacpdu *);
 142 static void     lacp_sm_rx_update_selected(struct lacp_port *,
 143                     const struct lacpdu *);
 144 static void     lacp_sm_rx_record_default(struct lacp_port *);
 145 static void     lacp_sm_rx_update_default_selected(struct lacp_port *);
 146 static void     lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *,
 147                     const struct lacp_peerinfo *);
 148
 149 /* mux machine */
 150
 151 static void     lacp_sm_mux(struct lacp_port *);
 152 static void     lacp_set_mux(struct lacp_port *, enum lacp_mux_state);
 153 static void     lacp_sm_mux_timer(struct lacp_port *);
 154
 155 /* periodic transmit machine */
 156
 157 static void     lacp_sm_ptx_update_timeout(struct lacp_port *, uint8_t);
 158 static void     lacp_sm_ptx_tx_schedule(struct lacp_port *);
 159 static void     lacp_sm_ptx_timer(struct lacp_port *);
 160
 161 /* transmit machine */
 162
 163 static void     lacp_sm_tx(struct lacp_port *);
 164 static void     lacp_sm_assert_ntt(struct lacp_port *);
 165
 166 static void     lacp_run_timers(struct lacp_port *);
 167 static int      lacp_compare_peerinfo(const struct lacp_peerinfo *,
 168                     const struct lacp_peerinfo *);
 169 static int      lacp_compare_systemid(const struct lacp_systemid *,
 170                     const struct lacp_systemid *);
 171 static void     lacp_port_enable(struct lacp_port *);
 172 static void     lacp_port_disable(struct lacp_port *);
 173 static void     lacp_select(struct lacp_port *);
 174 static void     lacp_unselect(struct lacp_port *);
 175 static void     lacp_disable_collecting(struct lacp_port *);
 176 static void     lacp_enable_collecting(struct lacp_port *);
 177 static void     lacp_disable_distributing(struct lacp_port *);
 178 static void     lacp_enable_distributing(struct lacp_port *);
 179 static int      lacp_xmit_lacpdu(struct lacp_port *);
 180 static int      lacp_xmit_marker(struct lacp_port *);
 181
 182 /* Debugging */
 183
 184 static void     lacp_dump_lacpdu(const struct lacpdu *);
 185 static const char *lacp_format_partner(const struct lacp_peerinfo *, char *,
 186                     size_t);
 187 static const char *lacp_format_lagid(const struct lacp_peerinfo *,
 188                     const struct lacp_peerinfo *, char *, size_t);
 189 static const char *lacp_format_lagid_aggregator(const struct lacp_aggregator *,
 190                     char *, size_t);
 191 static const char *lacp_format_state(uint8_t, char *, size_t);
 192 static const char *lacp_format_mac(const uint8_t *, char *, size_t);
 193 static const char *lacp_format_systemid(const struct lacp_systemid *, char *,
 194                     size_t);
 195 static const char *lacp_format_portid(const struct lacp_portid *, char *,
 196                     size_t);
 197 static void     lacp_dprintf(const struct lacp_port *, const char *, ...)
 198                     __attribute__((__format__(__printf__, 2, 3)));
 199
 200 VNET_DEFINE_STATIC(int, lacp_debug);
 201 #define V_lacp_debug    VNET(lacp_debug)
 202 SYSCTL_NODE(_net_link_lagg, OID_AUTO, lacp, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
 203     "ieee802.3ad");
 204 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, debug, CTLFLAG_RWTUN | CTLFLAG_VNET,
 205     &VNET_NAME(lacp_debug), 0, "Enable LACP debug logging (1=debug, 2=trace)");
 206
 207 VNET_DEFINE_STATIC(int, lacp_default_strict_mode) = 1;
 208 SYSCTL_INT(_net_link_lagg_lacp, OID_AUTO, default_strict_mode,
 209     CTLFLAG_RWTUN | CTLFLAG_VNET, &VNET_NAME(lacp_default_strict_mode), 0,
 210     "LACP strict protocol compliance default");
 211 #define LACP_DPRINTF(a) if (V_lacp_debug & 0x01) { lacp_dprintf a ; }
 212 #define LACP_TRACE(a) if (V_lacp_debug & 0x02) { lacp_dprintf(a,"%s\n",__func__); }
 213 #define LACP_TPRINTF(a) if (V_lacp_debug & 0x04) { lacp_dprintf a ; }
 214
 215 /*
 216  * partner administration variables.
 217  * XXX should be configurable.
 218  */
 219
 220 static const struct lacp_peerinfo lacp_partner_admin_optimistic = {
 221         .lip_systemid = { .lsi_prio = 0xffff },
 222         .lip_portid = { .lpi_prio = 0xffff },
 223         .lip_state = LACP_STATE_SYNC | LACP_STATE_AGGREGATION |
 224             LACP_STATE_COLLECTING | LACP_STATE_DISTRIBUTING,
 225 };
 226
 227 static const struct lacp_peerinfo lacp_partner_admin_strict = {
 228         .lip_systemid = { .lsi_prio = 0xffff },
 229         .lip_portid = { .lpi_prio = 0xffff },
 230         .lip_state = 0,
 231 };
 232
 233 static const lacp_timer_func_t lacp_timer_funcs[LACP_NTIMER] = {
 234         [LACP_TIMER_CURRENT_WHILE] = lacp_sm_rx_timer,
 235         [LACP_TIMER_PERIODIC] = lacp_sm_ptx_timer,
 236         [LACP_TIMER_WAIT_WHILE] = lacp_sm_mux_timer,
 237 };
 238
 239 struct mbuf *
 240 lacp_input(struct lagg_port *lgp, struct mbuf *m)
 241 {
 242         struct lacp_port *lp = LACP_PORT(lgp);
 243         uint8_t subtype;
 244
 245         if (m->m_pkthdr.len < sizeof(struct ether_header) + sizeof(subtype)) {
 246                 m_freem(m);
 247                 return (NULL);
 248         }
 249
 250         m_copydata(m, sizeof(struct ether_header), sizeof(subtype), &subtype);
 251         switch (subtype) {
 252                 case SLOWPROTOCOLS_SUBTYPE_LACP:
 253                         lacp_pdu_input(lp, m);
 254                         return (NULL);
 255
 256                 case SLOWPROTOCOLS_SUBTYPE_MARKER:
 257                         lacp_marker_input(lp, m);
 258                         return (NULL);
 259         }
 260
 261         /* Not a subtype we are interested in */
 262         return (m);
 263 }
 264
 265 /*
 266  * lacp_pdu_input: process lacpdu
 267  */
 268 static int
 269 lacp_pdu_input(struct lacp_port *lp, struct mbuf *m)
 270 {
 271         struct lacp_softc *lsc = lp->lp_lsc;
 272         struct lacpdu *du;
 273         int error = 0;
 274
 275         if (m->m_pkthdr.len != sizeof(*du)) {
 276                 goto bad;
 277         }
 278
 279         if ((m->m_flags & M_MCAST) == 0) {
 280                 goto bad;
 281         }
 282
 283         if (m->m_len < sizeof(*du)) {
 284                 m = m_pullup(m, sizeof(*du));
 285                 if (m == NULL) {
 286                         return (ENOMEM);
 287                 }
 288         }
 289
 290         du = mtod(m, struct lacpdu *);
 291
 292         if (memcmp(&du->ldu_eh.ether_dhost,
 293             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
 294                 goto bad;
 295         }
 296
 297         /*
 298          * ignore the version for compatibility with
 299          * the future protocol revisions.
 300          */
 301 #if 0
 302         if (du->ldu_sph.sph_version != 1) {
 303                 goto bad;
 304         }
 305 #endif
 306
 307         /*
 308          * ignore tlv types for compatibility with
 309          * the future protocol revisions.
 310          */
 311         if (tlv_check(du, sizeof(*du), &du->ldu_tlv_actor,
 312             lacp_info_tlv_template, FALSE)) {
 313                 goto bad;
 314         }
 315
 316         if (V_lacp_debug > 0) {
 317                 lacp_dprintf(lp, "lacpdu receive\n");
 318                 lacp_dump_lacpdu(du);
 319         }
 320
 321         if ((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_rx_test) {
 322                 LACP_TPRINTF((lp, "Dropping RX PDU\n"));
 323                 goto bad;
 324         }
 325
 326         LACP_LOCK(lsc);
 327         lacp_sm_rx(lp, du);
 328         LACP_UNLOCK(lsc);
 329
 330         m_freem(m);
 331         return (error);
 332
 333 bad:
 334         m_freem(m);
 335         return (EINVAL);
 336 }
 337
 338 static void
 339 lacp_fill_actorinfo(struct lacp_port *lp, struct lacp_peerinfo *info)
 340 {
 341         struct lagg_port *lgp = lp->lp_lagg;
 342         struct lagg_softc *sc = lgp->lp_softc;
 343
 344         info->lip_systemid.lsi_prio = htons(LACP_SYSTEM_PRIO);
 345         memcpy(&info->lip_systemid.lsi_mac,
 346             IF_LLADDR(sc->sc_ifp), ETHER_ADDR_LEN);
 347         info->lip_portid.lpi_prio = htons(LACP_PORT_PRIO);
 348         info->lip_portid.lpi_portno = htons(lp->lp_ifp->if_index);
 349         info->lip_state = lp->lp_state;
 350 }
 351
 352 static void
 353 lacp_fill_markerinfo(struct lacp_port *lp, struct lacp_markerinfo *info)
 354 {
 355         struct ifnet *ifp = lp->lp_ifp;
 356
 357         /* Fill in the port index and system id (encoded as the MAC) */
 358         info->mi_rq_port = htons(ifp->if_index);
 359         memcpy(&info->mi_rq_system, lp->lp_systemid.lsi_mac, ETHER_ADDR_LEN);
 360         info->mi_rq_xid = htonl(0);
 361 }
 362
 363 static int
 364 lacp_xmit_lacpdu(struct lacp_port *lp)
 365 {
 366         struct lagg_port *lgp = lp->lp_lagg;
 367         struct mbuf *m;
 368         struct lacpdu *du;
 369         int error;
 370
 371         LACP_LOCK_ASSERT(lp->lp_lsc);
 372
 373         m = m_gethdr(M_NOWAIT, MT_DATA);
 374         if (m == NULL) {
 375                 return (ENOMEM);
 376         }
 377         m->m_len = m->m_pkthdr.len = sizeof(*du);
 378
 379         du = mtod(m, struct lacpdu *);
 380         memset(du, 0, sizeof(*du));
 381
 382         memcpy(&du->ldu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 383             ETHER_ADDR_LEN);
 384         memcpy(&du->ldu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 385         du->ldu_eh.ether_type = htons(ETHERTYPE_SLOW);
 386
 387         du->ldu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_LACP;
 388         du->ldu_sph.sph_version = 1;
 389
 390         TLV_SET(&du->ldu_tlv_actor, LACP_TYPE_ACTORINFO, sizeof(du->ldu_actor));
 391         du->ldu_actor = lp->lp_actor;
 392
 393         TLV_SET(&du->ldu_tlv_partner, LACP_TYPE_PARTNERINFO,
 394             sizeof(du->ldu_partner));
 395         du->ldu_partner = lp->lp_partner;
 396
 397         TLV_SET(&du->ldu_tlv_collector, LACP_TYPE_COLLECTORINFO,
 398             sizeof(du->ldu_collector));
 399         du->ldu_collector.lci_maxdelay = 0;
 400
 401         if (V_lacp_debug > 0) {
 402                 lacp_dprintf(lp, "lacpdu transmit\n");
 403                 lacp_dump_lacpdu(du);
 404         }
 405
 406         m->m_flags |= M_MCAST;
 407
 408         /*
 409          * XXX should use higher priority queue.
 410          * otherwise network congestion can break aggregation.
 411          */
 412
 413         error = lagg_enqueue(lp->lp_ifp, m);
 414         return (error);
 415 }
 416
 417 static int
 418 lacp_xmit_marker(struct lacp_port *lp)
 419 {
 420         struct lagg_port *lgp = lp->lp_lagg;
 421         struct mbuf *m;
 422         struct markerdu *mdu;
 423         int error;
 424
 425         LACP_LOCK_ASSERT(lp->lp_lsc);
 426
 427         m = m_gethdr(M_NOWAIT, MT_DATA);
 428         if (m == NULL) {
 429                 return (ENOMEM);
 430         }
 431         m->m_len = m->m_pkthdr.len = sizeof(*mdu);
 432
 433         mdu = mtod(m, struct markerdu *);
 434         memset(mdu, 0, sizeof(*mdu));
 435
 436         memcpy(&mdu->mdu_eh.ether_dhost, ethermulticastaddr_slowprotocols,
 437             ETHER_ADDR_LEN);
 438         memcpy(&mdu->mdu_eh.ether_shost, lgp->lp_lladdr, ETHER_ADDR_LEN);
 439         mdu->mdu_eh.ether_type = htons(ETHERTYPE_SLOW);
 440
 441         mdu->mdu_sph.sph_subtype = SLOWPROTOCOLS_SUBTYPE_MARKER;
 442         mdu->mdu_sph.sph_version = 1;
 443
 444         /* Bump the transaction id and copy over the marker info */
 445         lp->lp_marker.mi_rq_xid = htonl(ntohl(lp->lp_marker.mi_rq_xid) + 1);
 446         TLV_SET(&mdu->mdu_tlv, MARKER_TYPE_INFO, sizeof(mdu->mdu_info));
 447         mdu->mdu_info = lp->lp_marker;
 448
 449         LACP_DPRINTF((lp, "marker transmit, port=%u, sys=%6D, id=%u\n",
 450             ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system, ":",
 451             ntohl(mdu->mdu_info.mi_rq_xid)));
 452
 453         m->m_flags |= M_MCAST;
 454         error = lagg_enqueue(lp->lp_ifp, m);
 455         return (error);
 456 }
 457
 458 void
 459 lacp_linkstate(struct lagg_port *lgp)
 460 {
 461         struct lacp_port *lp = LACP_PORT(lgp);
 462         struct lacp_softc *lsc = lp->lp_lsc;
 463         struct ifnet *ifp = lgp->lp_ifp;
 464         struct ifmediareq ifmr;
 465         int error = 0;
 466         u_int media;
 467         uint8_t old_state;
 468         uint16_t old_key;
 469
 470         bzero((char *)&ifmr, sizeof(ifmr));
 471         error = (*ifp->if_ioctl)(ifp, SIOCGIFXMEDIA, (caddr_t)&ifmr);
 472         if (error != 0) {
 473                 bzero((char *)&ifmr, sizeof(ifmr));
 474                 error = (*ifp->if_ioctl)(ifp, SIOCGIFMEDIA, (caddr_t)&ifmr);
 475         }
 476         if (error != 0)
 477                 return;
 478
 479         LACP_LOCK(lsc);
 480         media = ifmr.ifm_active;
 481         LACP_DPRINTF((lp, "media changed 0x%x -> 0x%x, ether = %d, fdx = %d, "
 482             "link = %d\n", lp->lp_media, media, IFM_TYPE(media) == IFM_ETHER,
 483             (media & IFM_FDX) != 0, ifp->if_link_state == LINK_STATE_UP));
 484         old_state = lp->lp_state;
 485         old_key = lp->lp_key;
 486
 487         lp->lp_media = media;
 488         /*
 489          * If the port is not an active full duplex Ethernet link then it can
 490          * not be aggregated.
 491          */
 492         if (IFM_TYPE(media) != IFM_ETHER || (media & IFM_FDX) == 0 ||
 493             ifp->if_link_state != LINK_STATE_UP) {
 494                 lacp_port_disable(lp);
 495         } else {
 496                 lacp_port_enable(lp);
 497         }
 498         lp->lp_key = lacp_compose_key(lp);
 499
 500         if (old_state != lp->lp_state || old_key != lp->lp_key) {
 501                 LACP_DPRINTF((lp, "-> UNSELECTED\n"));
 502                 lp->lp_selected = LACP_UNSELECTED;
 503         }
 504         LACP_UNLOCK(lsc);
 505 }
 506
 507 static void
 508 lacp_tick(void *arg)
 509 {
 510         struct lacp_softc *lsc = arg;
 511         struct lacp_port *lp;
 512
 513         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 514                 if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0)
 515                         continue;
 516
 517                 CURVNET_SET(lp->lp_ifp->if_vnet);
 518                 lacp_run_timers(lp);
 519
 520                 lacp_select(lp);
 521                 lacp_sm_mux(lp);
 522                 lacp_sm_tx(lp);
 523                 lacp_sm_ptx_tx_schedule(lp);
 524                 CURVNET_RESTORE();
 525         }
 526         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 527 }
 528
 529 int
 530 lacp_port_create(struct lagg_port *lgp)
 531 {
 532         struct lagg_softc *sc = lgp->lp_softc;
 533         struct lacp_softc *lsc = LACP_SOFTC(sc);
 534         struct lacp_port *lp;
 535         struct ifnet *ifp = lgp->lp_ifp;
 536         struct sockaddr_dl sdl;
 537         struct ifmultiaddr *rifma = NULL;
 538         int error;
 539
 540         link_init_sdl(ifp, (struct sockaddr *)&sdl, IFT_ETHER);
 541         sdl.sdl_alen = ETHER_ADDR_LEN;
 542
 543         bcopy(&ethermulticastaddr_slowprotocols,
 544             LLADDR(&sdl), ETHER_ADDR_LEN);
 545         error = if_addmulti(ifp, (struct sockaddr *)&sdl, &rifma);
 546         if (error) {
 547                 printf("%s: ADDMULTI failed on %s\n", __func__,
 548                     lgp->lp_ifp->if_xname);
 549                 return (error);
 550         }
 551
 552         lp = malloc(sizeof(struct lacp_port),
 553             M_DEVBUF, M_NOWAIT|M_ZERO);
 554         if (lp == NULL)
 555                 return (ENOMEM);
 556
 557         LACP_LOCK(lsc);
 558         lgp->lp_psc = lp;
 559         lp->lp_ifp = ifp;
 560         lp->lp_lagg = lgp;
 561         lp->lp_lsc = lsc;
 562         lp->lp_ifma = rifma;
 563
 564         LIST_INSERT_HEAD(&lsc->lsc_ports, lp, lp_next);
 565
 566         lacp_fill_actorinfo(lp, &lp->lp_actor);
 567         lacp_fill_markerinfo(lp, &lp->lp_marker);
 568         lp->lp_state = LACP_STATE_ACTIVITY;
 569         lp->lp_aggregator = NULL;
 570         lacp_sm_rx_set_expired(lp);
 571         LACP_UNLOCK(lsc);
 572         lacp_linkstate(lgp);
 573
 574         return (0);
 575 }
 576
 577 void
 578 lacp_port_destroy(struct lagg_port *lgp)
 579 {
 580         struct lacp_port *lp = LACP_PORT(lgp);
 581         struct lacp_softc *lsc = lp->lp_lsc;
 582         int i;
 583
 584         LACP_LOCK(lsc);
 585         for (i = 0; i < LACP_NTIMER; i++) {
 586                 LACP_TIMER_DISARM(lp, i);
 587         }
 588
 589         lacp_disable_collecting(lp);
 590         lacp_disable_distributing(lp);
 591         lacp_unselect(lp);
 592
 593         LIST_REMOVE(lp, lp_next);
 594         LACP_UNLOCK(lsc);
 595
 596         /* The address may have already been removed by if_purgemaddrs() */
 597         if (!lgp->lp_detaching)
 598                 if_delmulti_ifma(lp->lp_ifma);
 599
 600         free(lp, M_DEVBUF);
 601 }
 602
 603 void
 604 lacp_req(struct lagg_softc *sc, void *data)
 605 {
 606         struct lacp_opreq *req = (struct lacp_opreq *)data;
 607         struct lacp_softc *lsc = LACP_SOFTC(sc);
 608         struct lacp_aggregator *la;
 609
 610         bzero(req, sizeof(struct lacp_opreq));
 611
 612         /*
 613          * If the LACP softc is NULL, return with the opreq structure full of
 614          * zeros.  It is normal for the softc to be NULL while the lagg is
 615          * being destroyed.
 616          */
 617         if (NULL == lsc)
 618                 return;
 619
 620         la = lsc->lsc_active_aggregator;
 621         LACP_LOCK(lsc);
 622         if (la != NULL) {
 623                 req->actor_prio = ntohs(la->la_actor.lip_systemid.lsi_prio);
 624                 memcpy(&req->actor_mac, &la->la_actor.lip_systemid.lsi_mac,
 625                     ETHER_ADDR_LEN);
 626                 req->actor_key = ntohs(la->la_actor.lip_key);
 627                 req->actor_portprio = ntohs(la->la_actor.lip_portid.lpi_prio);
 628                 req->actor_portno = ntohs(la->la_actor.lip_portid.lpi_portno);
 629                 req->actor_state = la->la_actor.lip_state;
 630
 631                 req->partner_prio = ntohs(la->la_partner.lip_systemid.lsi_prio);
 632                 memcpy(&req->partner_mac, &la->la_partner.lip_systemid.lsi_mac,
 633                     ETHER_ADDR_LEN);
 634                 req->partner_key = ntohs(la->la_partner.lip_key);
 635                 req->partner_portprio = ntohs(la->la_partner.lip_portid.lpi_prio);
 636                 req->partner_portno = ntohs(la->la_partner.lip_portid.lpi_portno);
 637                 req->partner_state = la->la_partner.lip_state;
 638         }
 639         LACP_UNLOCK(lsc);
 640 }
 641
 642 void
 643 lacp_portreq(struct lagg_port *lgp, void *data)
 644 {
 645         struct lacp_opreq *req = (struct lacp_opreq *)data;
 646         struct lacp_port *lp = LACP_PORT(lgp);
 647         struct lacp_softc *lsc = lp->lp_lsc;
 648
 649         LACP_LOCK(lsc);
 650         req->actor_prio = ntohs(lp->lp_actor.lip_systemid.lsi_prio);
 651         memcpy(&req->actor_mac, &lp->lp_actor.lip_systemid.lsi_mac,
 652             ETHER_ADDR_LEN);
 653         req->actor_key = ntohs(lp->lp_actor.lip_key);
 654         req->actor_portprio = ntohs(lp->lp_actor.lip_portid.lpi_prio);
 655         req->actor_portno = ntohs(lp->lp_actor.lip_portid.lpi_portno);
 656         req->actor_state = lp->lp_actor.lip_state;
 657
 658         req->partner_prio = ntohs(lp->lp_partner.lip_systemid.lsi_prio);
 659         memcpy(&req->partner_mac, &lp->lp_partner.lip_systemid.lsi_mac,
 660             ETHER_ADDR_LEN);
 661         req->partner_key = ntohs(lp->lp_partner.lip_key);
 662         req->partner_portprio = ntohs(lp->lp_partner.lip_portid.lpi_prio);
 663         req->partner_portno = ntohs(lp->lp_partner.lip_portid.lpi_portno);
 664         req->partner_state = lp->lp_partner.lip_state;
 665         LACP_UNLOCK(lsc);
 666 }
 667
 668 static void
 669 lacp_disable_collecting(struct lacp_port *lp)
 670 {
 671         LACP_DPRINTF((lp, "collecting disabled\n"));
 672         lp->lp_state &= ~LACP_STATE_COLLECTING;
 673 }
 674
 675 static void
 676 lacp_enable_collecting(struct lacp_port *lp)
 677 {
 678         LACP_DPRINTF((lp, "collecting enabled\n"));
 679         lp->lp_state |= LACP_STATE_COLLECTING;
 680 }
 681
 682 static void
 683 lacp_disable_distributing(struct lacp_port *lp)
 684 {
 685         struct lacp_aggregator *la = lp->lp_aggregator;
 686         struct lacp_softc *lsc = lp->lp_lsc;
 687         struct lagg_softc *sc = lsc->lsc_softc;
 688         char buf[LACP_LAGIDSTR_MAX+1];
 689
 690         LACP_LOCK_ASSERT(lsc);
 691
 692         if (la == NULL || (lp->lp_state & LACP_STATE_DISTRIBUTING) == 0) {
 693                 return;
 694         }
 695
 696         KASSERT(!TAILQ_EMPTY(&la->la_ports), ("no aggregator ports"));
 697         KASSERT(la->la_nports > 0, ("nports invalid (%d)", la->la_nports));
 698         KASSERT(la->la_refcnt >= la->la_nports, ("aggregator refcnt invalid"));
 699
 700         LACP_DPRINTF((lp, "disable distributing on aggregator %s, "
 701             "nports %d -> %d\n",
 702             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 703             la->la_nports, la->la_nports - 1));
 704
 705         TAILQ_REMOVE(&la->la_ports, lp, lp_dist_q);
 706         la->la_nports--;
 707         sc->sc_active = la->la_nports;
 708
 709         if (lsc->lsc_active_aggregator == la) {
 710                 lacp_suppress_distributing(lsc, la);
 711                 lacp_select_active_aggregator(lsc);
 712                 /* regenerate the port map, the active aggregator has changed */
 713                 lacp_update_portmap(lsc);
 714         }
 715
 716         lp->lp_state &= ~LACP_STATE_DISTRIBUTING;
 717         if_link_state_change(sc->sc_ifp,
 718             sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
 719 }
 720
 721 static void
 722 lacp_enable_distributing(struct lacp_port *lp)
 723 {
 724         struct lacp_aggregator *la = lp->lp_aggregator;
 725         struct lacp_softc *lsc = lp->lp_lsc;
 726         struct lagg_softc *sc = lsc->lsc_softc;
 727         char buf[LACP_LAGIDSTR_MAX+1];
 728
 729         LACP_LOCK_ASSERT(lsc);
 730
 731         if ((lp->lp_state & LACP_STATE_DISTRIBUTING) != 0) {
 732                 return;
 733         }
 734
 735         LACP_DPRINTF((lp, "enable distributing on aggregator %s, "
 736             "nports %d -> %d\n",
 737             lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 738             la->la_nports, la->la_nports + 1));
 739
 740         KASSERT(la->la_refcnt > la->la_nports, ("aggregator refcnt invalid"));
 741         TAILQ_INSERT_HEAD(&la->la_ports, lp, lp_dist_q);
 742         la->la_nports++;
 743         sc->sc_active = la->la_nports;
 744
 745         lp->lp_state |= LACP_STATE_DISTRIBUTING;
 746
 747         if (lsc->lsc_active_aggregator == la) {
 748                 lacp_suppress_distributing(lsc, la);
 749                 lacp_update_portmap(lsc);
 750         } else
 751                 /* try to become the active aggregator */
 752                 lacp_select_active_aggregator(lsc);
 753
 754         if_link_state_change(sc->sc_ifp,
 755             sc->sc_active ? LINK_STATE_UP : LINK_STATE_DOWN);
 756 }
 757
 758 static void
 759 lacp_transit_expire(void *vp)
 760 {
 761         struct lacp_softc *lsc = vp;
 762
 763         LACP_LOCK_ASSERT(lsc);
 764
 765         CURVNET_SET(lsc->lsc_softc->sc_ifp->if_vnet);
 766         LACP_TRACE(NULL);
 767         CURVNET_RESTORE();
 768
 769         lsc->lsc_suppress_distributing = FALSE;
 770 }
 771
 772 void
 773 lacp_attach(struct lagg_softc *sc)
 774 {
 775         struct lacp_softc *lsc;
 776
 777         lsc = malloc(sizeof(struct lacp_softc), M_DEVBUF, M_WAITOK | M_ZERO);
 778
 779         sc->sc_psc = lsc;
 780         lsc->lsc_softc = sc;
 781
 782         lsc->lsc_hashkey = m_ether_tcpip_hash_init();
 783         lsc->lsc_active_aggregator = NULL;
 784         lsc->lsc_strict_mode = VNET(lacp_default_strict_mode);
 785         LACP_LOCK_INIT(lsc);
 786         TAILQ_INIT(&lsc->lsc_aggregators);
 787         LIST_INIT(&lsc->lsc_ports);
 788
 789         callout_init_mtx(&lsc->lsc_transit_callout, &lsc->lsc_mtx, 0);
 790         callout_init_mtx(&lsc->lsc_callout, &lsc->lsc_mtx, 0);
 791
 792         /* if the lagg is already up then do the same */
 793         if (sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING)
 794                 lacp_init(sc);
 795 }
 796
 797 void
 798 lacp_detach(void *psc)
 799 {
 800         struct lacp_softc *lsc = (struct lacp_softc *)psc;
 801
 802         KASSERT(TAILQ_EMPTY(&lsc->lsc_aggregators),
 803             ("aggregators still active"));
 804         KASSERT(lsc->lsc_active_aggregator == NULL,
 805             ("aggregator still attached"));
 806
 807         callout_drain(&lsc->lsc_transit_callout);
 808         callout_drain(&lsc->lsc_callout);
 809
 810         LACP_LOCK_DESTROY(lsc);
 811         free(lsc, M_DEVBUF);
 812 }
 813
 814 void
 815 lacp_init(struct lagg_softc *sc)
 816 {
 817         struct lacp_softc *lsc = LACP_SOFTC(sc);
 818
 819         LACP_LOCK(lsc);
 820         callout_reset(&lsc->lsc_callout, hz, lacp_tick, lsc);
 821         LACP_UNLOCK(lsc);
 822 }
 823
 824 void
 825 lacp_stop(struct lagg_softc *sc)
 826 {
 827         struct lacp_softc *lsc = LACP_SOFTC(sc);
 828
 829         LACP_LOCK(lsc);
 830         callout_stop(&lsc->lsc_transit_callout);
 831         callout_stop(&lsc->lsc_callout);
 832         LACP_UNLOCK(lsc);
 833 }
 834
 835 struct lagg_port *
 836 lacp_select_tx_port_by_hash(struct lagg_softc *sc, uint32_t hash,
 837     uint8_t numa_domain, int *err)
 838 {
 839         struct lacp_softc *lsc = LACP_SOFTC(sc);
 840         struct lacp_portmap *pm;
 841         struct lacp_port *lp;
 842         struct lacp_port **map;
 843         int count;
 844
 845         if (__predict_false(lsc->lsc_suppress_distributing)) {
 846                 LACP_DPRINTF((NULL, "%s: waiting transit\n", __func__));
 847                 *err = ENOBUFS;
 848                 return (NULL);
 849         }
 850
 851         pm = &lsc->lsc_pmap[lsc->lsc_activemap];
 852         if (pm->pm_count == 0) {
 853                 LACP_DPRINTF((NULL, "%s: no active aggregator\n", __func__));
 854                 *err = ENETDOWN;
 855                 return (NULL);
 856         }
 857
 858 #ifdef NUMA
 859         if ((sc->sc_opts & LAGG_OPT_USE_NUMA) &&
 860             pm->pm_num_dom > 1 && numa_domain < MAXMEMDOM) {
 861                 count = pm->pm_numa[numa_domain].count;
 862                 if (count > 0) {
 863                         map = pm->pm_numa[numa_domain].map;
 864                 } else {
 865                         /* No ports on this domain; use global hash. */
 866                         map = pm->pm_map;
 867                         count = pm->pm_count;
 868                 }
 869         } else
 870 #endif
 871         {
 872                 map = pm->pm_map;
 873                 count = pm->pm_count;
 874         }
 875
 876         hash %= count;
 877         lp = map[hash];
 878
 879         return (lp->lp_lagg);
 880 }
 881
 882 struct lagg_port *
 883 lacp_select_tx_port(struct lagg_softc *sc, struct mbuf *m, int *err)
 884 {
 885         struct lacp_softc *lsc = LACP_SOFTC(sc);
 886         uint32_t hash;
 887         uint8_t numa_domain;
 888
 889         if ((sc->sc_opts & LAGG_OPT_USE_FLOWID) &&
 890             M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
 891                 hash = m->m_pkthdr.flowid >> sc->flowid_shift;
 892         else
 893                 hash = m_ether_tcpip_hash(sc->sc_flags, m, lsc->lsc_hashkey);
 894
 895         numa_domain = m->m_pkthdr.numa_domain;
 896         return (lacp_select_tx_port_by_hash(sc, hash, numa_domain, err));
 897 }
 898
 899 /*
 900  * lacp_suppress_distributing: drop transmit packets for a while
 901  * to preserve packet ordering.
 902  */
 903
 904 static void
 905 lacp_suppress_distributing(struct lacp_softc *lsc, struct lacp_aggregator *la)
 906 {
 907         struct lacp_port *lp;
 908
 909         if (lsc->lsc_active_aggregator != la) {
 910                 return;
 911         }
 912
 913         LACP_TRACE(NULL);
 914
 915         lsc->lsc_suppress_distributing = TRUE;
 916
 917         /* send a marker frame down each port to verify the queues are empty */
 918         LIST_FOREACH(lp, &lsc->lsc_ports, lp_next) {
 919                 lp->lp_flags |= LACP_PORT_MARK;
 920                 if (lacp_xmit_marker(lp) != 0)
 921                         lp->lp_flags &= ~LACP_PORT_MARK;
 922         }
 923
 924         /* set a timeout for the marker frames */
 925         callout_reset(&lsc->lsc_transit_callout,
 926             LACP_TRANSIT_DELAY * hz / 1000, lacp_transit_expire, lsc);
 927 }
 928
 929 static int
 930 lacp_compare_peerinfo(const struct lacp_peerinfo *a,
 931     const struct lacp_peerinfo *b)
 932 {
 933         return (memcmp(a, b, offsetof(struct lacp_peerinfo, lip_state)));
 934 }
 935
 936 static int
 937 lacp_compare_systemid(const struct lacp_systemid *a,
 938     const struct lacp_systemid *b)
 939 {
 940         return (memcmp(a, b, sizeof(*a)));
 941 }
 942
 943 #if 0   /* unused */
 944 static int
 945 lacp_compare_portid(const struct lacp_portid *a,
 946     const struct lacp_portid *b)
 947 {
 948         return (memcmp(a, b, sizeof(*a)));
 949 }
 950 #endif
 951
 952 static uint64_t
 953 lacp_aggregator_bandwidth(struct lacp_aggregator *la)
 954 {
 955         struct lacp_port *lp;
 956         uint64_t speed;
 957
 958         lp = TAILQ_FIRST(&la->la_ports);
 959         if (lp == NULL) {
 960                 return (0);
 961         }
 962
 963         speed = ifmedia_baudrate(lp->lp_media);
 964         speed *= la->la_nports;
 965         if (speed == 0) {
 966                 LACP_DPRINTF((lp, "speed 0? media=0x%x nports=%d\n",
 967                     lp->lp_media, la->la_nports));
 968         }
 969
 970         return (speed);
 971 }
 972
 973 /*
 974  * lacp_select_active_aggregator: select an aggregator to be used to transmit
 975  * packets from lagg(4) interface.
 976  */
 977
 978 static void
 979 lacp_select_active_aggregator(struct lacp_softc *lsc)
 980 {
 981         struct lacp_aggregator *la;
 982         struct lacp_aggregator *best_la = NULL;
 983         uint64_t best_speed = 0;
 984         char buf[LACP_LAGIDSTR_MAX+1];
 985
 986         LACP_TRACE(NULL);
 987
 988         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
 989                 uint64_t speed;
 990
 991                 if (la->la_nports == 0) {
 992                         continue;
 993                 }
 994
 995                 speed = lacp_aggregator_bandwidth(la);
 996                 LACP_DPRINTF((NULL, "%s, speed=%jd, nports=%d\n",
 997                     lacp_format_lagid_aggregator(la, buf, sizeof(buf)),
 998                     speed, la->la_nports));
 999
1000                 /*
1001                  * This aggregator is chosen if the partner has a better
1002                  * system priority or, the total aggregated speed is higher
1003                  * or, it is already the chosen aggregator
1004                  */
1005                 if ((best_la != NULL && LACP_SYS_PRI(la->la_partner) <
1006                     LACP_SYS_PRI(best_la->la_partner)) ||
1007                     speed > best_speed ||
1008                     (speed == best_speed &&
1009                     la == lsc->lsc_active_aggregator)) {
1010                         best_la = la;
1011                         best_speed = speed;
1012                 }
1013         }
1014
1015         KASSERT(best_la == NULL || best_la->la_nports > 0,
1016             ("invalid aggregator refcnt"));
1017         KASSERT(best_la == NULL || !TAILQ_EMPTY(&best_la->la_ports),
1018             ("invalid aggregator list"));
1019
1020         if (lsc->lsc_active_aggregator != best_la) {
1021                 LACP_DPRINTF((NULL, "active aggregator changed\n"));
1022                 LACP_DPRINTF((NULL, "old %s\n",
1023                     lacp_format_lagid_aggregator(lsc->lsc_active_aggregator,
1024                     buf, sizeof(buf))));
1025         } else {
1026                 LACP_DPRINTF((NULL, "active aggregator not changed\n"));
1027         }
1028         LACP_DPRINTF((NULL, "new %s\n",
1029             lacp_format_lagid_aggregator(best_la, buf, sizeof(buf))));
1030
1031         if (lsc->lsc_active_aggregator != best_la) {
1032                 lsc->lsc_active_aggregator = best_la;
1033                 lacp_update_portmap(lsc);
1034                 if (best_la) {
1035                         lacp_suppress_distributing(lsc, best_la);
1036                 }
1037         }
1038 }
1039
1040 /*
1041  * Updated the inactive portmap array with the new list of ports and
1042  * make it live.
1043  */
1044 static void
1045 lacp_update_portmap(struct lacp_softc *lsc)
1046 {
1047         struct lagg_softc *sc = lsc->lsc_softc;
1048         struct lacp_aggregator *la;
1049         struct lacp_portmap *p;
1050         struct lacp_port *lp;
1051         uint64_t speed;
1052         u_int newmap;
1053         int i;
1054 #ifdef NUMA
1055         int count;
1056         uint8_t domain;
1057 #endif
1058
1059         newmap = lsc->lsc_activemap == 0 ? 1 : 0;
1060         p = &lsc->lsc_pmap[newmap];
1061         la = lsc->lsc_active_aggregator;
1062         speed = 0;
1063         bzero(p, sizeof(struct lacp_portmap));
1064
1065         if (la != NULL && la->la_nports > 0) {
1066                 p->pm_count = la->la_nports;
1067                 i = 0;
1068                 TAILQ_FOREACH(lp, &la->la_ports, lp_dist_q) {
1069                         p->pm_map[i++] = lp;
1070 #ifdef NUMA
1071                         domain = lp->lp_ifp->if_numa_domain;
1072                         if (domain >= MAXMEMDOM)
1073                                 continue;
1074                         count = p->pm_numa[domain].count;
1075                         p->pm_numa[domain].map[count] = lp;
1076                         p->pm_numa[domain].count++;
1077 #endif
1078                 }
1079                 KASSERT(i == p->pm_count, ("Invalid port count"));
1080
1081 #ifdef NUMA
1082                 for (i = 0; i < MAXMEMDOM; i++) {
1083                         if (p->pm_numa[i].count != 0)
1084                                 p->pm_num_dom++;
1085                 }
1086 #endif
1087                 speed = lacp_aggregator_bandwidth(la);
1088         }
1089         sc->sc_ifp->if_baudrate = speed;
1090         EVENTHANDLER_INVOKE(ifnet_event, sc->sc_ifp,
1091             IFNET_EVENT_UPDATE_BAUDRATE);
1092
1093         /* switch the active portmap over */
1094         atomic_store_rel_int(&lsc->lsc_activemap, newmap);
1095         LACP_DPRINTF((NULL, "Set table %d with %d ports\n",
1096                     lsc->lsc_activemap,
1097                     lsc->lsc_pmap[lsc->lsc_activemap].pm_count));
1098 }
1099
1100 static uint16_t
1101 lacp_compose_key(struct lacp_port *lp)
1102 {
1103         struct lagg_port *lgp = lp->lp_lagg;
1104         struct lagg_softc *sc = lgp->lp_softc;
1105         u_int media = lp->lp_media;
1106         uint16_t key;
1107
1108         if ((lp->lp_state & LACP_STATE_AGGREGATION) == 0) {
1109                 /*
1110                  * non-aggregatable links should have unique keys.
1111                  *
1112                  * XXX this isn't really unique as if_index is 16 bit.
1113                  */
1114
1115                 /* bit 0..14:   (some bits of) if_index of this port */
1116                 key = lp->lp_ifp->if_index;
1117                 /* bit 15:      1 */
1118                 key |= 0x8000;
1119         } else {
1120                 u_int subtype = IFM_SUBTYPE(media);
1121
1122                 KASSERT(IFM_TYPE(media) == IFM_ETHER, ("invalid media type"));
1123                 KASSERT((media & IFM_FDX) != 0, ("aggregating HDX interface"));
1124
1125                 /* bit 0..4:    IFM_SUBTYPE modulo speed */
1126                 switch (subtype) {
1127                 case IFM_10_T:
1128                 case IFM_10_2:
1129                 case IFM_10_5:
1130                 case IFM_10_STP:
1131                 case IFM_10_FL:
1132                         key = IFM_10_T;
1133                         break;
1134                 case IFM_100_TX:
1135                 case IFM_100_FX:
1136                 case IFM_100_T4:
1137                 case IFM_100_VG:
1138                 case IFM_100_T2:
1139                 case IFM_100_T:
1140                 case IFM_100_SGMII:
1141                         key = IFM_100_TX;
1142                         break;
1143                 case IFM_1000_SX:
1144                 case IFM_1000_LX:
1145                 case IFM_1000_CX:
1146                 case IFM_1000_T:
1147                 case IFM_1000_KX:
1148                 case IFM_1000_SGMII:
1149                 case IFM_1000_CX_SGMII:
1150                         key = IFM_1000_SX;
1151                         break;
1152                 case IFM_10G_LR:
1153                 case IFM_10G_SR:
1154                 case IFM_10G_CX4:
1155                 case IFM_10G_TWINAX:
1156                 case IFM_10G_TWINAX_LONG:
1157                 case IFM_10G_LRM:
1158                 case IFM_10G_T:
1159                 case IFM_10G_KX4:
1160                 case IFM_10G_KR:
1161                 case IFM_10G_CR1:
1162                 case IFM_10G_ER:
1163                 case IFM_10G_SFI:
1164                 case IFM_10G_AOC:
1165                         key = IFM_10G_LR;
1166                         break;
1167                 case IFM_20G_KR2:
1168                         key = IFM_20G_KR2;
1169                         break;
1170                 case IFM_2500_KX:
1171                 case IFM_2500_T:
1172                 case IFM_2500_X:
1173                         key = IFM_2500_KX;
1174                         break;
1175                 case IFM_5000_T:
1176                 case IFM_5000_KR:
1177                 case IFM_5000_KR_S:
1178                 case IFM_5000_KR1:
1179                         key = IFM_5000_T;
1180                         break;
1181                 case IFM_50G_PCIE:
1182                 case IFM_50G_CR2:
1183                 case IFM_50G_KR2:
1184                 case IFM_50G_KR4:
1185                 case IFM_50G_SR2:
1186                 case IFM_50G_LR2:
1187                 case IFM_50G_LAUI2_AC:
1188                 case IFM_50G_LAUI2:
1189                 case IFM_50G_AUI2_AC:
1190                 case IFM_50G_AUI2:
1191                 case IFM_50G_CP:
1192                 case IFM_50G_SR:
1193                 case IFM_50G_LR:
1194                 case IFM_50G_FR:
1195                 case IFM_50G_KR_PAM4:
1196                 case IFM_50G_AUI1_AC:
1197                 case IFM_50G_AUI1:
1198                         key = IFM_50G_PCIE;
1199                         break;
1200                 case IFM_56G_R4:
1201                         key = IFM_56G_R4;
1202                         break;
1203                 case IFM_25G_PCIE:
1204                 case IFM_25G_CR:
1205                 case IFM_25G_KR:
1206                 case IFM_25G_SR:
1207                 case IFM_25G_LR:
1208                 case IFM_25G_ACC:
1209                 case IFM_25G_AOC:
1210                 case IFM_25G_T:
1211                 case IFM_25G_CR_S:
1212                 case IFM_25G_CR1:
1213                 case IFM_25G_KR_S:
1214                 case IFM_25G_AUI:
1215                 case IFM_25G_KR1:
1216                         key = IFM_25G_PCIE;
1217                         break;
1218                 case IFM_40G_CR4:
1219                 case IFM_40G_SR4:
1220                 case IFM_40G_LR4:
1221                 case IFM_40G_LM4:
1222                 case IFM_40G_XLPPI:
1223                 case IFM_40G_KR4:
1224                 case IFM_40G_XLAUI:
1225                 case IFM_40G_XLAUI_AC:
1226                 case IFM_40G_ER4:
1227                         key = IFM_40G_CR4;
1228                         break;
1229                 case IFM_100G_CR4:
1230                 case IFM_100G_SR4:
1231                 case IFM_100G_KR4:
1232                 case IFM_100G_LR4:
1233                 case IFM_100G_CAUI4_AC:
1234                 case IFM_100G_CAUI4:
1235                 case IFM_100G_AUI4_AC:
1236                 case IFM_100G_AUI4:
1237                 case IFM_100G_CR_PAM4:
1238                 case IFM_100G_KR_PAM4:
1239                 case IFM_100G_CP2:
1240                 case IFM_100G_SR2:
1241                 case IFM_100G_DR:
1242                 case IFM_100G_KR2_PAM4:
1243                 case IFM_100G_CAUI2_AC:
1244                 case IFM_100G_CAUI2:
1245                 case IFM_100G_AUI2_AC:
1246                 case IFM_100G_AUI2:
1247                         key = IFM_100G_CR4;
1248                         break;
1249                 case IFM_200G_CR4_PAM4:
1250                 case IFM_200G_SR4:
1251                 case IFM_200G_FR4:
1252                 case IFM_200G_LR4:
1253                 case IFM_200G_DR4:
1254                 case IFM_200G_KR4_PAM4:
1255                 case IFM_200G_AUI4_AC:
1256                 case IFM_200G_AUI4:
1257                 case IFM_200G_AUI8_AC:
1258                 case IFM_200G_AUI8:
1259                         key = IFM_200G_CR4_PAM4;
1260                         break;
1261                 case IFM_400G_FR8:
1262                 case IFM_400G_LR8:
1263                 case IFM_400G_DR4:
1264                 case IFM_400G_AUI8_AC:
1265                 case IFM_400G_AUI8:
1266                         key = IFM_400G_FR8;
1267                         break;
1268                 default:
1269                         key = subtype;
1270                         break;
1271                 }
1272                 /* bit 5..14:   (some bits of) if_index of lagg device */
1273                 key |= 0x7fe0 & ((sc->sc_ifp->if_index) << 5);
1274                 /* bit 15:      0 */
1275         }
1276         return (htons(key));
1277 }
1278
1279 static void
1280 lacp_aggregator_addref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1281 {
1282         char buf[LACP_LAGIDSTR_MAX+1];
1283
1284         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1285             __func__,
1286             lacp_format_lagid(&la->la_actor, &la->la_partner,
1287             buf, sizeof(buf)),
1288             la->la_refcnt, la->la_refcnt + 1));
1289
1290         KASSERT(la->la_refcnt > 0, ("refcount <= 0"));
1291         la->la_refcnt++;
1292         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcount"));
1293 }
1294
1295 static void
1296 lacp_aggregator_delref(struct lacp_softc *lsc, struct lacp_aggregator *la)
1297 {
1298         char buf[LACP_LAGIDSTR_MAX+1];
1299
1300         LACP_DPRINTF((NULL, "%s: lagid=%s, refcnt %d -> %d\n",
1301             __func__,
1302             lacp_format_lagid(&la->la_actor, &la->la_partner,
1303             buf, sizeof(buf)),
1304             la->la_refcnt, la->la_refcnt - 1));
1305
1306         KASSERT(la->la_refcnt > la->la_nports, ("invalid refcnt"));
1307         la->la_refcnt--;
1308         if (la->la_refcnt > 0) {
1309                 return;
1310         }
1311
1312         KASSERT(la->la_refcnt == 0, ("refcount not zero"));
1313         KASSERT(lsc->lsc_active_aggregator != la, ("aggregator active"));
1314
1315         TAILQ_REMOVE(&lsc->lsc_aggregators, la, la_q);
1316
1317         free(la, M_DEVBUF);
1318 }
1319
1320 /*
1321  * lacp_aggregator_get: allocate an aggregator.
1322  */
1323
1324 static struct lacp_aggregator *
1325 lacp_aggregator_get(struct lacp_softc *lsc, struct lacp_port *lp)
1326 {
1327         struct lacp_aggregator *la;
1328
1329         la = malloc(sizeof(*la), M_DEVBUF, M_NOWAIT);
1330         if (la) {
1331                 la->la_refcnt = 1;
1332                 la->la_nports = 0;
1333                 TAILQ_INIT(&la->la_ports);
1334                 la->la_pending = 0;
1335                 TAILQ_INSERT_TAIL(&lsc->lsc_aggregators, la, la_q);
1336         }
1337
1338         return (la);
1339 }
1340
1341 /*
1342  * lacp_fill_aggregator_id: setup a newly allocated aggregator from a port.
1343  */
1344
1345 static void
1346 lacp_fill_aggregator_id(struct lacp_aggregator *la, const struct lacp_port *lp)
1347 {
1348         lacp_fill_aggregator_id_peer(&la->la_partner, &lp->lp_partner);
1349         lacp_fill_aggregator_id_peer(&la->la_actor, &lp->lp_actor);
1350
1351         la->la_actor.lip_state = lp->lp_state & LACP_STATE_AGGREGATION;
1352 }
1353
1354 static void
1355 lacp_fill_aggregator_id_peer(struct lacp_peerinfo *lpi_aggr,
1356     const struct lacp_peerinfo *lpi_port)
1357 {
1358         memset(lpi_aggr, 0, sizeof(*lpi_aggr));
1359         lpi_aggr->lip_systemid = lpi_port->lip_systemid;
1360         lpi_aggr->lip_key = lpi_port->lip_key;
1361 }
1362
1363 /*
1364  * lacp_aggregator_is_compatible: check if a port can join to an aggregator.
1365  */
1366
1367 static int
1368 lacp_aggregator_is_compatible(const struct lacp_aggregator *la,
1369     const struct lacp_port *lp)
1370 {
1371         if (!(lp->lp_state & LACP_STATE_AGGREGATION) ||
1372             !(lp->lp_partner.lip_state & LACP_STATE_AGGREGATION)) {
1373                 return (0);
1374         }
1375
1376         if (!(la->la_actor.lip_state & LACP_STATE_AGGREGATION)) {
1377                 return (0);
1378         }
1379
1380         if (!lacp_peerinfo_is_compatible(&la->la_partner, &lp->lp_partner)) {
1381                 return (0);
1382         }
1383
1384         if (!lacp_peerinfo_is_compatible(&la->la_actor, &lp->lp_actor)) {
1385                 return (0);
1386         }
1387
1388         return (1);
1389 }
1390
1391 static int
1392 lacp_peerinfo_is_compatible(const struct lacp_peerinfo *a,
1393     const struct lacp_peerinfo *b)
1394 {
1395         if (memcmp(&a->lip_systemid, &b->lip_systemid,
1396             sizeof(a->lip_systemid))) {
1397                 return (0);
1398         }
1399
1400         if (memcmp(&a->lip_key, &b->lip_key, sizeof(a->lip_key))) {
1401                 return (0);
1402         }
1403
1404         return (1);
1405 }
1406
1407 static void
1408 lacp_port_enable(struct lacp_port *lp)
1409 {
1410         lp->lp_state |= LACP_STATE_AGGREGATION;
1411 }
1412
1413 static void
1414 lacp_port_disable(struct lacp_port *lp)
1415 {
1416         lacp_set_mux(lp, LACP_MUX_DETACHED);
1417
1418         lp->lp_state &= ~LACP_STATE_AGGREGATION;
1419         lp->lp_selected = LACP_UNSELECTED;
1420         lacp_sm_rx_record_default(lp);
1421         lp->lp_partner.lip_state &= ~LACP_STATE_AGGREGATION;
1422         lp->lp_state &= ~LACP_STATE_EXPIRED;
1423 }
1424
1425 /*
1426  * lacp_select: select an aggregator.  create one if necessary.
1427  */
1428 static void
1429 lacp_select(struct lacp_port *lp)
1430 {
1431         struct lacp_softc *lsc = lp->lp_lsc;
1432         struct lacp_aggregator *la;
1433         char buf[LACP_LAGIDSTR_MAX+1];
1434
1435         if (lp->lp_aggregator) {
1436                 return;
1437         }
1438
1439         /* If we haven't heard from our peer, skip this step. */
1440         if (lp->lp_state & LACP_STATE_DEFAULTED)
1441                 return;
1442
1443         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1444             ("timer_wait_while still active"));
1445
1446         LACP_DPRINTF((lp, "port lagid=%s\n",
1447             lacp_format_lagid(&lp->lp_actor, &lp->lp_partner,
1448             buf, sizeof(buf))));
1449
1450         TAILQ_FOREACH(la, &lsc->lsc_aggregators, la_q) {
1451                 if (lacp_aggregator_is_compatible(la, lp)) {
1452                         break;
1453                 }
1454         }
1455
1456         if (la == NULL) {
1457                 la = lacp_aggregator_get(lsc, lp);
1458                 if (la == NULL) {
1459                         LACP_DPRINTF((lp, "aggregator creation failed\n"));
1460
1461                         /*
1462                          * will retry on the next tick.
1463                          */
1464
1465                         return;
1466                 }
1467                 lacp_fill_aggregator_id(la, lp);
1468                 LACP_DPRINTF((lp, "aggregator created\n"));
1469         } else {
1470                 LACP_DPRINTF((lp, "compatible aggregator found\n"));
1471                 if (la->la_refcnt == LACP_MAX_PORTS)
1472                         return;
1473                 lacp_aggregator_addref(lsc, la);
1474         }
1475
1476         LACP_DPRINTF((lp, "aggregator lagid=%s\n",
1477             lacp_format_lagid(&la->la_actor, &la->la_partner,
1478             buf, sizeof(buf))));
1479
1480         lp->lp_aggregator = la;
1481         lp->lp_selected = LACP_SELECTED;
1482 }
1483
1484 /*
1485  * lacp_unselect: finish unselect/detach process.
1486  */
1487
1488 static void
1489 lacp_unselect(struct lacp_port *lp)
1490 {
1491         struct lacp_softc *lsc = lp->lp_lsc;
1492         struct lacp_aggregator *la = lp->lp_aggregator;
1493
1494         KASSERT(!LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1495             ("timer_wait_while still active"));
1496
1497         if (la == NULL) {
1498                 return;
1499         }
1500
1501         lp->lp_aggregator = NULL;
1502         lacp_aggregator_delref(lsc, la);
1503 }
1504
1505 /* mux machine */
1506
1507 static void
1508 lacp_sm_mux(struct lacp_port *lp)
1509 {
1510         struct lagg_port *lgp = lp->lp_lagg;
1511         struct lagg_softc *sc = lgp->lp_softc;
1512         enum lacp_mux_state new_state;
1513         boolean_t p_sync =
1514                     (lp->lp_partner.lip_state & LACP_STATE_SYNC) != 0;
1515         boolean_t p_collecting =
1516             (lp->lp_partner.lip_state & LACP_STATE_COLLECTING) != 0;
1517         enum lacp_selected selected = lp->lp_selected;
1518         struct lacp_aggregator *la;
1519
1520         if (V_lacp_debug > 1)
1521                 lacp_dprintf(lp, "%s: state= 0x%x, selected= 0x%x, "
1522                     "p_sync= 0x%x, p_collecting= 0x%x\n", __func__,
1523                     lp->lp_mux_state, selected, p_sync, p_collecting);
1524
1525 re_eval:
1526         la = lp->lp_aggregator;
1527         KASSERT(lp->lp_mux_state == LACP_MUX_DETACHED || la != NULL,
1528             ("MUX not detached"));
1529         new_state = lp->lp_mux_state;
1530         switch (lp->lp_mux_state) {
1531         case LACP_MUX_DETACHED:
1532                 if (selected != LACP_UNSELECTED) {
1533                         new_state = LACP_MUX_WAITING;
1534                 }
1535                 break;
1536         case LACP_MUX_WAITING:
1537                 KASSERT(la->la_pending > 0 ||
1538                     !LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE),
1539                     ("timer_wait_while still active"));
1540                 if (selected == LACP_SELECTED && la->la_pending == 0) {
1541                         new_state = LACP_MUX_ATTACHED;
1542                 } else if (selected == LACP_UNSELECTED) {
1543                         new_state = LACP_MUX_DETACHED;
1544                 }
1545                 break;
1546         case LACP_MUX_ATTACHED:
1547                 if (selected == LACP_SELECTED && p_sync) {
1548                         new_state = LACP_MUX_COLLECTING;
1549                 } else if (selected != LACP_SELECTED) {
1550                         new_state = LACP_MUX_DETACHED;
1551                 }
1552                 break;
1553         case LACP_MUX_COLLECTING:
1554                 if (selected == LACP_SELECTED && p_sync && p_collecting) {
1555                         new_state = LACP_MUX_DISTRIBUTING;
1556                 } else if (selected != LACP_SELECTED || !p_sync) {
1557                         new_state = LACP_MUX_ATTACHED;
1558                 }
1559                 break;
1560         case LACP_MUX_DISTRIBUTING:
1561                 if (selected != LACP_SELECTED || !p_sync || !p_collecting) {
1562                         new_state = LACP_MUX_COLLECTING;
1563                         lacp_dprintf(lp, "Interface stopped DISTRIBUTING, possible flapping\n");
1564                         sc->sc_flapping++;
1565                 }
1566                 break;
1567         default:
1568                 panic("%s: unknown state", __func__);
1569         }
1570
1571         if (lp->lp_mux_state == new_state) {
1572                 return;
1573         }
1574
1575         lacp_set_mux(lp, new_state);
1576         goto re_eval;
1577 }
1578
1579 static void
1580 lacp_set_mux(struct lacp_port *lp, enum lacp_mux_state new_state)
1581 {
1582         struct lacp_aggregator *la = lp->lp_aggregator;
1583
1584         if (lp->lp_mux_state == new_state) {
1585                 return;
1586         }
1587
1588         switch (new_state) {
1589         case LACP_MUX_DETACHED:
1590                 lp->lp_state &= ~LACP_STATE_SYNC;
1591                 lacp_disable_distributing(lp);
1592                 lacp_disable_collecting(lp);
1593                 lacp_sm_assert_ntt(lp);
1594                 /* cancel timer */
1595                 if (LACP_TIMER_ISARMED(lp, LACP_TIMER_WAIT_WHILE)) {
1596                         KASSERT(la->la_pending > 0,
1597                             ("timer_wait_while not active"));
1598                         la->la_pending--;
1599                 }
1600                 LACP_TIMER_DISARM(lp, LACP_TIMER_WAIT_WHILE);
1601                 lacp_unselect(lp);
1602                 break;
1603         case LACP_MUX_WAITING:
1604                 LACP_TIMER_ARM(lp, LACP_TIMER_WAIT_WHILE,
1605                     LACP_AGGREGATE_WAIT_TIME);
1606                 la->la_pending++;
1607                 break;
1608         case LACP_MUX_ATTACHED:
1609                 lp->lp_state |= LACP_STATE_SYNC;
1610                 lacp_disable_collecting(lp);
1611                 lacp_sm_assert_ntt(lp);
1612                 break;
1613         case LACP_MUX_COLLECTING:
1614                 lacp_enable_collecting(lp);
1615                 lacp_disable_distributing(lp);
1616                 lacp_sm_assert_ntt(lp);
1617                 break;
1618         case LACP_MUX_DISTRIBUTING:
1619                 lacp_enable_distributing(lp);
1620                 break;
1621         default:
1622                 panic("%s: unknown state", __func__);
1623         }
1624
1625         LACP_DPRINTF((lp, "mux_state %d -> %d\n", lp->lp_mux_state, new_state));
1626
1627         lp->lp_mux_state = new_state;
1628 }
1629
1630 static void
1631 lacp_sm_mux_timer(struct lacp_port *lp)
1632 {
1633         struct lacp_aggregator *la = lp->lp_aggregator;
1634         char buf[LACP_LAGIDSTR_MAX+1];
1635
1636         KASSERT(la->la_pending > 0, ("no pending event"));
1637
1638         LACP_DPRINTF((lp, "%s: aggregator %s, pending %d -> %d\n", __func__,
1639             lacp_format_lagid(&la->la_actor, &la->la_partner,
1640             buf, sizeof(buf)),
1641             la->la_pending, la->la_pending - 1));
1642
1643         la->la_pending--;
1644 }
1645
1646 /* periodic transmit machine */
1647
1648 static void
1649 lacp_sm_ptx_update_timeout(struct lacp_port *lp, uint8_t oldpstate)
1650 {
1651         if (LACP_STATE_EQ(oldpstate, lp->lp_partner.lip_state,
1652             LACP_STATE_TIMEOUT)) {
1653                 return;
1654         }
1655
1656         LACP_DPRINTF((lp, "partner timeout changed\n"));
1657
1658         /*
1659          * FAST_PERIODIC -> SLOW_PERIODIC
1660          * or
1661          * SLOW_PERIODIC (-> PERIODIC_TX) -> FAST_PERIODIC
1662          *
1663          * let lacp_sm_ptx_tx_schedule to update timeout.
1664          */
1665
1666         LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1667
1668         /*
1669          * if timeout has been shortened, assert NTT.
1670          */
1671
1672         if ((lp->lp_partner.lip_state & LACP_STATE_TIMEOUT)) {
1673                 lacp_sm_assert_ntt(lp);
1674         }
1675 }
1676
1677 static void
1678 lacp_sm_ptx_tx_schedule(struct lacp_port *lp)
1679 {
1680         int timeout;
1681
1682         if (!(lp->lp_state & LACP_STATE_ACTIVITY) &&
1683             !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY)) {
1684                 /*
1685                  * NO_PERIODIC
1686                  */
1687
1688                 LACP_TIMER_DISARM(lp, LACP_TIMER_PERIODIC);
1689                 return;
1690         }
1691
1692         if (LACP_TIMER_ISARMED(lp, LACP_TIMER_PERIODIC)) {
1693                 return;
1694         }
1695
1696         timeout = (lp->lp_partner.lip_state & LACP_STATE_TIMEOUT) ?
1697             LACP_FAST_PERIODIC_TIME : LACP_SLOW_PERIODIC_TIME;
1698
1699         LACP_TIMER_ARM(lp, LACP_TIMER_PERIODIC, timeout);
1700 }
1701
1702 static void
1703 lacp_sm_ptx_timer(struct lacp_port *lp)
1704 {
1705         lacp_sm_assert_ntt(lp);
1706 }
1707
1708 static void
1709 lacp_sm_rx(struct lacp_port *lp, const struct lacpdu *du)
1710 {
1711         int timeout;
1712
1713         /*
1714          * check LACP_DISABLED first
1715          */
1716
1717         if (!(lp->lp_state & LACP_STATE_AGGREGATION)) {
1718                 return;
1719         }
1720
1721         /*
1722          * check loopback condition.
1723          */
1724
1725         if (!lacp_compare_systemid(&du->ldu_actor.lip_systemid,
1726             &lp->lp_actor.lip_systemid)) {
1727                 return;
1728         }
1729
1730         /*
1731          * EXPIRED, DEFAULTED, CURRENT -> CURRENT
1732          */
1733
1734         microuptime(&lp->lp_last_lacpdu_rx);
1735         lacp_sm_rx_update_selected(lp, du);
1736         lacp_sm_rx_update_ntt(lp, du);
1737         lacp_sm_rx_record_pdu(lp, du);
1738
1739         timeout = (lp->lp_state & LACP_STATE_TIMEOUT) ?
1740             LACP_SHORT_TIMEOUT_TIME : LACP_LONG_TIMEOUT_TIME;
1741         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, timeout);
1742
1743         lp->lp_state &= ~LACP_STATE_EXPIRED;
1744
1745         /*
1746          * kick transmit machine without waiting the next tick.
1747          */
1748
1749         lacp_sm_tx(lp);
1750 }
1751
1752 static void
1753 lacp_sm_rx_set_expired(struct lacp_port *lp)
1754 {
1755         lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1756         lp->lp_partner.lip_state |= LACP_STATE_TIMEOUT;
1757         LACP_TIMER_ARM(lp, LACP_TIMER_CURRENT_WHILE, LACP_SHORT_TIMEOUT_TIME);
1758         lp->lp_state |= LACP_STATE_EXPIRED;
1759 }
1760
1761 static void
1762 lacp_sm_rx_timer(struct lacp_port *lp)
1763 {
1764         if ((lp->lp_state & LACP_STATE_EXPIRED) == 0) {
1765                 /* CURRENT -> EXPIRED */
1766                 LACP_DPRINTF((lp, "%s: CURRENT -> EXPIRED\n", __func__));
1767                 lacp_sm_rx_set_expired(lp);
1768         } else {
1769                 /* EXPIRED -> DEFAULTED */
1770                 LACP_DPRINTF((lp, "%s: EXPIRED -> DEFAULTED\n", __func__));
1771                 lacp_sm_rx_update_default_selected(lp);
1772                 lacp_sm_rx_record_default(lp);
1773                 lp->lp_state &= ~LACP_STATE_EXPIRED;
1774         }
1775 }
1776
1777 static void
1778 lacp_sm_rx_record_pdu(struct lacp_port *lp, const struct lacpdu *du)
1779 {
1780         boolean_t active;
1781         uint8_t oldpstate;
1782         char buf[LACP_STATESTR_MAX+1];
1783
1784         LACP_TRACE(lp);
1785
1786         oldpstate = lp->lp_partner.lip_state;
1787
1788         active = (du->ldu_actor.lip_state & LACP_STATE_ACTIVITY)
1789             || ((lp->lp_state & LACP_STATE_ACTIVITY) &&
1790             (du->ldu_partner.lip_state & LACP_STATE_ACTIVITY));
1791
1792         lp->lp_partner = du->ldu_actor;
1793         if (active &&
1794             ((LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1795             LACP_STATE_AGGREGATION) &&
1796             !lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner))
1797             || (du->ldu_partner.lip_state & LACP_STATE_AGGREGATION) == 0)) {
1798                 /*
1799                  * XXX Maintain legacy behavior of leaving the
1800                  * LACP_STATE_SYNC bit unchanged from the partner's
1801                  * advertisement if lsc_strict_mode is false.
1802                  * TODO: We should re-examine the concept of the "strict mode"
1803                  * to ensure it makes sense to maintain a non-strict mode.
1804                  */
1805                 if (lp->lp_lsc->lsc_strict_mode)
1806                         lp->lp_partner.lip_state |= LACP_STATE_SYNC;
1807         } else {
1808                 lp->lp_partner.lip_state &= ~LACP_STATE_SYNC;
1809         }
1810
1811         lp->lp_state &= ~LACP_STATE_DEFAULTED;
1812
1813         if (oldpstate != lp->lp_partner.lip_state) {
1814                 LACP_DPRINTF((lp, "old pstate %s\n",
1815                     lacp_format_state(oldpstate, buf, sizeof(buf))));
1816                 LACP_DPRINTF((lp, "new pstate %s\n",
1817                     lacp_format_state(lp->lp_partner.lip_state, buf,
1818                     sizeof(buf))));
1819         }
1820
1821         lacp_sm_ptx_update_timeout(lp, oldpstate);
1822 }
1823
1824 static void
1825 lacp_sm_rx_update_ntt(struct lacp_port *lp, const struct lacpdu *du)
1826 {
1827
1828         LACP_TRACE(lp);
1829
1830         if (lacp_compare_peerinfo(&lp->lp_actor, &du->ldu_partner) ||
1831             !LACP_STATE_EQ(lp->lp_state, du->ldu_partner.lip_state,
1832             LACP_STATE_ACTIVITY | LACP_STATE_SYNC | LACP_STATE_AGGREGATION)) {
1833                 LACP_DPRINTF((lp, "%s: assert ntt\n", __func__));
1834                 lacp_sm_assert_ntt(lp);
1835         }
1836 }
1837
1838 static void
1839 lacp_sm_rx_record_default(struct lacp_port *lp)
1840 {
1841         uint8_t oldpstate;
1842
1843         LACP_TRACE(lp);
1844
1845         oldpstate = lp->lp_partner.lip_state;
1846         if (lp->lp_lsc->lsc_strict_mode)
1847                 lp->lp_partner = lacp_partner_admin_strict;
1848         else
1849                 lp->lp_partner = lacp_partner_admin_optimistic;
1850         lp->lp_state |= LACP_STATE_DEFAULTED;
1851         lacp_sm_ptx_update_timeout(lp, oldpstate);
1852 }
1853
1854 static void
1855 lacp_sm_rx_update_selected_from_peerinfo(struct lacp_port *lp,
1856     const struct lacp_peerinfo *info)
1857 {
1858
1859         LACP_TRACE(lp);
1860
1861         if (lacp_compare_peerinfo(&lp->lp_partner, info) ||
1862             !LACP_STATE_EQ(lp->lp_partner.lip_state, info->lip_state,
1863             LACP_STATE_AGGREGATION)) {
1864                 lp->lp_selected = LACP_UNSELECTED;
1865                 /* mux machine will clean up lp->lp_aggregator */
1866         }
1867 }
1868
1869 static void
1870 lacp_sm_rx_update_selected(struct lacp_port *lp, const struct lacpdu *du)
1871 {
1872
1873         LACP_TRACE(lp);
1874
1875         lacp_sm_rx_update_selected_from_peerinfo(lp, &du->ldu_actor);
1876 }
1877
1878 static void
1879 lacp_sm_rx_update_default_selected(struct lacp_port *lp)
1880 {
1881
1882         LACP_TRACE(lp);
1883
1884         if (lp->lp_lsc->lsc_strict_mode)
1885                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1886                     &lacp_partner_admin_strict);
1887         else
1888                 lacp_sm_rx_update_selected_from_peerinfo(lp,
1889                     &lacp_partner_admin_optimistic);
1890 }
1891
1892 /* transmit machine */
1893
1894 static void
1895 lacp_sm_tx(struct lacp_port *lp)
1896 {
1897         int error = 0;
1898
1899         if (!(lp->lp_state & LACP_STATE_AGGREGATION)
1900 #if 1
1901             || (!(lp->lp_state & LACP_STATE_ACTIVITY)
1902             && !(lp->lp_partner.lip_state & LACP_STATE_ACTIVITY))
1903 #endif
1904             ) {
1905                 lp->lp_flags &= ~LACP_PORT_NTT;
1906         }
1907
1908         if (!(lp->lp_flags & LACP_PORT_NTT)) {
1909                 return;
1910         }
1911
1912         /* Rate limit to 3 PDUs per LACP_FAST_PERIODIC_TIME */
1913         if (ppsratecheck(&lp->lp_last_lacpdu, &lp->lp_lacpdu_sent,
1914                     (3 / LACP_FAST_PERIODIC_TIME)) == 0) {
1915                 LACP_DPRINTF((lp, "rate limited pdu\n"));
1916                 return;
1917         }
1918
1919         if (((1 << lp->lp_ifp->if_dunit) & lp->lp_lsc->lsc_debug.lsc_tx_test) == 0) {
1920                 error = lacp_xmit_lacpdu(lp);
1921         } else {
1922                 LACP_TPRINTF((lp, "Dropping TX PDU\n"));
1923         }
1924
1925         if (error == 0) {
1926                 lp->lp_flags &= ~LACP_PORT_NTT;
1927         } else {
1928                 LACP_DPRINTF((lp, "lacpdu transmit failure, error %d\n",
1929                     error));
1930         }
1931 }
1932
1933 static void
1934 lacp_sm_assert_ntt(struct lacp_port *lp)
1935 {
1936
1937         lp->lp_flags |= LACP_PORT_NTT;
1938 }
1939
1940 static void
1941 lacp_run_timers(struct lacp_port *lp)
1942 {
1943         int i;
1944         struct timeval time_diff;
1945
1946         for (i = 0; i < LACP_NTIMER; i++) {
1947                 KASSERT(lp->lp_timer[i] >= 0,
1948                     ("invalid timer value %d", lp->lp_timer[i]));
1949                 if (lp->lp_timer[i] == 0) {
1950                         continue;
1951                 } else {
1952                         if (i == LACP_TIMER_CURRENT_WHILE) {
1953                                 microuptime(&time_diff);
1954                                 timevalsub(&time_diff, &lp->lp_last_lacpdu_rx);
1955                                 if (time_diff.tv_sec) {
1956                                         /* At least one sec has elapsed since last LACP packet. */
1957                                         --lp->lp_timer[i];
1958                                 }
1959                         } else {
1960                                 --lp->lp_timer[i];
1961                         }
1962
1963                         if ((lp->lp_timer[i] <= 0) && (lacp_timer_funcs[i])) {
1964                                 (*lacp_timer_funcs[i])(lp);
1965                         }
1966                 }
1967         }
1968 }
1969
1970 int
1971 lacp_marker_input(struct lacp_port *lp, struct mbuf *m)
1972 {
1973         struct lacp_softc *lsc = lp->lp_lsc;
1974         struct lagg_port *lgp = lp->lp_lagg;
1975         struct lacp_port *lp2;
1976         struct markerdu *mdu;
1977         int error = 0;
1978         int pending = 0;
1979
1980         if (m->m_pkthdr.len != sizeof(*mdu)) {
1981                 goto bad;
1982         }
1983
1984         if ((m->m_flags & M_MCAST) == 0) {
1985                 goto bad;
1986         }
1987
1988         if (m->m_len < sizeof(*mdu)) {
1989                 m = m_pullup(m, sizeof(*mdu));
1990                 if (m == NULL) {
1991                         return (ENOMEM);
1992                 }
1993         }
1994
1995         mdu = mtod(m, struct markerdu *);
1996
1997         if (memcmp(&mdu->mdu_eh.ether_dhost,
1998             &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN)) {
1999                 goto bad;
2000         }
2001
2002         if (mdu->mdu_sph.sph_version != 1) {
2003                 goto bad;
2004         }
2005
2006         switch (mdu->mdu_tlv.tlv_type) {
2007         case MARKER_TYPE_INFO:
2008                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2009                     marker_info_tlv_template, TRUE)) {
2010                         goto bad;
2011                 }
2012                 mdu->mdu_tlv.tlv_type = MARKER_TYPE_RESPONSE;
2013                 memcpy(&mdu->mdu_eh.ether_dhost,
2014                     &ethermulticastaddr_slowprotocols, ETHER_ADDR_LEN);
2015                 memcpy(&mdu->mdu_eh.ether_shost,
2016                     lgp->lp_lladdr, ETHER_ADDR_LEN);
2017                 error = lagg_enqueue(lp->lp_ifp, m);
2018                 break;
2019
2020         case MARKER_TYPE_RESPONSE:
2021                 if (tlv_check(mdu, sizeof(*mdu), &mdu->mdu_tlv,
2022                     marker_response_tlv_template, TRUE)) {
2023                         goto bad;
2024                 }
2025                 LACP_DPRINTF((lp, "marker response, port=%u, sys=%6D, id=%u\n",
2026                     ntohs(mdu->mdu_info.mi_rq_port), mdu->mdu_info.mi_rq_system,
2027                     ":", ntohl(mdu->mdu_info.mi_rq_xid)));
2028
2029                 /* Verify that it is the last marker we sent out */
2030                 if (memcmp(&mdu->mdu_info, &lp->lp_marker,
2031                     sizeof(struct lacp_markerinfo)))
2032                         goto bad;
2033
2034                 LACP_LOCK(lsc);
2035                 lp->lp_flags &= ~LACP_PORT_MARK;
2036
2037                 if (lsc->lsc_suppress_distributing) {
2038                         /* Check if any ports are waiting for a response */
2039                         LIST_FOREACH(lp2, &lsc->lsc_ports, lp_next) {
2040                                 if (lp2->lp_flags & LACP_PORT_MARK) {
2041                                         pending = 1;
2042                                         break;
2043                                 }
2044                         }
2045
2046                         if (pending == 0) {
2047                                 /* All interface queues are clear */
2048                                 LACP_DPRINTF((NULL, "queue flush complete\n"));
2049                                 lsc->lsc_suppress_distributing = FALSE;
2050                         }
2051                 }
2052                 LACP_UNLOCK(lsc);
2053                 m_freem(m);
2054                 break;
2055
2056         default:
2057                 goto bad;
2058         }
2059
2060         return (error);
2061
2062 bad:
2063         LACP_DPRINTF((lp, "bad marker frame\n"));
2064         m_freem(m);
2065         return (EINVAL);
2066 }
2067
2068 static int
2069 tlv_check(const void *p, size_t size, const struct tlvhdr *tlv,
2070     const struct tlv_template *tmpl, boolean_t check_type)
2071 {
2072         while (/* CONSTCOND */ 1) {
2073                 if ((const char *)tlv - (const char *)p + sizeof(*tlv) > size) {
2074                         return (EINVAL);
2075                 }
2076                 if ((check_type && tlv->tlv_type != tmpl->tmpl_type) ||
2077                     tlv->tlv_length != tmpl->tmpl_length) {
2078                         return (EINVAL);
2079                 }
2080                 if (tmpl->tmpl_type == 0) {
2081                         break;
2082                 }
2083                 tlv = (const struct tlvhdr *)
2084                     ((const char *)tlv + tlv->tlv_length);
2085                 tmpl++;
2086         }
2087
2088         return (0);
2089 }
2090
2091 /* Debugging */
2092 const char *
2093 lacp_format_mac(const uint8_t *mac, char *buf, size_t buflen)
2094 {
2095         snprintf(buf, buflen, "%02X-%02X-%02X-%02X-%02X-%02X",
2096             (int)mac[0],
2097             (int)mac[1],
2098             (int)mac[2],
2099             (int)mac[3],
2100             (int)mac[4],
2101             (int)mac[5]);
2102
2103         return (buf);
2104 }
2105
2106 const char *
2107 lacp_format_systemid(const struct lacp_systemid *sysid,
2108     char *buf, size_t buflen)
2109 {
2110         char macbuf[LACP_MACSTR_MAX+1];
2111
2112         snprintf(buf, buflen, "%04X,%s",
2113             ntohs(sysid->lsi_prio),
2114             lacp_format_mac(sysid->lsi_mac, macbuf, sizeof(macbuf)));
2115
2116         return (buf);
2117 }
2118
2119 const char *
2120 lacp_format_portid(const struct lacp_portid *portid, char *buf, size_t buflen)
2121 {
2122         snprintf(buf, buflen, "%04X,%04X",
2123             ntohs(portid->lpi_prio),
2124             ntohs(portid->lpi_portno));
2125
2126         return (buf);
2127 }
2128
2129 const char *
2130 lacp_format_partner(const struct lacp_peerinfo *peer, char *buf, size_t buflen)
2131 {
2132         char sysid[LACP_SYSTEMIDSTR_MAX+1];
2133         char portid[LACP_PORTIDSTR_MAX+1];
2134
2135         snprintf(buf, buflen, "(%s,%04X,%s)",
2136             lacp_format_systemid(&peer->lip_systemid, sysid, sizeof(sysid)),
2137             ntohs(peer->lip_key),
2138             lacp_format_portid(&peer->lip_portid, portid, sizeof(portid)));
2139
2140         return (buf);
2141 }
2142
2143 const char *
2144 lacp_format_lagid(const struct lacp_peerinfo *a,
2145     const struct lacp_peerinfo *b, char *buf, size_t buflen)
2146 {
2147         char astr[LACP_PARTNERSTR_MAX+1];
2148         char bstr[LACP_PARTNERSTR_MAX+1];
2149
2150 #if 0
2151         /*
2152          * there's a convention to display small numbered peer
2153          * in the left.
2154          */
2155
2156         if (lacp_compare_peerinfo(a, b) > 0) {
2157                 const struct lacp_peerinfo *t;
2158
2159                 t = a;
2160                 a = b;
2161                 b = t;
2162         }
2163 #endif
2164
2165         snprintf(buf, buflen, "[%s,%s]",
2166             lacp_format_partner(a, astr, sizeof(astr)),
2167             lacp_format_partner(b, bstr, sizeof(bstr)));
2168
2169         return (buf);
2170 }
2171
2172 const char *
2173 lacp_format_lagid_aggregator(const struct lacp_aggregator *la,
2174     char *buf, size_t buflen)
2175 {
2176         if (la == NULL) {
2177                 return ("(none)");
2178         }
2179
2180         return (lacp_format_lagid(&la->la_actor, &la->la_partner, buf, buflen));
2181 }
2182
2183 const char *
2184 lacp_format_state(uint8_t state, char *buf, size_t buflen)
2185 {
2186         snprintf(buf, buflen, "%b", state, LACP_STATE_BITS);
2187         return (buf);
2188 }
2189
2190 static void
2191 lacp_dump_lacpdu(const struct lacpdu *du)
2192 {
2193         char buf[LACP_PARTNERSTR_MAX+1];
2194         char buf2[LACP_STATESTR_MAX+1];
2195
2196         printf("actor=%s\n",
2197             lacp_format_partner(&du->ldu_actor, buf, sizeof(buf)));
2198         printf("actor.state=%s\n",
2199             lacp_format_state(du->ldu_actor.lip_state, buf2, sizeof(buf2)));
2200         printf("partner=%s\n",
2201             lacp_format_partner(&du->ldu_partner, buf, sizeof(buf)));
2202         printf("partner.state=%s\n",
2203             lacp_format_state(du->ldu_partner.lip_state, buf2, sizeof(buf2)));
2204
2205         printf("maxdelay=%d\n", ntohs(du->ldu_collector.lci_maxdelay));
2206 }
2207
2208 static void
2209 lacp_dprintf(const struct lacp_port *lp, const char *fmt, ...)
2210 {
2211         va_list va;
2212
2213         if (lp) {
2214                 printf("%s: ", lp->lp_ifp->if_xname);
2215         }
2216
2217         va_start(va, fmt);
2218         vprintf(fmt, va);
2219         va_end(va);
2220 }