drivers/net/bonding/rte_eth_bond_8023ad.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2015 Intel Corporation
   3  */
   4
   5 #include <stddef.h>
   6 #include <string.h>
   7 #include <stdbool.h>
   8
   9 #include <rte_alarm.h>
  10 #include <rte_malloc.h>
  11 #include <rte_errno.h>
  12 #include <rte_cycles.h>
  13 #include <rte_compat.h>
  14
  15 #include "rte_eth_bond_private.h"
  16
  17 static void bond_mode_8023ad_ext_periodic_cb(void *arg);
  18 #ifdef RTE_LIBRTE_BOND_DEBUG_8023AD
  19
  20 #define MODE4_DEBUG(fmt, ...)                           \
  21         rte_log(RTE_LOG_DEBUG, bond_logtype,            \
  22                 "%6u [Port %u: %s] " fmt,               \
  23                 bond_dbg_get_time_diff_ms(), slave_id,  \
  24                 __func__, ##__VA_ARGS__)
  25
  26 static uint64_t start_time;
  27
  28 static unsigned
  29 bond_dbg_get_time_diff_ms(void)
  30 {
  31         uint64_t now;
  32
  33         now = rte_rdtsc();
  34         if (start_time == 0)
  35                 start_time = now;
  36
  37         return ((now - start_time) * 1000) / rte_get_tsc_hz();
  38 }
  39
  40 static void
  41 bond_print_lacp(struct lacpdu *l)
  42 {
  43         char a_address[18];
  44         char p_address[18];
  45         char a_state[256] = { 0 };
  46         char p_state[256] = { 0 };
  47
  48         static const char * const state_labels[] = {
  49                 "ACT", "TIMEOUT", "AGG", "SYNC", "COL", "DIST", "DEF", "EXP"
  50         };
  51
  52         int a_len = 0;
  53         int p_len = 0;
  54         uint8_t i;
  55         uint8_t *addr;
  56
  57         addr = l->actor.port_params.system.addr_bytes;
  58         snprintf(a_address, sizeof(a_address), "%02X:%02X:%02X:%02X:%02X:%02X",
  59                 addr[0], addr[1], addr[2], addr[3], addr[4], addr[5]);
  60
  61         addr = l->partner.port_params.system.addr_bytes;
  62         snprintf(p_address, sizeof(p_address), "%02X:%02X:%02X:%02X:%02X:%02X",
  63                 addr[0], addr[1], addr[2], addr[3], addr[4], addr[5]);
  64
  65         for (i = 0; i < 8; i++) {
  66                 if ((l->actor.state >> i) & 1) {
  67                         a_len += snprintf(&a_state[a_len], RTE_DIM(a_state) - a_len, "%s ",
  68                                 state_labels[i]);
  69                 }
  70
  71                 if ((l->partner.state >> i) & 1) {
  72                         p_len += snprintf(&p_state[p_len], RTE_DIM(p_state) - p_len, "%s ",
  73                                 state_labels[i]);
  74                 }
  75         }
  76
  77         if (a_len && a_state[a_len-1] == ' ')
  78                 a_state[a_len-1] = '\0';
  79
  80         if (p_len && p_state[p_len-1] == ' ')
  81                 p_state[p_len-1] = '\0';
  82
  83         RTE_BOND_LOG(DEBUG,
  84                      "LACP: {\n"
  85                      "  subtype= %02X\n"
  86                      "  ver_num=%02X\n"
  87                      "  actor={ tlv=%02X, len=%02X\n"
  88                      "    pri=%04X, system=%s, key=%04X, p_pri=%04X p_num=%04X\n"
  89                      "       state={ %s }\n"
  90                      "  }\n"
  91                      "  partner={ tlv=%02X, len=%02X\n"
  92                      "    pri=%04X, system=%s, key=%04X, p_pri=%04X p_num=%04X\n"
  93                      "       state={ %s }\n"
  94                      "  }\n"
  95                      "  collector={info=%02X, length=%02X, max_delay=%04X\n, "
  96                      "type_term=%02X, terminator_length = %02X }",
  97                      l->subtype,
  98                      l->version_number,
  99                      l->actor.tlv_type_info,
 100                      l->actor.info_length,
 101                      l->actor.port_params.system_priority,
 102                      a_address,
 103                      l->actor.port_params.key,
 104                      l->actor.port_params.port_priority,
 105                      l->actor.port_params.port_number,
 106                      a_state,
 107                      l->partner.tlv_type_info,
 108                      l->partner.info_length,
 109                      l->partner.port_params.system_priority,
 110                      p_address,
 111                      l->partner.port_params.key,
 112                      l->partner.port_params.port_priority,
 113                      l->partner.port_params.port_number,
 114                      p_state,
 115                      l->tlv_type_collector_info,
 116                      l->collector_info_length,
 117                      l->collector_max_delay,
 118                      l->tlv_type_terminator,
 119                      l->terminator_length);
 120
 121 }
 122
 123 #define BOND_PRINT_LACP(lacpdu) bond_print_lacp(lacpdu)
 124 #else
 125 #define BOND_PRINT_LACP(lacpdu) do { } while (0)
 126 #define MODE4_DEBUG(fmt, ...) do { } while (0)
 127 #endif
 128
 129 static const struct ether_addr lacp_mac_addr = {
 130         .addr_bytes = { 0x01, 0x80, 0xC2, 0x00, 0x00, 0x02 }
 131 };
 132
 133 struct port mode_8023ad_ports[RTE_MAX_ETHPORTS];
 134
 135 static void
 136 timer_cancel(uint64_t *timer)
 137 {
 138         *timer = 0;
 139 }
 140
 141 static void
 142 timer_set(uint64_t *timer, uint64_t timeout)
 143 {
 144         *timer = rte_rdtsc() + timeout;
 145 }
 146
 147 /* Forces given timer to be in expired state. */
 148 static void
 149 timer_force_expired(uint64_t *timer)
 150 {
 151         *timer = rte_rdtsc();
 152 }
 153
 154 static bool
 155 timer_is_stopped(uint64_t *timer)
 156 {
 157         return *timer == 0;
 158 }
 159
 160 static bool
 161 timer_is_expired(uint64_t *timer)
 162 {
 163         return *timer < rte_rdtsc();
 164 }
 165
 166 /* Timer is in running state if it is not stopped nor expired */
 167 static bool
 168 timer_is_running(uint64_t *timer)
 169 {
 170         return !timer_is_stopped(timer) && !timer_is_expired(timer);
 171 }
 172
 173 static void
 174 set_warning_flags(struct port *port, uint16_t flags)
 175 {
 176         int retval;
 177         uint16_t old;
 178         uint16_t new_flag = 0;
 179
 180         do {
 181                 old = port->warnings_to_show;
 182                 new_flag = old | flags;
 183                 retval = rte_atomic16_cmpset(&port->warnings_to_show, old, new_flag);
 184         } while (unlikely(retval == 0));
 185 }
 186
 187 static void
 188 show_warnings(uint16_t slave_id)
 189 {
 190         struct port *port = &mode_8023ad_ports[slave_id];
 191         uint8_t warnings;
 192
 193         do {
 194                 warnings = port->warnings_to_show;
 195         } while (rte_atomic16_cmpset(&port->warnings_to_show, warnings, 0) == 0);
 196
 197         if (!warnings)
 198                 return;
 199
 200         if (!timer_is_expired(&port->warning_timer))
 201                 return;
 202
 203
 204         timer_set(&port->warning_timer, BOND_8023AD_WARNINGS_PERIOD_MS *
 205                         rte_get_tsc_hz() / 1000);
 206
 207         if (warnings & WRN_RX_QUEUE_FULL) {
 208                 RTE_BOND_LOG(DEBUG,
 209                              "Slave %u: failed to enqueue LACP packet into RX ring.\n"
 210                              "Receive and transmit functions must be invoked on bonded"
 211                              "interface at least 10 times per second or LACP will notwork correctly",
 212                              slave_id);
 213         }
 214
 215         if (warnings & WRN_TX_QUEUE_FULL) {
 216                 RTE_BOND_LOG(DEBUG,
 217                              "Slave %u: failed to enqueue LACP packet into TX ring.\n"
 218                              "Receive and transmit functions must be invoked on bonded"
 219                              "interface at least 10 times per second or LACP will not work correctly",
 220                              slave_id);
 221         }
 222
 223         if (warnings & WRN_RX_MARKER_TO_FAST)
 224                 RTE_BOND_LOG(INFO, "Slave %u: marker to early - ignoring.",
 225                              slave_id);
 226
 227         if (warnings & WRN_UNKNOWN_SLOW_TYPE) {
 228                 RTE_BOND_LOG(INFO,
 229                         "Slave %u: ignoring unknown slow protocol frame type",
 230                              slave_id);
 231         }
 232
 233         if (warnings & WRN_UNKNOWN_MARKER_TYPE)
 234                 RTE_BOND_LOG(INFO, "Slave %u: ignoring unknown marker type",
 235                              slave_id);
 236
 237         if (warnings & WRN_NOT_LACP_CAPABLE)
 238                 MODE4_DEBUG("Port %u is not LACP capable!\n", slave_id);
 239 }
 240
 241 static void
 242 record_default(struct port *port)
 243 {
 244         /* Record default parameters for partner. Partner admin parameters
 245          * are not implemented so set them to arbitrary default (last known) and
 246          * mark actor that parner is in defaulted state. */
 247         port->partner_state = STATE_LACP_ACTIVE;
 248         ACTOR_STATE_SET(port, DEFAULTED);
 249 }
 250
 251 /** Function handles rx state machine.
 252  *
 253  * This function implements Receive State Machine from point 5.4.12 in
 254  * 802.1AX documentation. It should be called periodically.
 255  *
 256  * @param lacpdu                LACPDU received.
 257  * @param port                  Port on which LACPDU was received.
 258  */
 259 static void
 260 rx_machine(struct bond_dev_private *internals, uint16_t slave_id,
 261                 struct lacpdu *lacp)
 262 {
 263         struct port *agg, *port = &mode_8023ad_ports[slave_id];
 264         uint64_t timeout;
 265
 266         if (SM_FLAG(port, BEGIN)) {
 267                 /* Initialize stuff */
 268                 MODE4_DEBUG("-> INITIALIZE\n");
 269                 SM_FLAG_CLR(port, MOVED);
 270                 port->selected = UNSELECTED;
 271
 272                 record_default(port);
 273
 274                 ACTOR_STATE_CLR(port, EXPIRED);
 275                 timer_cancel(&port->current_while_timer);
 276
 277                 /* DISABLED: On initialization partner is out of sync */
 278                 PARTNER_STATE_CLR(port, SYNCHRONIZATION);
 279
 280                 /* LACP DISABLED stuff if LACP not enabled on this port */
 281                 if (!SM_FLAG(port, LACP_ENABLED))
 282                         PARTNER_STATE_CLR(port, AGGREGATION);
 283                 else
 284                         PARTNER_STATE_SET(port, AGGREGATION);
 285         }
 286
 287         if (!SM_FLAG(port, LACP_ENABLED)) {
 288                 /* Update parameters only if state changed */
 289                 if (!timer_is_stopped(&port->current_while_timer)) {
 290                         port->selected = UNSELECTED;
 291                         record_default(port);
 292                         PARTNER_STATE_CLR(port, AGGREGATION);
 293                         ACTOR_STATE_CLR(port, EXPIRED);
 294                         timer_cancel(&port->current_while_timer);
 295                 }
 296                 return;
 297         }
 298
 299         if (lacp) {
 300                 MODE4_DEBUG("LACP -> CURRENT\n");
 301                 BOND_PRINT_LACP(lacp);
 302                 /* Update selected flag. If partner parameters are defaulted assume they
 303                  * are match. If not defaulted  compare LACP actor with ports parner
 304                  * params. */
 305                 if (!ACTOR_STATE(port, DEFAULTED) &&
 306                         (ACTOR_STATE(port, AGGREGATION) != PARTNER_STATE(port, AGGREGATION)
 307                         || memcmp(&port->partner, &lacp->actor.port_params,
 308                                 sizeof(port->partner)) != 0)) {
 309                         MODE4_DEBUG("selected <- UNSELECTED\n");
 310                         port->selected = UNSELECTED;
 311                 }
 312
 313                 /* Record this PDU actor params as partner params */
 314                 memcpy(&port->partner, &lacp->actor.port_params,
 315                         sizeof(struct port_params));
 316                 port->partner_state = lacp->actor.state;
 317
 318                 /* Partner parameters are not defaulted any more */
 319                 ACTOR_STATE_CLR(port, DEFAULTED);
 320
 321                 /* If LACP partner params match this port actor params */
 322                 agg = &mode_8023ad_ports[port->aggregator_port_id];
 323                 bool match = port->actor.system_priority ==
 324                         lacp->partner.port_params.system_priority &&
 325                         is_same_ether_addr(&agg->actor.system,
 326                         &lacp->partner.port_params.system) &&
 327                         port->actor.port_priority ==
 328                         lacp->partner.port_params.port_priority &&
 329                         port->actor.port_number ==
 330                         lacp->partner.port_params.port_number;
 331
 332                 /* Update NTT if partners information are outdated (xored and masked
 333                  * bits are set)*/
 334                 uint8_t state_mask = STATE_LACP_ACTIVE | STATE_LACP_SHORT_TIMEOUT |
 335                         STATE_SYNCHRONIZATION | STATE_AGGREGATION;
 336
 337                 if (((port->actor_state ^ lacp->partner.state) & state_mask) ||
 338                                 match == false) {
 339                         SM_FLAG_SET(port, NTT);
 340                 }
 341
 342                 /* If LACP partner params match this port actor params */
 343                 if (match == true && ACTOR_STATE(port, AGGREGATION) ==
 344                                 PARTNER_STATE(port,     AGGREGATION))
 345                         PARTNER_STATE_SET(port, SYNCHRONIZATION);
 346                 else if (!PARTNER_STATE(port, AGGREGATION) && ACTOR_STATE(port,
 347                                 AGGREGATION))
 348                         PARTNER_STATE_SET(port, SYNCHRONIZATION);
 349                 else
 350                         PARTNER_STATE_CLR(port, SYNCHRONIZATION);
 351
 352                 if (ACTOR_STATE(port, LACP_SHORT_TIMEOUT))
 353                         timeout = internals->mode4.short_timeout;
 354                 else
 355                         timeout = internals->mode4.long_timeout;
 356
 357                 timer_set(&port->current_while_timer, timeout);
 358                 ACTOR_STATE_CLR(port, EXPIRED);
 359                 return; /* No state change */
 360         }
 361
 362         /* If CURRENT state timer is not running (stopped or expired)
 363          * transit to EXPIRED state from DISABLED or CURRENT */
 364         if (!timer_is_running(&port->current_while_timer)) {
 365                 ACTOR_STATE_SET(port, EXPIRED);
 366                 PARTNER_STATE_CLR(port, SYNCHRONIZATION);
 367                 PARTNER_STATE_SET(port, LACP_SHORT_TIMEOUT);
 368                 timer_set(&port->current_while_timer, internals->mode4.short_timeout);
 369         }
 370 }
 371
 372 /**
 373  * Function handles periodic tx state machine.
 374  *
 375  * Function implements Periodic Transmission state machine from point 5.4.13
 376  * in 802.1AX documentation. It should be called periodically.
 377  *
 378  * @param port                  Port to handle state machine.
 379  */
 380 static void
 381 periodic_machine(struct bond_dev_private *internals, uint16_t slave_id)
 382 {
 383         struct port *port = &mode_8023ad_ports[slave_id];
 384         /* Calculate if either site is LACP enabled */
 385         uint64_t timeout;
 386         uint8_t active = ACTOR_STATE(port, LACP_ACTIVE) ||
 387                 PARTNER_STATE(port, LACP_ACTIVE);
 388
 389         uint8_t is_partner_fast, was_partner_fast;
 390         /* No periodic is on BEGIN, LACP DISABLE or when both sides are pasive */
 391         if (SM_FLAG(port, BEGIN) || !SM_FLAG(port, LACP_ENABLED) || !active) {
 392                 timer_cancel(&port->periodic_timer);
 393                 timer_force_expired(&port->tx_machine_timer);
 394                 SM_FLAG_CLR(port, PARTNER_SHORT_TIMEOUT);
 395
 396                 MODE4_DEBUG("-> NO_PERIODIC ( %s%s%s)\n",
 397                         SM_FLAG(port, BEGIN) ? "begind " : "",
 398                         SM_FLAG(port, LACP_ENABLED) ? "" : "LACP disabled ",
 399                         active ? "LACP active " : "LACP pasive ");
 400                 return;
 401         }
 402
 403         is_partner_fast = PARTNER_STATE(port, LACP_SHORT_TIMEOUT);
 404         was_partner_fast = SM_FLAG(port, PARTNER_SHORT_TIMEOUT);
 405
 406         /* If periodic timer is not started, transit from NO PERIODIC to FAST/SLOW.
 407          * Other case: check if timer expire or partners settings changed. */
 408         if (!timer_is_stopped(&port->periodic_timer)) {
 409                 if (timer_is_expired(&port->periodic_timer)) {
 410                         SM_FLAG_SET(port, NTT);
 411                 } else if (is_partner_fast != was_partner_fast) {
 412                         /* Partners timeout  was slow and now it is fast -> send LACP.
 413                          * In other case (was fast and now it is slow) just switch
 414                          * timeout to slow without forcing send of LACP (because standard
 415                          * say so)*/
 416                         if (is_partner_fast)
 417                                 SM_FLAG_SET(port, NTT);
 418                 } else
 419                         return; /* Nothing changed */
 420         }
 421
 422         /* Handle state transition to FAST/SLOW LACP timeout */
 423         if (is_partner_fast) {
 424                 timeout = internals->mode4.fast_periodic_timeout;
 425                 SM_FLAG_SET(port, PARTNER_SHORT_TIMEOUT);
 426         } else {
 427                 timeout = internals->mode4.slow_periodic_timeout;
 428                 SM_FLAG_CLR(port, PARTNER_SHORT_TIMEOUT);
 429         }
 430
 431         timer_set(&port->periodic_timer, timeout);
 432 }
 433
 434 /**
 435  * Function handles mux state machine.
 436  *
 437  * Function implements Mux Machine from point 5.4.15 in 802.1AX documentation.
 438  * It should be called periodically.
 439  *
 440  * @param port                  Port to handle state machine.
 441  */
 442 static void
 443 mux_machine(struct bond_dev_private *internals, uint16_t slave_id)
 444 {
 445         struct port *port = &mode_8023ad_ports[slave_id];
 446
 447         /* Save current state for later use */
 448         const uint8_t state_mask = STATE_SYNCHRONIZATION | STATE_DISTRIBUTING |
 449                 STATE_COLLECTING;
 450
 451         /* Enter DETACHED state on BEGIN condition or from any other state if
 452          * port was unselected */
 453         if (SM_FLAG(port, BEGIN) ||
 454                         port->selected == UNSELECTED || (port->selected == STANDBY &&
 455                                 (port->actor_state & state_mask) != 0)) {
 456                 /* detach mux from aggregator */
 457                 port->actor_state &= ~state_mask;
 458                 /* Set ntt to true if BEGIN condition or transition from any other state
 459                  * which is indicated that wait_while_timer was started */
 460                 if (SM_FLAG(port, BEGIN) ||
 461                                 !timer_is_stopped(&port->wait_while_timer)) {
 462                         SM_FLAG_SET(port, NTT);
 463                         MODE4_DEBUG("-> DETACHED\n");
 464                 }
 465                 timer_cancel(&port->wait_while_timer);
 466         }
 467
 468         if (timer_is_stopped(&port->wait_while_timer)) {
 469                 if (port->selected == SELECTED || port->selected == STANDBY) {
 470                         timer_set(&port->wait_while_timer,
 471                                 internals->mode4.aggregate_wait_timeout);
 472
 473                         MODE4_DEBUG("DETACHED -> WAITING\n");
 474                 }
 475                 /* Waiting state entered */
 476                 return;
 477         }
 478
 479         /* Transit next state if port is ready */
 480         if (!timer_is_expired(&port->wait_while_timer))
 481                 return;
 482
 483         if ((ACTOR_STATE(port, DISTRIBUTING) || ACTOR_STATE(port, COLLECTING)) &&
 484                 !PARTNER_STATE(port, SYNCHRONIZATION)) {
 485                 /* If in COLLECTING or DISTRIBUTING state and partner becomes out of
 486                  * sync transit to ATACHED state.  */
 487                 ACTOR_STATE_CLR(port, DISTRIBUTING);
 488                 ACTOR_STATE_CLR(port, COLLECTING);
 489                 /* Clear actor sync to activate transit ATACHED in condition bellow */
 490                 ACTOR_STATE_CLR(port, SYNCHRONIZATION);
 491                 MODE4_DEBUG("Out of sync -> ATTACHED\n");
 492         }
 493
 494         if (!ACTOR_STATE(port, SYNCHRONIZATION)) {
 495                 /* attach mux to aggregator */
 496                 RTE_ASSERT((port->actor_state & (STATE_COLLECTING |
 497                         STATE_DISTRIBUTING)) == 0);
 498
 499                 ACTOR_STATE_SET(port, SYNCHRONIZATION);
 500                 SM_FLAG_SET(port, NTT);
 501                 MODE4_DEBUG("ATTACHED Entered\n");
 502         } else if (!ACTOR_STATE(port, COLLECTING)) {
 503                 /* Start collecting if in sync */
 504                 if (PARTNER_STATE(port, SYNCHRONIZATION)) {
 505                         MODE4_DEBUG("ATTACHED -> COLLECTING\n");
 506                         ACTOR_STATE_SET(port, COLLECTING);
 507                         SM_FLAG_SET(port, NTT);
 508                 }
 509         } else if (ACTOR_STATE(port, COLLECTING)) {
 510                 /* Check if partner is in COLLECTING state. If so this port can
 511                  * distribute frames to it */
 512                 if (!ACTOR_STATE(port, DISTRIBUTING)) {
 513                         if (PARTNER_STATE(port, COLLECTING)) {
 514                                 /* Enable  DISTRIBUTING if partner is collecting */
 515                                 ACTOR_STATE_SET(port, DISTRIBUTING);
 516                                 SM_FLAG_SET(port, NTT);
 517                                 MODE4_DEBUG("COLLECTING -> DISTRIBUTING\n");
 518                                 RTE_BOND_LOG(INFO,
 519                                         "Bond %u: slave id %u distributing started.",
 520                                         internals->port_id, slave_id);
 521                         }
 522                 } else {
 523                         if (!PARTNER_STATE(port, COLLECTING)) {
 524                                 /* Disable DISTRIBUTING (enter COLLECTING state) if partner
 525                                  * is not collecting */
 526                                 ACTOR_STATE_CLR(port, DISTRIBUTING);
 527                                 SM_FLAG_SET(port, NTT);
 528                                 MODE4_DEBUG("DISTRIBUTING -> COLLECTING\n");
 529                                 RTE_BOND_LOG(INFO,
 530                                         "Bond %u: slave id %u distributing stopped.",
 531                                         internals->port_id, slave_id);
 532                         }
 533                 }
 534         }
 535 }
 536
 537 /**
 538  * Function handles transmit state machine.
 539  *
 540  * Function implements Transmit Machine from point 5.4.16 in 802.1AX
 541  * documentation.
 542  *
 543  * @param port
 544  */
 545 static void
 546 tx_machine(struct bond_dev_private *internals, uint16_t slave_id)
 547 {
 548         struct port *agg, *port = &mode_8023ad_ports[slave_id];
 549
 550         struct rte_mbuf *lacp_pkt = NULL;
 551         struct lacpdu_header *hdr;
 552         struct lacpdu *lacpdu;
 553
 554         /* If periodic timer is not running periodic machine is in NO PERIODIC and
 555          * according to 802.3ax standard tx machine should not transmit any frames
 556          * and set ntt to false. */
 557         if (timer_is_stopped(&port->periodic_timer))
 558                 SM_FLAG_CLR(port, NTT);
 559
 560         if (!SM_FLAG(port, NTT))
 561                 return;
 562
 563         if (!timer_is_expired(&port->tx_machine_timer))
 564                 return;
 565
 566         lacp_pkt = rte_pktmbuf_alloc(port->mbuf_pool);
 567         if (lacp_pkt == NULL) {
 568                 RTE_BOND_LOG(ERR, "Failed to allocate LACP packet from pool");
 569                 return;
 570         }
 571
 572         lacp_pkt->data_len = sizeof(*hdr);
 573         lacp_pkt->pkt_len = sizeof(*hdr);
 574
 575         hdr = rte_pktmbuf_mtod(lacp_pkt, struct lacpdu_header *);
 576
 577         /* Source and destination MAC */
 578         ether_addr_copy(&lacp_mac_addr, &hdr->eth_hdr.d_addr);
 579         rte_eth_macaddr_get(slave_id, &hdr->eth_hdr.s_addr);
 580         hdr->eth_hdr.ether_type = rte_cpu_to_be_16(ETHER_TYPE_SLOW);
 581
 582         lacpdu = &hdr->lacpdu;
 583         memset(lacpdu, 0, sizeof(*lacpdu));
 584
 585         /* Initialize LACP part */
 586         lacpdu->subtype = SLOW_SUBTYPE_LACP;
 587         lacpdu->version_number = 1;
 588
 589         /* ACTOR */
 590         lacpdu->actor.tlv_type_info = TLV_TYPE_ACTOR_INFORMATION;
 591         lacpdu->actor.info_length = sizeof(struct lacpdu_actor_partner_params);
 592         memcpy(&hdr->lacpdu.actor.port_params, &port->actor,
 593                         sizeof(port->actor));
 594         agg = &mode_8023ad_ports[port->aggregator_port_id];
 595         ether_addr_copy(&agg->actor.system, &hdr->lacpdu.actor.port_params.system);
 596         lacpdu->actor.state = port->actor_state;
 597
 598         /* PARTNER */
 599         lacpdu->partner.tlv_type_info = TLV_TYPE_PARTNER_INFORMATION;
 600         lacpdu->partner.info_length = sizeof(struct lacpdu_actor_partner_params);
 601         memcpy(&lacpdu->partner.port_params, &port->partner,
 602                         sizeof(struct port_params));
 603         lacpdu->partner.state = port->partner_state;
 604
 605         /* Other fields */
 606         lacpdu->tlv_type_collector_info = TLV_TYPE_COLLECTOR_INFORMATION;
 607         lacpdu->collector_info_length = 0x10;
 608         lacpdu->collector_max_delay = 0;
 609
 610         lacpdu->tlv_type_terminator = TLV_TYPE_TERMINATOR_INFORMATION;
 611         lacpdu->terminator_length = 0;
 612
 613         MODE4_DEBUG("Sending LACP frame\n");
 614         BOND_PRINT_LACP(lacpdu);
 615
 616         if (internals->mode4.dedicated_queues.enabled == 0) {
 617                 int retval = rte_ring_enqueue(port->tx_ring, lacp_pkt);
 618                 if (retval != 0) {
 619                         /* If TX ring full, drop packet and free message.
 620                            Retransmission will happen in next function call. */
 621                         rte_pktmbuf_free(lacp_pkt);
 622                         set_warning_flags(port, WRN_TX_QUEUE_FULL);
 623                         return;
 624                 }
 625         } else {
 626                 uint16_t pkts_sent = rte_eth_tx_burst(slave_id,
 627                                 internals->mode4.dedicated_queues.tx_qid,
 628                                 &lacp_pkt, 1);
 629                 if (pkts_sent != 1) {
 630                         rte_pktmbuf_free(lacp_pkt);
 631                         set_warning_flags(port, WRN_TX_QUEUE_FULL);
 632                         return;
 633                 }
 634         }
 635
 636
 637         timer_set(&port->tx_machine_timer, internals->mode4.tx_period_timeout);
 638         SM_FLAG_CLR(port, NTT);
 639 }
 640
 641 static uint8_t
 642 max_index(uint64_t *a, int n)
 643 {
 644         if (n <= 0)
 645                 return -1;
 646
 647         int i, max_i = 0;
 648         uint64_t max = a[0];
 649
 650         for (i = 1; i < n; ++i) {
 651                 if (a[i] > max) {
 652                         max = a[i];
 653                         max_i = i;
 654                 }
 655         }
 656
 657         return max_i;
 658 }
 659
 660 /**
 661  * Function assigns port to aggregator.
 662  *
 663  * @param bond_dev_private      Pointer to bond_dev_private structure.
 664  * @param port_pos                      Port to assign.
 665  */
 666 static void
 667 selection_logic(struct bond_dev_private *internals, uint8_t slave_id)
 668 {
 669         struct port *agg, *port;
 670         uint16_t slaves_count, new_agg_id, i, j = 0;
 671         uint16_t *slaves;
 672         uint64_t agg_bandwidth[8] = {0};
 673         uint64_t agg_count[8] = {0};
 674         uint16_t default_slave = 0;
 675         uint8_t mode_count_id, mode_band_id;
 676         struct rte_eth_link link_info;
 677
 678         slaves = internals->active_slaves;
 679         slaves_count = internals->active_slave_count;
 680         port = &mode_8023ad_ports[slave_id];
 681
 682         /* Search for aggregator suitable for this port */
 683         for (i = 0; i < slaves_count; ++i) {
 684                 agg = &mode_8023ad_ports[slaves[i]];
 685                 /* Skip ports that are not aggreagators */
 686                 if (agg->aggregator_port_id != slaves[i])
 687                         continue;
 688
 689                 agg_count[agg->aggregator_port_id] += 1;
 690                 rte_eth_link_get_nowait(slaves[i], &link_info);
 691                 agg_bandwidth[agg->aggregator_port_id] += link_info.link_speed;
 692
 693                 /* Actors system ID is not checked since all slave device have the same
 694                  * ID (MAC address). */
 695                 if ((agg->actor.key == port->actor.key &&
 696                         agg->partner.system_priority == port->partner.system_priority &&
 697                         is_same_ether_addr(&agg->partner.system, &port->partner.system) == 1
 698                         && (agg->partner.key == port->partner.key)) &&
 699                         is_zero_ether_addr(&port->partner.system) != 1 &&
 700                         (agg->actor.key &
 701                                 rte_cpu_to_be_16(BOND_LINK_FULL_DUPLEX_KEY)) != 0) {
 702
 703                         if (j == 0)
 704                                 default_slave = i;
 705                         j++;
 706                 }
 707         }
 708
 709         switch (internals->mode4.agg_selection) {
 710         case AGG_COUNT:
 711                 mode_count_id = max_index(
 712                                 (uint64_t *)agg_count, slaves_count);
 713                 new_agg_id = mode_count_id;
 714                 break;
 715         case AGG_BANDWIDTH:
 716                 mode_band_id = max_index(
 717                                 (uint64_t *)agg_bandwidth, slaves_count);
 718                 new_agg_id = mode_band_id;
 719                 break;
 720         case AGG_STABLE:
 721                 if (default_slave == slaves_count)
 722                         new_agg_id = slave_id;
 723                 else
 724                         new_agg_id = slaves[default_slave];
 725                 break;
 726         default:
 727                 if (default_slave == slaves_count)
 728                         new_agg_id = slave_id;
 729                 else
 730                         new_agg_id = slaves[default_slave];
 731                 break;
 732         }
 733
 734         if (new_agg_id != port->aggregator_port_id) {
 735                 port->aggregator_port_id = new_agg_id;
 736
 737                 MODE4_DEBUG("-> SELECTED: ID=%3u\n"
 738                         "\t%s aggregator ID=%3u\n",
 739                         port->aggregator_port_id,
 740                         port->aggregator_port_id == slave_id ?
 741                                 "aggregator not found, using default" : "aggregator found",
 742                         port->aggregator_port_id);
 743         }
 744
 745         port->selected = SELECTED;
 746 }
 747
 748 /* Function maps DPDK speed to bonding speed stored in key field */
 749 static uint16_t
 750 link_speed_key(uint16_t speed) {
 751         uint16_t key_speed;
 752
 753         switch (speed) {
 754         case ETH_SPEED_NUM_NONE:
 755                 key_speed = 0x00;
 756                 break;
 757         case ETH_SPEED_NUM_10M:
 758                 key_speed = BOND_LINK_SPEED_KEY_10M;
 759                 break;
 760         case ETH_SPEED_NUM_100M:
 761                 key_speed = BOND_LINK_SPEED_KEY_100M;
 762                 break;
 763         case ETH_SPEED_NUM_1G:
 764                 key_speed = BOND_LINK_SPEED_KEY_1000M;
 765                 break;
 766         case ETH_SPEED_NUM_10G:
 767                 key_speed = BOND_LINK_SPEED_KEY_10G;
 768                 break;
 769         case ETH_SPEED_NUM_20G:
 770                 key_speed = BOND_LINK_SPEED_KEY_20G;
 771                 break;
 772         case ETH_SPEED_NUM_40G:
 773                 key_speed = BOND_LINK_SPEED_KEY_40G;
 774                 break;
 775         default:
 776                 /* Unknown speed*/
 777                 key_speed = 0xFFFF;
 778         }
 779
 780         return key_speed;
 781 }
 782
 783 static void
 784 rx_machine_update(struct bond_dev_private *internals, uint8_t slave_id,
 785                 struct rte_mbuf *lacp_pkt) {
 786         struct lacpdu_header *lacp;
 787
 788         if (lacp_pkt != NULL) {
 789                 lacp = rte_pktmbuf_mtod(lacp_pkt, struct lacpdu_header *);
 790                 RTE_ASSERT(lacp->lacpdu.subtype == SLOW_SUBTYPE_LACP);
 791
 792                 /* This is LACP frame so pass it to rx_machine */
 793                 rx_machine(internals, slave_id, &lacp->lacpdu);
 794                 rte_pktmbuf_free(lacp_pkt);
 795         } else
 796                 rx_machine(internals, slave_id, NULL);
 797 }
 798
 799 static void
 800 bond_mode_8023ad_periodic_cb(void *arg)
 801 {
 802         struct rte_eth_dev *bond_dev = arg;
 803         struct bond_dev_private *internals = bond_dev->data->dev_private;
 804         struct port *port;
 805         struct rte_eth_link link_info;
 806         struct ether_addr slave_addr;
 807         struct rte_mbuf *lacp_pkt = NULL;
 808
 809         uint8_t i, slave_id;
 810
 811
 812         /* Update link status on each port */
 813         for (i = 0; i < internals->active_slave_count; i++) {
 814                 uint16_t key;
 815
 816                 slave_id = internals->active_slaves[i];
 817                 rte_eth_link_get_nowait(slave_id, &link_info);
 818                 rte_eth_macaddr_get(slave_id, &slave_addr);
 819
 820                 if (link_info.link_status != 0) {
 821                         key = link_speed_key(link_info.link_speed) << 1;
 822                         if (link_info.link_duplex == ETH_LINK_FULL_DUPLEX)
 823                                 key |= BOND_LINK_FULL_DUPLEX_KEY;
 824                 } else
 825                         key = 0;
 826
 827                 port = &mode_8023ad_ports[slave_id];
 828
 829                 key = rte_cpu_to_be_16(key);
 830                 if (key != port->actor.key) {
 831                         if (!(key & rte_cpu_to_be_16(BOND_LINK_FULL_DUPLEX_KEY)))
 832                                 set_warning_flags(port, WRN_NOT_LACP_CAPABLE);
 833
 834                         port->actor.key = key;
 835                         SM_FLAG_SET(port, NTT);
 836                 }
 837
 838                 if (!is_same_ether_addr(&port->actor.system, &slave_addr)) {
 839                         ether_addr_copy(&slave_addr, &port->actor.system);
 840                         if (port->aggregator_port_id == slave_id)
 841                                 SM_FLAG_SET(port, NTT);
 842                 }
 843         }
 844
 845         for (i = 0; i < internals->active_slave_count; i++) {
 846                 slave_id = internals->active_slaves[i];
 847                 port = &mode_8023ad_ports[slave_id];
 848
 849                 if ((port->actor.key &
 850                                 rte_cpu_to_be_16(BOND_LINK_FULL_DUPLEX_KEY)) == 0) {
 851
 852                         SM_FLAG_SET(port, BEGIN);
 853
 854                         /* LACP is disabled on half duples or link is down */
 855                         if (SM_FLAG(port, LACP_ENABLED)) {
 856                                 /* If port was enabled set it to BEGIN state */
 857                                 SM_FLAG_CLR(port, LACP_ENABLED);
 858                                 ACTOR_STATE_CLR(port, DISTRIBUTING);
 859                                 ACTOR_STATE_CLR(port, COLLECTING);
 860                         }
 861
 862                         /* Skip this port processing */
 863                         continue;
 864                 }
 865
 866                 SM_FLAG_SET(port, LACP_ENABLED);
 867
 868                 if (internals->mode4.dedicated_queues.enabled == 0) {
 869                         /* Find LACP packet to this port. Do not check subtype,
 870                          * it is done in function that queued packet
 871                          */
 872                         int retval = rte_ring_dequeue(port->rx_ring,
 873                                         (void **)&lacp_pkt);
 874
 875                         if (retval != 0)
 876                                 lacp_pkt = NULL;
 877
 878                         rx_machine_update(internals, slave_id, lacp_pkt);
 879                 } else {
 880                         uint16_t rx_count = rte_eth_rx_burst(slave_id,
 881                                         internals->mode4.dedicated_queues.rx_qid,
 882                                         &lacp_pkt, 1);
 883
 884                         if (rx_count == 1)
 885                                 bond_mode_8023ad_handle_slow_pkt(internals,
 886                                                 slave_id, lacp_pkt);
 887                         else
 888                                 rx_machine_update(internals, slave_id, NULL);
 889                 }
 890
 891                 periodic_machine(internals, slave_id);
 892                 mux_machine(internals, slave_id);
 893                 tx_machine(internals, slave_id);
 894                 selection_logic(internals, slave_id);
 895
 896                 SM_FLAG_CLR(port, BEGIN);
 897                 show_warnings(slave_id);
 898         }
 899
 900         rte_eal_alarm_set(internals->mode4.update_timeout_us,
 901                         bond_mode_8023ad_periodic_cb, arg);
 902 }
 903
 904 void
 905 bond_mode_8023ad_activate_slave(struct rte_eth_dev *bond_dev,
 906                                 uint16_t slave_id)
 907 {
 908         struct bond_dev_private *internals = bond_dev->data->dev_private;
 909
 910         struct port *port = &mode_8023ad_ports[slave_id];
 911         struct port_params initial = {
 912                         .system = { { 0 } },
 913                         .system_priority = rte_cpu_to_be_16(0xFFFF),
 914                         .key = rte_cpu_to_be_16(BOND_LINK_FULL_DUPLEX_KEY),
 915                         .port_priority = rte_cpu_to_be_16(0x00FF),
 916                         .port_number = 0,
 917         };
 918
 919         char mem_name[RTE_ETH_NAME_MAX_LEN];
 920         int socket_id;
 921         unsigned element_size;
 922         uint32_t total_tx_desc;
 923         struct bond_tx_queue *bd_tx_q;
 924         uint16_t q_id;
 925
 926         /* Given slave mus not be in active list */
 927         RTE_ASSERT(find_slave_by_id(internals->active_slaves,
 928         internals->active_slave_count, slave_id) == internals->active_slave_count);
 929         RTE_SET_USED(internals); /* used only for assert when enabled */
 930
 931         memcpy(&port->actor, &initial, sizeof(struct port_params));
 932         /* Standard requires that port ID must be grater than 0.
 933          * Add 1 do get corresponding port_number */
 934         port->actor.port_number = rte_cpu_to_be_16(slave_id + 1);
 935
 936         memcpy(&port->partner, &initial, sizeof(struct port_params));
 937
 938         /* default states */
 939         port->actor_state = STATE_AGGREGATION | STATE_LACP_ACTIVE | STATE_DEFAULTED;
 940         port->partner_state = STATE_LACP_ACTIVE | STATE_AGGREGATION;
 941         port->sm_flags = SM_FLAGS_BEGIN;
 942
 943         /* use this port as agregator */
 944         port->aggregator_port_id = slave_id;
 945         rte_eth_promiscuous_enable(slave_id);
 946
 947         timer_cancel(&port->warning_timer);
 948
 949         if (port->mbuf_pool != NULL)
 950                 return;
 951
 952         RTE_ASSERT(port->rx_ring == NULL);
 953         RTE_ASSERT(port->tx_ring == NULL);
 954
 955         socket_id = rte_eth_dev_socket_id(slave_id);
 956         if (socket_id == (int)LCORE_ID_ANY)
 957                 socket_id = rte_socket_id();
 958
 959         element_size = sizeof(struct slow_protocol_frame) +
 960                                 RTE_PKTMBUF_HEADROOM;
 961
 962         /* The size of the mempool should be at least:
 963          * the sum of the TX descriptors + BOND_MODE_8023AX_SLAVE_TX_PKTS */
 964         total_tx_desc = BOND_MODE_8023AX_SLAVE_TX_PKTS;
 965         for (q_id = 0; q_id < bond_dev->data->nb_tx_queues; q_id++) {
 966                 bd_tx_q = (struct bond_tx_queue*)bond_dev->data->tx_queues[q_id];
 967                 total_tx_desc += bd_tx_q->nb_tx_desc;
 968         }
 969
 970         snprintf(mem_name, RTE_DIM(mem_name), "slave_port%u_pool", slave_id);
 971         port->mbuf_pool = rte_pktmbuf_pool_create(mem_name, total_tx_desc,
 972                 RTE_MEMPOOL_CACHE_MAX_SIZE >= 32 ?
 973                         32 : RTE_MEMPOOL_CACHE_MAX_SIZE,
 974                 0, element_size, socket_id);
 975
 976         /* Any memory allocation failure in initialization is critical because
 977          * resources can't be free, so reinitialization is impossible. */
 978         if (port->mbuf_pool == NULL) {
 979                 rte_panic("Slave %u: Failed to create memory pool '%s': %s\n",
 980                         slave_id, mem_name, rte_strerror(rte_errno));
 981         }
 982
 983         snprintf(mem_name, RTE_DIM(mem_name), "slave_%u_rx", slave_id);
 984         port->rx_ring = rte_ring_create(mem_name,
 985                         rte_align32pow2(BOND_MODE_8023AX_SLAVE_RX_PKTS), socket_id, 0);
 986
 987         if (port->rx_ring == NULL) {
 988                 rte_panic("Slave %u: Failed to create rx ring '%s': %s\n", slave_id,
 989                         mem_name, rte_strerror(rte_errno));
 990         }
 991
 992         /* TX ring is at least one pkt longer to make room for marker packet. */
 993         snprintf(mem_name, RTE_DIM(mem_name), "slave_%u_tx", slave_id);
 994         port->tx_ring = rte_ring_create(mem_name,
 995                         rte_align32pow2(BOND_MODE_8023AX_SLAVE_TX_PKTS + 1), socket_id, 0);
 996
 997         if (port->tx_ring == NULL) {
 998                 rte_panic("Slave %u: Failed to create tx ring '%s': %s\n", slave_id,
 999                         mem_name, rte_strerror(rte_errno));
1000         }
1001 }
1002
1003 int
1004 bond_mode_8023ad_deactivate_slave(struct rte_eth_dev *bond_dev __rte_unused,
1005                 uint16_t slave_id)
1006 {
1007         void *pkt = NULL;
1008         struct port *port = NULL;
1009         uint8_t old_partner_state;
1010
1011         port = &mode_8023ad_ports[slave_id];
1012
1013         ACTOR_STATE_CLR(port, AGGREGATION);
1014         port->selected = UNSELECTED;
1015
1016         old_partner_state = port->partner_state;
1017         record_default(port);
1018
1019         /* If partner timeout state changes then disable timer */
1020         if (!((old_partner_state ^ port->partner_state) &
1021                         STATE_LACP_SHORT_TIMEOUT))
1022                 timer_cancel(&port->current_while_timer);
1023
1024         PARTNER_STATE_CLR(port, AGGREGATION);
1025         ACTOR_STATE_CLR(port, EXPIRED);
1026
1027         /* flush rx/tx rings */
1028         while (rte_ring_dequeue(port->rx_ring, &pkt) == 0)
1029                 rte_pktmbuf_free((struct rte_mbuf *)pkt);
1030
1031         while (rte_ring_dequeue(port->tx_ring, &pkt) == 0)
1032                         rte_pktmbuf_free((struct rte_mbuf *)pkt);
1033         return 0;
1034 }
1035
1036 void
1037 bond_mode_8023ad_mac_address_update(struct rte_eth_dev *bond_dev)
1038 {
1039         struct bond_dev_private *internals = bond_dev->data->dev_private;
1040         struct ether_addr slave_addr;
1041         struct port *slave, *agg_slave;
1042         uint16_t slave_id, i, j;
1043
1044         bond_mode_8023ad_stop(bond_dev);
1045
1046         for (i = 0; i < internals->active_slave_count; i++) {
1047                 slave_id = internals->active_slaves[i];
1048                 slave = &mode_8023ad_ports[slave_id];
1049                 rte_eth_macaddr_get(slave_id, &slave_addr);
1050
1051                 if (is_same_ether_addr(&slave_addr, &slave->actor.system))
1052                         continue;
1053
1054                 ether_addr_copy(&slave_addr, &slave->actor.system);
1055                 /* Do nothing if this port is not an aggregator. In other case
1056                  * Set NTT flag on every port that use this aggregator. */
1057                 if (slave->aggregator_port_id != slave_id)
1058                         continue;
1059
1060                 for (j = 0; j < internals->active_slave_count; j++) {
1061                         agg_slave = &mode_8023ad_ports[internals->active_slaves[j]];
1062                         if (agg_slave->aggregator_port_id == slave_id)
1063                                 SM_FLAG_SET(agg_slave, NTT);
1064                 }
1065         }
1066
1067         if (bond_dev->data->dev_started)
1068                 bond_mode_8023ad_start(bond_dev);
1069 }
1070
1071 static void
1072 bond_mode_8023ad_conf_get(struct rte_eth_dev *dev,
1073                 struct rte_eth_bond_8023ad_conf *conf)
1074 {
1075         struct bond_dev_private *internals = dev->data->dev_private;
1076         struct mode8023ad_private *mode4 = &internals->mode4;
1077         uint64_t ms_ticks = rte_get_tsc_hz() / 1000;
1078
1079         conf->fast_periodic_ms = mode4->fast_periodic_timeout / ms_ticks;
1080         conf->slow_periodic_ms = mode4->slow_periodic_timeout / ms_ticks;
1081         conf->short_timeout_ms = mode4->short_timeout / ms_ticks;
1082         conf->long_timeout_ms = mode4->long_timeout / ms_ticks;
1083         conf->aggregate_wait_timeout_ms = mode4->aggregate_wait_timeout / ms_ticks;
1084         conf->tx_period_ms = mode4->tx_period_timeout / ms_ticks;
1085         conf->update_timeout_ms = mode4->update_timeout_us / 1000;
1086         conf->rx_marker_period_ms = mode4->rx_marker_timeout / ms_ticks;
1087         conf->slowrx_cb = mode4->slowrx_cb;
1088         conf->agg_selection = mode4->agg_selection;
1089 }
1090
1091 static void
1092 bond_mode_8023ad_conf_get_default(struct rte_eth_bond_8023ad_conf *conf)
1093 {
1094         conf->fast_periodic_ms = BOND_8023AD_FAST_PERIODIC_MS;
1095         conf->slow_periodic_ms = BOND_8023AD_SLOW_PERIODIC_MS;
1096         conf->short_timeout_ms = BOND_8023AD_SHORT_TIMEOUT_MS;
1097         conf->long_timeout_ms = BOND_8023AD_LONG_TIMEOUT_MS;
1098         conf->aggregate_wait_timeout_ms = BOND_8023AD_AGGREGATE_WAIT_TIMEOUT_MS;
1099         conf->tx_period_ms = BOND_8023AD_TX_MACHINE_PERIOD_MS;
1100         conf->rx_marker_period_ms = BOND_8023AD_RX_MARKER_PERIOD_MS;
1101         conf->update_timeout_ms = BOND_MODE_8023AX_UPDATE_TIMEOUT_MS;
1102         conf->slowrx_cb = NULL;
1103         conf->agg_selection = AGG_STABLE;
1104 }
1105
1106 static void
1107 bond_mode_8023ad_conf_assign(struct mode8023ad_private *mode4,
1108                 struct rte_eth_bond_8023ad_conf *conf)
1109 {
1110         uint64_t ms_ticks = rte_get_tsc_hz() / 1000;
1111
1112         mode4->fast_periodic_timeout = conf->fast_periodic_ms * ms_ticks;
1113         mode4->slow_periodic_timeout = conf->slow_periodic_ms * ms_ticks;
1114         mode4->short_timeout = conf->short_timeout_ms * ms_ticks;
1115         mode4->long_timeout = conf->long_timeout_ms * ms_ticks;
1116         mode4->aggregate_wait_timeout = conf->aggregate_wait_timeout_ms * ms_ticks;
1117         mode4->tx_period_timeout = conf->tx_period_ms * ms_ticks;
1118         mode4->rx_marker_timeout = conf->rx_marker_period_ms * ms_ticks;
1119         mode4->update_timeout_us = conf->update_timeout_ms * 1000;
1120
1121         mode4->dedicated_queues.enabled = 0;
1122         mode4->dedicated_queues.rx_qid = UINT16_MAX;
1123         mode4->dedicated_queues.tx_qid = UINT16_MAX;
1124 }
1125
1126 void
1127 bond_mode_8023ad_setup(struct rte_eth_dev *dev,
1128                 struct rte_eth_bond_8023ad_conf *conf)
1129 {
1130         struct rte_eth_bond_8023ad_conf def_conf;
1131         struct bond_dev_private *internals = dev->data->dev_private;
1132         struct mode8023ad_private *mode4 = &internals->mode4;
1133
1134         if (conf == NULL) {
1135                 conf = &def_conf;
1136                 bond_mode_8023ad_conf_get_default(conf);
1137         }
1138
1139         bond_mode_8023ad_stop(dev);
1140         bond_mode_8023ad_conf_assign(mode4, conf);
1141         mode4->slowrx_cb = conf->slowrx_cb;
1142         mode4->agg_selection = AGG_STABLE;
1143
1144         if (dev->data->dev_started)
1145                 bond_mode_8023ad_start(dev);
1146 }
1147
1148 int
1149 bond_mode_8023ad_enable(struct rte_eth_dev *bond_dev)
1150 {
1151         struct bond_dev_private *internals = bond_dev->data->dev_private;
1152         uint8_t i;
1153
1154         for (i = 0; i < internals->active_slave_count; i++)
1155                 bond_mode_8023ad_activate_slave(bond_dev,
1156                                 internals->active_slaves[i]);
1157
1158         return 0;
1159 }
1160
1161 int
1162 bond_mode_8023ad_start(struct rte_eth_dev *bond_dev)
1163 {
1164         struct bond_dev_private *internals = bond_dev->data->dev_private;
1165         struct mode8023ad_private *mode4 = &internals->mode4;
1166         static const uint64_t us = BOND_MODE_8023AX_UPDATE_TIMEOUT_MS * 1000;
1167
1168         if (mode4->slowrx_cb)
1169                 return rte_eal_alarm_set(us, &bond_mode_8023ad_ext_periodic_cb,
1170                                          bond_dev);
1171
1172         return rte_eal_alarm_set(us, &bond_mode_8023ad_periodic_cb, bond_dev);
1173 }
1174
1175 void
1176 bond_mode_8023ad_stop(struct rte_eth_dev *bond_dev)
1177 {
1178         struct bond_dev_private *internals = bond_dev->data->dev_private;
1179         struct mode8023ad_private *mode4 = &internals->mode4;
1180
1181         if (mode4->slowrx_cb) {
1182                 rte_eal_alarm_cancel(&bond_mode_8023ad_ext_periodic_cb,
1183                                      bond_dev);
1184                 return;
1185         }
1186         rte_eal_alarm_cancel(&bond_mode_8023ad_periodic_cb, bond_dev);
1187 }
1188
1189 void
1190 bond_mode_8023ad_handle_slow_pkt(struct bond_dev_private *internals,
1191                                   uint16_t slave_id, struct rte_mbuf *pkt)
1192 {
1193         struct mode8023ad_private *mode4 = &internals->mode4;
1194         struct port *port = &mode_8023ad_ports[slave_id];
1195         struct marker_header *m_hdr;
1196         uint64_t marker_timer, old_marker_timer;
1197         int retval;
1198         uint8_t wrn, subtype;
1199         /* If packet is a marker, we send response now by reusing given packet
1200          * and update only source MAC, destination MAC is multicast so don't
1201          * update it. Other frames will be handled later by state machines */
1202         subtype = rte_pktmbuf_mtod(pkt,
1203                         struct slow_protocol_frame *)->slow_protocol.subtype;
1204
1205         if (subtype == SLOW_SUBTYPE_MARKER) {
1206                 m_hdr = rte_pktmbuf_mtod(pkt, struct marker_header *);
1207
1208                 if (likely(m_hdr->marker.tlv_type_marker != MARKER_TLV_TYPE_INFO)) {
1209                         wrn = WRN_UNKNOWN_MARKER_TYPE;
1210                         goto free_out;
1211                 }
1212
1213                 /* Setup marker timer. Do it in loop in case concurrent access. */
1214                 do {
1215                         old_marker_timer = port->rx_marker_timer;
1216                         if (!timer_is_expired(&old_marker_timer)) {
1217                                 wrn = WRN_RX_MARKER_TO_FAST;
1218                                 goto free_out;
1219                         }
1220
1221                         timer_set(&marker_timer, mode4->rx_marker_timeout);
1222                         retval = rte_atomic64_cmpset(&port->rx_marker_timer,
1223                                 old_marker_timer, marker_timer);
1224                 } while (unlikely(retval == 0));
1225
1226                 m_hdr->marker.tlv_type_marker = MARKER_TLV_TYPE_RESP;
1227                 rte_eth_macaddr_get(slave_id, &m_hdr->eth_hdr.s_addr);
1228
1229                 if (internals->mode4.dedicated_queues.enabled == 0) {
1230                         int retval = rte_ring_enqueue(port->tx_ring, pkt);
1231                         if (retval != 0) {
1232                                 /* reset timer */
1233                                 port->rx_marker_timer = 0;
1234                                 wrn = WRN_TX_QUEUE_FULL;
1235                                 goto free_out;
1236                         }
1237                 } else {
1238                         /* Send packet directly to the slow queue */
1239                         uint16_t tx_count = rte_eth_tx_burst(slave_id,
1240                                         internals->mode4.dedicated_queues.tx_qid,
1241                                         &pkt, 1);
1242                         if (tx_count != 1) {
1243                                 /* reset timer */
1244                                 port->rx_marker_timer = 0;
1245                                 wrn = WRN_TX_QUEUE_FULL;
1246                                 goto free_out;
1247                         }
1248                 }
1249         } else if (likely(subtype == SLOW_SUBTYPE_LACP)) {
1250                 if (internals->mode4.dedicated_queues.enabled == 0) {
1251                         int retval = rte_ring_enqueue(port->rx_ring, pkt);
1252                         if (retval != 0) {
1253                                 /* If RX fing full free lacpdu message and drop packet */
1254                                 wrn = WRN_RX_QUEUE_FULL;
1255                                 goto free_out;
1256                         }
1257                 } else
1258                         rx_machine_update(internals, slave_id, pkt);
1259         } else {
1260                 wrn = WRN_UNKNOWN_SLOW_TYPE;
1261                 goto free_out;
1262         }
1263
1264         return;
1265
1266 free_out:
1267         set_warning_flags(port, wrn);
1268         rte_pktmbuf_free(pkt);
1269 }
1270
1271 int
1272 rte_eth_bond_8023ad_conf_get(uint16_t port_id,
1273                 struct rte_eth_bond_8023ad_conf *conf)
1274 {
1275         struct rte_eth_dev *bond_dev;
1276
1277         if (valid_bonded_port_id(port_id) != 0)
1278                 return -EINVAL;
1279
1280         if (conf == NULL)
1281                 return -EINVAL;
1282
1283         bond_dev = &rte_eth_devices[port_id];
1284         bond_mode_8023ad_conf_get(bond_dev, conf);
1285         return 0;
1286 }
1287
1288 int
1289 rte_eth_bond_8023ad_agg_selection_set(uint16_t port_id,
1290                 enum rte_bond_8023ad_agg_selection agg_selection)
1291 {
1292         struct rte_eth_dev *bond_dev;
1293         struct bond_dev_private *internals;
1294         struct mode8023ad_private *mode4;
1295
1296         bond_dev = &rte_eth_devices[port_id];
1297         internals = bond_dev->data->dev_private;
1298
1299         if (valid_bonded_port_id(port_id) != 0)
1300                 return -EINVAL;
1301         if (internals->mode != 4)
1302                 return -EINVAL;
1303
1304         mode4 = &internals->mode4;
1305         if (agg_selection == AGG_COUNT || agg_selection == AGG_BANDWIDTH
1306                         || agg_selection == AGG_STABLE)
1307                 mode4->agg_selection = agg_selection;
1308         return 0;
1309 }
1310
1311 int rte_eth_bond_8023ad_agg_selection_get(uint16_t port_id)
1312 {
1313         struct rte_eth_dev *bond_dev;
1314         struct bond_dev_private *internals;
1315         struct mode8023ad_private *mode4;
1316
1317         bond_dev = &rte_eth_devices[port_id];
1318         internals = bond_dev->data->dev_private;
1319
1320         if (valid_bonded_port_id(port_id) != 0)
1321                 return -EINVAL;
1322         if (internals->mode != 4)
1323                 return -EINVAL;
1324         mode4 = &internals->mode4;
1325
1326         return mode4->agg_selection;
1327 }
1328
1329
1330
1331 static int
1332 bond_8023ad_setup_validate(uint16_t port_id,
1333                 struct rte_eth_bond_8023ad_conf *conf)
1334 {
1335         if (valid_bonded_port_id(port_id) != 0)
1336                 return -EINVAL;
1337
1338         if (conf != NULL) {
1339                 /* Basic sanity check */
1340                 if (conf->slow_periodic_ms == 0 ||
1341                                 conf->fast_periodic_ms >= conf->slow_periodic_ms ||
1342                                 conf->long_timeout_ms == 0 ||
1343                                 conf->short_timeout_ms >= conf->long_timeout_ms ||
1344                                 conf->aggregate_wait_timeout_ms == 0 ||
1345                                 conf->tx_period_ms == 0 ||
1346                                 conf->rx_marker_period_ms == 0 ||
1347                                 conf->update_timeout_ms == 0) {
1348                         RTE_BOND_LOG(ERR, "given mode 4 configuration is invalid");
1349                         return -EINVAL;
1350                 }
1351         }
1352
1353         return 0;
1354 }
1355
1356
1357 int
1358 rte_eth_bond_8023ad_setup(uint16_t port_id,
1359                 struct rte_eth_bond_8023ad_conf *conf)
1360 {
1361         struct rte_eth_dev *bond_dev;
1362         int err;
1363
1364         err = bond_8023ad_setup_validate(port_id, conf);
1365         if (err != 0)
1366                 return err;
1367
1368         bond_dev = &rte_eth_devices[port_id];
1369         bond_mode_8023ad_setup(bond_dev, conf);
1370
1371         return 0;
1372 }
1373
1374
1375
1376
1377
1378 int
1379 rte_eth_bond_8023ad_slave_info(uint16_t port_id, uint16_t slave_id,
1380                 struct rte_eth_bond_8023ad_slave_info *info)
1381 {
1382         struct rte_eth_dev *bond_dev;
1383         struct bond_dev_private *internals;
1384         struct port *port;
1385
1386         if (info == NULL || valid_bonded_port_id(port_id) != 0 ||
1387                         rte_eth_bond_mode_get(port_id) != BONDING_MODE_8023AD)
1388                 return -EINVAL;
1389
1390         bond_dev = &rte_eth_devices[port_id];
1391
1392         internals = bond_dev->data->dev_private;
1393         if (find_slave_by_id(internals->active_slaves,
1394                         internals->active_slave_count, slave_id) ==
1395                                 internals->active_slave_count)
1396                 return -EINVAL;
1397
1398         port = &mode_8023ad_ports[slave_id];
1399         info->selected = port->selected;
1400
1401         info->actor_state = port->actor_state;
1402         rte_memcpy(&info->actor, &port->actor, sizeof(port->actor));
1403
1404         info->partner_state = port->partner_state;
1405         rte_memcpy(&info->partner, &port->partner, sizeof(port->partner));
1406
1407         info->agg_port_id = port->aggregator_port_id;
1408         return 0;
1409 }
1410
1411 static int
1412 bond_8023ad_ext_validate(uint16_t port_id, uint16_t slave_id)
1413 {
1414         struct rte_eth_dev *bond_dev;
1415         struct bond_dev_private *internals;
1416         struct mode8023ad_private *mode4;
1417
1418         if (rte_eth_bond_mode_get(port_id) != BONDING_MODE_8023AD)
1419                 return -EINVAL;
1420
1421         bond_dev = &rte_eth_devices[port_id];
1422
1423         if (!bond_dev->data->dev_started)
1424                 return -EINVAL;
1425
1426         internals = bond_dev->data->dev_private;
1427         if (find_slave_by_id(internals->active_slaves,
1428                         internals->active_slave_count, slave_id) ==
1429                                 internals->active_slave_count)
1430                 return -EINVAL;
1431
1432         mode4 = &internals->mode4;
1433         if (mode4->slowrx_cb == NULL)
1434                 return -EINVAL;
1435
1436         return 0;
1437 }
1438
1439 int
1440 rte_eth_bond_8023ad_ext_collect(uint16_t port_id, uint16_t slave_id,
1441                                 int enabled)
1442 {
1443         struct port *port;
1444         int res;
1445
1446         res = bond_8023ad_ext_validate(port_id, slave_id);
1447         if (res != 0)
1448                 return res;
1449
1450         port = &mode_8023ad_ports[slave_id];
1451
1452         if (enabled)
1453                 ACTOR_STATE_SET(port, COLLECTING);
1454         else
1455                 ACTOR_STATE_CLR(port, COLLECTING);
1456
1457         return 0;
1458 }
1459
1460 int
1461 rte_eth_bond_8023ad_ext_distrib(uint16_t port_id, uint16_t slave_id,
1462                                 int enabled)
1463 {
1464         struct port *port;
1465         int res;
1466
1467         res = bond_8023ad_ext_validate(port_id, slave_id);
1468         if (res != 0)
1469                 return res;
1470
1471         port = &mode_8023ad_ports[slave_id];
1472
1473         if (enabled)
1474                 ACTOR_STATE_SET(port, DISTRIBUTING);
1475         else
1476                 ACTOR_STATE_CLR(port, DISTRIBUTING);
1477
1478         return 0;
1479 }
1480
1481 int
1482 rte_eth_bond_8023ad_ext_distrib_get(uint16_t port_id, uint16_t slave_id)
1483 {
1484         struct port *port;
1485         int err;
1486
1487         err = bond_8023ad_ext_validate(port_id, slave_id);
1488         if (err != 0)
1489                 return err;
1490
1491         port = &mode_8023ad_ports[slave_id];
1492         return ACTOR_STATE(port, DISTRIBUTING);
1493 }
1494
1495 int
1496 rte_eth_bond_8023ad_ext_collect_get(uint16_t port_id, uint16_t slave_id)
1497 {
1498         struct port *port;
1499         int err;
1500
1501         err = bond_8023ad_ext_validate(port_id, slave_id);
1502         if (err != 0)
1503                 return err;
1504
1505         port = &mode_8023ad_ports[slave_id];
1506         return ACTOR_STATE(port, COLLECTING);
1507 }
1508
1509 int
1510 rte_eth_bond_8023ad_ext_slowtx(uint16_t port_id, uint16_t slave_id,
1511                 struct rte_mbuf *lacp_pkt)
1512 {
1513         struct port *port;
1514         int res;
1515
1516         res = bond_8023ad_ext_validate(port_id, slave_id);
1517         if (res != 0)
1518                 return res;
1519
1520         port = &mode_8023ad_ports[slave_id];
1521
1522         if (rte_pktmbuf_pkt_len(lacp_pkt) < sizeof(struct lacpdu_header))
1523                 return -EINVAL;
1524
1525         struct lacpdu_header *lacp;
1526
1527         /* only enqueue LACPDUs */
1528         lacp = rte_pktmbuf_mtod(lacp_pkt, struct lacpdu_header *);
1529         if (lacp->lacpdu.subtype != SLOW_SUBTYPE_LACP)
1530                 return -EINVAL;
1531
1532         MODE4_DEBUG("sending LACP frame\n");
1533
1534         return rte_ring_enqueue(port->tx_ring, lacp_pkt);
1535 }
1536
1537 static void
1538 bond_mode_8023ad_ext_periodic_cb(void *arg)
1539 {
1540         struct rte_eth_dev *bond_dev = arg;
1541         struct bond_dev_private *internals = bond_dev->data->dev_private;
1542         struct mode8023ad_private *mode4 = &internals->mode4;
1543         struct port *port;
1544         void *pkt = NULL;
1545         uint16_t i, slave_id;
1546
1547         for (i = 0; i < internals->active_slave_count; i++) {
1548                 slave_id = internals->active_slaves[i];
1549                 port = &mode_8023ad_ports[slave_id];
1550
1551                 if (rte_ring_dequeue(port->rx_ring, &pkt) == 0) {
1552                         struct rte_mbuf *lacp_pkt = pkt;
1553                         struct lacpdu_header *lacp;
1554
1555                         lacp = rte_pktmbuf_mtod(lacp_pkt,
1556                                                 struct lacpdu_header *);
1557                         RTE_VERIFY(lacp->lacpdu.subtype == SLOW_SUBTYPE_LACP);
1558
1559                         /* This is LACP frame so pass it to rx callback.
1560                          * Callback is responsible for freeing mbuf.
1561                          */
1562                         mode4->slowrx_cb(slave_id, lacp_pkt);
1563                 }
1564         }
1565
1566         rte_eal_alarm_set(internals->mode4.update_timeout_us,
1567                         bond_mode_8023ad_ext_periodic_cb, arg);
1568 }
1569
1570 int
1571 rte_eth_bond_8023ad_dedicated_queues_enable(uint16_t port)
1572 {
1573         int retval = 0;
1574         struct rte_eth_dev *dev = &rte_eth_devices[port];
1575         struct bond_dev_private *internals = (struct bond_dev_private *)
1576                 dev->data->dev_private;
1577
1578         if (check_for_bonded_ethdev(dev) != 0)
1579                 return -1;
1580
1581         if (bond_8023ad_slow_pkt_hw_filter_supported(port) != 0)
1582                 return -1;
1583
1584         /* Device must be stopped to set up slow queue */
1585         if (dev->data->dev_started)
1586                 return -1;
1587
1588         internals->mode4.dedicated_queues.enabled = 1;
1589
1590         bond_ethdev_mode_set(dev, internals->mode);
1591         return retval;
1592 }
1593
1594 int
1595 rte_eth_bond_8023ad_dedicated_queues_disable(uint16_t port)
1596 {
1597         int retval = 0;
1598         struct rte_eth_dev *dev = &rte_eth_devices[port];
1599         struct bond_dev_private *internals = (struct bond_dev_private *)
1600                 dev->data->dev_private;
1601
1602         if (check_for_bonded_ethdev(dev) != 0)
1603                 return -1;
1604
1605         /* Device must be stopped to set up slow queue */
1606         if (dev->data->dev_started)
1607                 return -1;
1608
1609         internals->mode4.dedicated_queues.enabled = 0;
1610
1611         bond_ethdev_mode_set(dev, internals->mode);
1612
1613         return retval;
1614 }