examples/multi_process/l2fwd_fork/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4 #define _GNU_SOURCE
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <unistd.h>
   8 #include <string.h>
   9 #include <stdint.h>
  10 #include <sched.h>
  11 #include <inttypes.h>
  12 #include <sys/types.h>
  13 #include <sys/queue.h>
  14 #include <netinet/in.h>
  15 #include <setjmp.h>
  16 #include <stdarg.h>
  17 #include <ctype.h>
  18 #include <errno.h>
  19 #include <getopt.h>
  20
  21 #include <rte_common.h>
  22 #include <rte_log.h>
  23 #include <rte_memory.h>
  24 #include <rte_memcpy.h>
  25 #include <rte_eal.h>
  26 #include <rte_launch.h>
  27 #include <rte_atomic.h>
  28 #include <rte_spinlock.h>
  29 #include <rte_cycles.h>
  30 #include <rte_prefetch.h>
  31 #include <rte_lcore.h>
  32 #include <rte_per_lcore.h>
  33 #include <rte_branch_prediction.h>
  34 #include <rte_interrupts.h>
  35 #include <rte_random.h>
  36 #include <rte_debug.h>
  37 #include <rte_ether.h>
  38 #include <rte_ethdev.h>
  39 #include <rte_ring.h>
  40 #include <rte_mempool.h>
  41 #include <rte_mbuf.h>
  42 #include <rte_malloc.h>
  43
  44 #include "flib.h"
  45
  46 #define RTE_LOGTYPE_L2FWD RTE_LOGTYPE_USER1
  47 #define MBUF_NAME       "mbuf_pool_%d"
  48 #define MBUF_DATA_SIZE  RTE_MBUF_DEFAULT_BUF_SIZE
  49 #define NB_MBUF   8192
  50 #define RING_MASTER_NAME        "l2fwd_ring_m2s_"
  51 #define RING_SLAVE_NAME         "l2fwd_ring_s2m_"
  52 #define MAX_NAME_LEN    32
  53 /* RECREATE flag indicate needs initialize resource and launch slave_core again */
  54 #define SLAVE_RECREATE_FLAG 0x1
  55 /* RESTART flag indicate needs restart port and send START command again */
  56 #define SLAVE_RESTART_FLAG 0x2
  57 #define INVALID_MAPPING_ID      ((unsigned)LCORE_ID_ANY)
  58 /* Maximum message buffer per slave */
  59 #define NB_CORE_MSGBUF  32
  60 enum l2fwd_cmd{
  61         CMD_START,
  62         CMD_STOP,
  63 };
  64
  65 #define MAX_PKT_BURST 32
  66 #define BURST_TX_DRAIN_US 100 /* TX drain every ~100us */
  67
  68 /*
  69  * Configurable number of RX/TX ring descriptors
  70  */
  71 #define RTE_TEST_RX_DESC_DEFAULT 128
  72 #define RTE_TEST_TX_DESC_DEFAULT 512
  73 static uint16_t nb_rxd = RTE_TEST_RX_DESC_DEFAULT;
  74 static uint16_t nb_txd = RTE_TEST_TX_DESC_DEFAULT;
  75
  76 /* ethernet addresses of ports */
  77 static struct ether_addr l2fwd_ports_eth_addr[RTE_MAX_ETHPORTS];
  78
  79 /* mask of enabled ports */
  80 static uint32_t l2fwd_enabled_port_mask = 0;
  81
  82 /* list of enabled ports */
  83 static uint32_t l2fwd_dst_ports[RTE_MAX_ETHPORTS];
  84
  85 static unsigned int l2fwd_rx_queue_per_lcore = 1;
  86
  87 struct mbuf_table {
  88         unsigned len;
  89         struct rte_mbuf *m_table[MAX_PKT_BURST];
  90 };
  91
  92 #define MAX_RX_QUEUE_PER_LCORE 16
  93 #define MAX_TX_QUEUE_PER_PORT 16
  94 struct lcore_queue_conf {
  95         unsigned n_rx_port;
  96         unsigned rx_port_list[MAX_RX_QUEUE_PER_LCORE];
  97 } __rte_cache_aligned;
  98 struct lcore_queue_conf lcore_queue_conf[RTE_MAX_LCORE];
  99
 100 struct rte_eth_dev_tx_buffer *tx_buffer[RTE_MAX_ETHPORTS];
 101
 102 struct lcore_resource_struct {
 103         int enabled;    /* Only set in case this lcore involved into packet forwarding */
 104         int flags;          /* Set only slave need to restart or recreate */
 105         unsigned lcore_id;  /*  lcore ID */
 106         unsigned pair_id;       /* dependency lcore ID on port */
 107         char ring_name[2][MAX_NAME_LEN];
 108         /* ring[0] for master send cmd, slave read */
 109         /* ring[1] for slave send ack, master read */
 110         struct rte_ring *ring[2];
 111         int port_num;                                   /* Total port numbers */
 112         /* Port id for that lcore to receive packets */
 113         uint16_t port[RTE_MAX_ETHPORTS];
 114 }__attribute__((packed)) __rte_cache_aligned;
 115
 116 static struct lcore_resource_struct lcore_resource[RTE_MAX_LCORE];
 117 static struct rte_mempool *message_pool;
 118 static rte_spinlock_t res_lock = RTE_SPINLOCK_INITIALIZER;
 119 /* use floating processes */
 120 static int float_proc = 0;
 121 /* Save original cpu affinity */
 122 struct cpu_aff_arg{
 123         cpu_set_t set;
 124         size_t size;
 125 }cpu_aff;
 126
 127 static const struct rte_eth_conf port_conf = {
 128         .rxmode = {
 129                 .split_hdr_size = 0,
 130                 .header_split   = 0, /**< Header Split disabled */
 131                 .hw_ip_checksum = 0, /**< IP checksum offload disabled */
 132                 .hw_vlan_filter = 0, /**< VLAN filtering disabled */
 133                 .jumbo_frame    = 0, /**< Jumbo Frame Support disabled */
 134                 .hw_strip_crc   = 1, /**< CRC stripped by hardware */
 135         },
 136         .txmode = {
 137                 .mq_mode = ETH_MQ_TX_NONE,
 138         },
 139 };
 140
 141 static struct rte_mempool * l2fwd_pktmbuf_pool[RTE_MAX_ETHPORTS];
 142
 143 /* Per-port statistics struct */
 144 struct l2fwd_port_statistics {
 145         uint64_t tx;
 146         uint64_t rx;
 147         uint64_t dropped;
 148 } __rte_cache_aligned;
 149 struct l2fwd_port_statistics *port_statistics;
 150 /**
 151  * pointer to lcore ID mapping array, used to return lcore id in case slave
 152  * process exited unexpectedly, use only floating process option applied
 153  **/
 154 unsigned *mapping_id;
 155
 156 /* A tsc-based timer responsible for triggering statistics printout */
 157 #define TIMER_MILLISECOND 2000000ULL /* around 1ms at 2 Ghz */
 158 #define MAX_TIMER_PERIOD 86400 /* 1 day max */
 159 static int64_t timer_period = 10 * TIMER_MILLISECOND * 1000; /* default period is 10 seconds */
 160
 161 static int l2fwd_launch_one_lcore(void *dummy);
 162
 163 /* Print out statistics on packets dropped */
 164 static void
 165 print_stats(void)
 166 {
 167         uint64_t total_packets_dropped, total_packets_tx, total_packets_rx;
 168         unsigned portid;
 169
 170         total_packets_dropped = 0;
 171         total_packets_tx = 0;
 172         total_packets_rx = 0;
 173
 174         const char clr[] = { 27, '[', '2', 'J', '\0' };
 175         const char topLeft[] = { 27, '[', '1', ';', '1', 'H','\0' };
 176
 177                 /* Clear screen and move to top left */
 178         printf("%s%s", clr, topLeft);
 179
 180         printf("\nPort statistics ====================================");
 181
 182         for (portid = 0; portid < RTE_MAX_ETHPORTS; portid++) {
 183                 /* skip disabled ports */
 184                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
 185                         continue;
 186                 printf("\nStatistics for port %u ------------------------------"
 187                            "\nPackets sent: %24"PRIu64
 188                            "\nPackets received: %20"PRIu64
 189                            "\nPackets dropped: %21"PRIu64,
 190                            portid,
 191                            port_statistics[portid].tx,
 192                            port_statistics[portid].rx,
 193                            port_statistics[portid].dropped);
 194
 195                 total_packets_dropped += port_statistics[portid].dropped;
 196                 total_packets_tx += port_statistics[portid].tx;
 197                 total_packets_rx += port_statistics[portid].rx;
 198         }
 199         printf("\nAggregate statistics ==============================="
 200                    "\nTotal packets sent: %18"PRIu64
 201                    "\nTotal packets received: %14"PRIu64
 202                    "\nTotal packets dropped: %15"PRIu64,
 203                    total_packets_tx,
 204                    total_packets_rx,
 205                    total_packets_dropped);
 206         printf("\n====================================================\n");
 207 }
 208
 209 static int
 210 clear_cpu_affinity(void)
 211 {
 212         int s;
 213
 214         s = sched_setaffinity(0, cpu_aff.size, &cpu_aff.set);
 215         if (s != 0) {
 216                 printf("sched_setaffinity failed:%s\n", strerror(errno));
 217                 return -1;
 218         }
 219
 220         return 0;
 221 }
 222
 223 static int
 224 get_cpu_affinity(void)
 225 {
 226         int s;
 227
 228         cpu_aff.size = sizeof(cpu_set_t);
 229         CPU_ZERO(&cpu_aff.set);
 230
 231         s = sched_getaffinity(0, cpu_aff.size, &cpu_aff.set);
 232         if (s != 0) {
 233                 printf("sched_getaffinity failed:%s\n", strerror(errno));
 234                 return -1;
 235         }
 236
 237         return 0;
 238 }
 239
 240 /**
 241  * This fnciton demonstrates the approach to create ring in first instance
 242  * or re-attach an existed ring in later instance.
 243  **/
 244 static struct rte_ring *
 245 create_ring(const char *name, unsigned count,
 246                                         int socket_id,unsigned flags)
 247 {
 248         struct rte_ring *ring;
 249
 250         if (name == NULL)
 251                 return NULL;
 252
 253         /* If already create, just attached it */
 254         if (likely((ring = rte_ring_lookup(name)) != NULL))
 255                 return ring;
 256
 257         /* First call it, create one */
 258         return rte_ring_create(name, count, socket_id, flags);
 259 }
 260
 261 /* Malloc with rte_malloc on structures that shared by master and slave */
 262 static int
 263 l2fwd_malloc_shared_struct(void)
 264 {
 265         port_statistics = rte_zmalloc("port_stat",
 266                                                 sizeof(struct l2fwd_port_statistics) * RTE_MAX_ETHPORTS,
 267                                                 0);
 268         if (port_statistics == NULL)
 269                 return -1;
 270
 271         /* allocate  mapping_id array */
 272         if (float_proc) {
 273                 int i;
 274                 mapping_id = rte_malloc("mapping_id", sizeof(unsigned) * RTE_MAX_LCORE,
 275                                                                 0);
 276
 277                 if (mapping_id == NULL)
 278                         return -1;
 279
 280                 for (i = 0 ;i < RTE_MAX_LCORE; i++)
 281                         mapping_id[i] = INVALID_MAPPING_ID;
 282         }
 283         return 0;
 284 }
 285
 286 /* Create ring which used for communicate among master and slave */
 287 static int
 288 create_ms_ring(unsigned slaveid)
 289 {
 290         unsigned flag = RING_F_SP_ENQ | RING_F_SC_DEQ;
 291         struct lcore_resource_struct *res = &lcore_resource[slaveid];
 292         unsigned socketid = rte_socket_id();
 293
 294         /* Always assume create ring on master socket_id */
 295         /* Default only create a ring size 32 */
 296         snprintf(res->ring_name[0], MAX_NAME_LEN, "%s%u",
 297                         RING_MASTER_NAME, slaveid);
 298         if ((res->ring[0] = create_ring(res->ring_name[0], NB_CORE_MSGBUF,
 299                                 socketid, flag)) == NULL) {
 300                 printf("Create m2s ring %s failed\n", res->ring_name[0]);
 301                 return -1;
 302         }
 303
 304         snprintf(res->ring_name[1], MAX_NAME_LEN, "%s%u",
 305                         RING_SLAVE_NAME, slaveid);
 306         if ((res->ring[1] = create_ring(res->ring_name[1], NB_CORE_MSGBUF,
 307                 socketid, flag)) == NULL) {
 308                 printf("Create s2m ring %s failed\n", res->ring_name[1]);
 309                 return -1;
 310         }
 311
 312         return 0;
 313 }
 314
 315 /* send command to pair in paired master and slave ring */
 316 static inline int
 317 sendcmd(unsigned slaveid, enum l2fwd_cmd cmd, int is_master)
 318 {
 319         struct lcore_resource_struct *res = &lcore_resource[slaveid];
 320         void *msg;
 321         int fd = !is_master;
 322
 323         /* Only check master, it must be enabled and running if it is slave */
 324         if (is_master && !res->enabled)
 325                 return -1;
 326
 327         if (res->ring[fd] == NULL)
 328                 return -1;
 329
 330         if (rte_mempool_get(message_pool, &msg) < 0) {
 331                 printf("Error to get message buffer\n");
 332                 return -1;
 333         }
 334
 335         *(enum l2fwd_cmd *)msg = cmd;
 336
 337         if (rte_ring_enqueue(res->ring[fd], msg) != 0) {
 338                 printf("Enqueue error\n");
 339                 rte_mempool_put(message_pool, msg);
 340                 return -1;
 341         }
 342
 343         return 0;
 344 }
 345
 346 /* Get command from pair in paired master and slave ring */
 347 static inline int
 348 getcmd(unsigned slaveid, enum l2fwd_cmd *cmd, int is_master)
 349 {
 350         struct lcore_resource_struct *res = &lcore_resource[slaveid];
 351         void *msg;
 352         int fd = !!is_master;
 353         int ret;
 354         /* Only check master, it must be enabled and running if it is slave */
 355         if (is_master && (!res->enabled))
 356                 return -1;
 357
 358         if (res->ring[fd] == NULL)
 359                 return -1;
 360
 361         ret = rte_ring_dequeue(res->ring[fd], &msg);
 362
 363         if (ret == 0) {
 364                 *cmd = *(enum l2fwd_cmd *)msg;
 365                 rte_mempool_put(message_pool, msg);
 366         }
 367         return ret;
 368 }
 369
 370 /* Master send command to slave and wait until ack received or error met */
 371 static int
 372 master_sendcmd_with_ack(unsigned slaveid, enum l2fwd_cmd cmd)
 373 {
 374         enum l2fwd_cmd ack_cmd;
 375         int ret = -1;
 376
 377         if (sendcmd(slaveid, cmd, 1) != 0)
 378                 rte_exit(EXIT_FAILURE, "Failed to send message\n");
 379
 380         /* Get ack */
 381         while (1) {
 382                 ret = getcmd(slaveid, &ack_cmd, 1);
 383                 if (ret == 0 && cmd == ack_cmd)
 384                         break;
 385
 386                 /* If slave not running yet, return an error */
 387                 if (flib_query_slave_status(slaveid) != ST_RUN) {
 388                         ret = -ENOENT;
 389                         break;
 390                 }
 391         }
 392
 393         return ret;
 394 }
 395
 396 /* restart all port that assigned to that slave lcore */
 397 static int
 398 reset_slave_all_ports(unsigned slaveid)
 399 {
 400         struct lcore_resource_struct *slave = &lcore_resource[slaveid];
 401         int i, ret = 0;
 402
 403         /* stop/start port */
 404         for (i = 0; i < slave->port_num; i++) {
 405                 char buf_name[RTE_MEMPOOL_NAMESIZE];
 406                 struct rte_mempool *pool;
 407                 printf("Stop port :%d\n", slave->port[i]);
 408                 rte_eth_dev_stop(slave->port[i]);
 409                 snprintf(buf_name, RTE_MEMPOOL_NAMESIZE, MBUF_NAME, slave->port[i]);
 410                 pool = rte_mempool_lookup(buf_name);
 411                 if (pool)
 412                         printf("Port %d mempool free object is %u(%u)\n", slave->port[i],
 413                                 rte_mempool_avail_count(pool),
 414                                 (unsigned int)NB_MBUF);
 415                 else
 416                         printf("Can't find mempool %s\n", buf_name);
 417
 418                 printf("Start port :%d\n", slave->port[i]);
 419                 ret = rte_eth_dev_start(slave->port[i]);
 420                 if (ret != 0)
 421                         break;
 422         }
 423         return ret;
 424 }
 425
 426 static int
 427 reset_shared_structures(unsigned slaveid)
 428 {
 429         int ret;
 430         /* Only port are shared resource here */
 431         ret = reset_slave_all_ports(slaveid);
 432
 433         return ret;
 434 }
 435
 436 /**
 437  * Call this function to re-create resource that needed for slave process that
 438  * exited in last instance
 439  **/
 440 static int
 441 init_slave_res(unsigned slaveid)
 442 {
 443         struct lcore_resource_struct *slave = &lcore_resource[slaveid];
 444         enum l2fwd_cmd cmd;
 445
 446         if (!slave->enabled) {
 447                 printf("Something wrong with lcore=%u enabled=%d\n",slaveid,
 448                         slave->enabled);
 449                 return -1;
 450         }
 451
 452         /* Initialize ring */
 453         if (create_ms_ring(slaveid) != 0)
 454                 rte_exit(EXIT_FAILURE, "failed to create ring for slave %u\n",
 455                                 slaveid);
 456
 457         /* drain un-read buffer if have */
 458         while (getcmd(slaveid, &cmd, 1) == 0);
 459         while (getcmd(slaveid, &cmd, 0) == 0);
 460
 461         return 0;
 462 }
 463
 464 static int
 465 recreate_one_slave(unsigned slaveid)
 466 {
 467         int ret = 0;
 468         /* Re-initialize resource for stalled slave */
 469         if ((ret = init_slave_res(slaveid)) != 0) {
 470                 printf("Init slave=%u failed\n", slaveid);
 471                 return ret;
 472         }
 473
 474         if ((ret = flib_remote_launch(l2fwd_launch_one_lcore, NULL, slaveid))
 475                 != 0)
 476                 printf("Launch slave %u failed\n", slaveid);
 477
 478         return ret;
 479 }
 480
 481 /**
 482  * remapping resource belong to slave_id to new lcore that gets from flib_assign_lcore_id(),
 483  * used only floating process option applied.
 484  *
 485  * @param slaveid
 486  *   original lcore_id that apply for remapping
 487  */
 488 static void
 489 remapping_slave_resource(unsigned slaveid, unsigned map_id)
 490 {
 491
 492         /* remapping lcore_resource */
 493         memcpy(&lcore_resource[map_id], &lcore_resource[slaveid],
 494                         sizeof(struct lcore_resource_struct));
 495
 496         /* remapping lcore_queue_conf */
 497         memcpy(&lcore_queue_conf[map_id], &lcore_queue_conf[slaveid],
 498                         sizeof(struct lcore_queue_conf));
 499 }
 500
 501 static int
 502 reset_pair(unsigned slaveid, unsigned pairid)
 503 {
 504         int ret;
 505         if ((ret = reset_shared_structures(slaveid)) != 0)
 506                 goto back;
 507
 508         if((ret = reset_shared_structures(pairid)) != 0)
 509                 goto back;
 510
 511         if (float_proc) {
 512                 unsigned map_id = mapping_id[slaveid];
 513
 514                 if (map_id != INVALID_MAPPING_ID) {
 515                         printf("%u return mapping id %u\n", slaveid, map_id);
 516                         flib_free_lcore_id(map_id);
 517                         mapping_id[slaveid] = INVALID_MAPPING_ID;
 518                 }
 519
 520                 map_id = mapping_id[pairid];
 521                 if (map_id != INVALID_MAPPING_ID) {
 522                         printf("%u return mapping id %u\n", pairid, map_id);
 523                         flib_free_lcore_id(map_id);
 524                         mapping_id[pairid] = INVALID_MAPPING_ID;
 525                 }
 526         }
 527
 528         if((ret = recreate_one_slave(slaveid)) != 0)
 529                 goto back;
 530
 531         ret = recreate_one_slave(pairid);
 532
 533 back:
 534         return ret;
 535 }
 536
 537 static void
 538 slave_exit_cb(unsigned slaveid, __attribute__((unused))int stat)
 539 {
 540         struct lcore_resource_struct *slave = &lcore_resource[slaveid];
 541
 542         printf("Get slave %u leave info\n", slaveid);
 543         if (!slave->enabled) {
 544                 printf("Lcore=%u not registered for it's exit\n", slaveid);
 545                 return;
 546         }
 547         rte_spinlock_lock(&res_lock);
 548
 549         /* Change the state and wait master to start them */
 550         slave->flags = SLAVE_RECREATE_FLAG;
 551
 552         rte_spinlock_unlock(&res_lock);
 553 }
 554
 555 static void
 556 l2fwd_simple_forward(struct rte_mbuf *m, unsigned portid)
 557 {
 558         struct ether_hdr *eth;
 559         void *tmp;
 560         unsigned dst_port;
 561         int sent;
 562         struct rte_eth_dev_tx_buffer *buffer;
 563
 564         dst_port = l2fwd_dst_ports[portid];
 565         eth = rte_pktmbuf_mtod(m, struct ether_hdr *);
 566
 567         /* 02:00:00:00:00:xx */
 568         tmp = &eth->d_addr.addr_bytes[0];
 569         *((uint64_t *)tmp) = 0x000000000002 + ((uint64_t)dst_port << 40);
 570
 571         /* src addr */
 572         ether_addr_copy(&l2fwd_ports_eth_addr[dst_port], &eth->s_addr);
 573
 574         buffer = tx_buffer[dst_port];
 575         sent = rte_eth_tx_buffer(dst_port, 0, buffer, m);
 576         if (sent)
 577                 port_statistics[dst_port].tx += sent;
 578 }
 579
 580 /* main processing loop */
 581 static void
 582 l2fwd_main_loop(void)
 583 {
 584         struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
 585         struct rte_mbuf *m;
 586         int sent;
 587         unsigned lcore_id;
 588         uint64_t prev_tsc, diff_tsc, cur_tsc;
 589         unsigned i, j, portid, nb_rx;
 590         struct lcore_queue_conf *qconf;
 591         const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) / US_PER_S *
 592                         BURST_TX_DRAIN_US;
 593         struct rte_eth_dev_tx_buffer *buffer;
 594
 595         prev_tsc = 0;
 596
 597         lcore_id = rte_lcore_id();
 598
 599         qconf = &lcore_queue_conf[lcore_id];
 600
 601         if (qconf->n_rx_port == 0) {
 602                 RTE_LOG(INFO, L2FWD, "lcore %u has nothing to do\n", lcore_id);
 603                 return;
 604         }
 605
 606         RTE_LOG(INFO, L2FWD, "entering main loop on lcore %u\n", lcore_id);
 607
 608         for (i = 0; i < qconf->n_rx_port; i++) {
 609                 portid = qconf->rx_port_list[i];
 610                 RTE_LOG(INFO, L2FWD, " -- lcoreid=%u portid=%u\n", lcore_id,
 611                         portid);
 612         }
 613
 614         while (1) {
 615                 enum l2fwd_cmd cmd;
 616                 cur_tsc = rte_rdtsc();
 617
 618                 if (unlikely(getcmd(lcore_id, &cmd, 0) == 0)) {
 619                         sendcmd(lcore_id, cmd, 0);
 620
 621                         /* If get stop command, stop forwarding and exit */
 622                         if (cmd == CMD_STOP) {
 623                                 return;
 624                         }
 625                 }
 626
 627                 /*
 628                  * TX burst queue drain
 629                  */
 630                 diff_tsc = cur_tsc - prev_tsc;
 631                 if (unlikely(diff_tsc > drain_tsc)) {
 632
 633                         for (i = 0; i < qconf->n_rx_port; i++) {
 634
 635                                 portid = l2fwd_dst_ports[qconf->rx_port_list[i]];
 636                                 buffer = tx_buffer[portid];
 637
 638                                 sent = rte_eth_tx_buffer_flush(portid, 0, buffer);
 639                                 if (sent)
 640                                         port_statistics[portid].tx += sent;
 641
 642                         }
 643
 644                         prev_tsc = cur_tsc;
 645                 }
 646
 647                 /*
 648                  * Read packet from RX queues
 649                  */
 650                 for (i = 0; i < qconf->n_rx_port; i++) {
 651
 652                         portid = qconf->rx_port_list[i];
 653                         nb_rx = rte_eth_rx_burst((uint8_t) portid, 0,
 654                                                  pkts_burst, MAX_PKT_BURST);
 655
 656                         port_statistics[portid].rx += nb_rx;
 657
 658                         for (j = 0; j < nb_rx; j++) {
 659                                 m = pkts_burst[j];
 660                                 rte_prefetch0(rte_pktmbuf_mtod(m, void *));
 661                                 l2fwd_simple_forward(m, portid);
 662                         }
 663                 }
 664         }
 665 }
 666
 667 static int
 668 l2fwd_launch_one_lcore(__attribute__((unused)) void *dummy)
 669 {
 670         unsigned lcore_id = rte_lcore_id();
 671
 672         if (float_proc) {
 673                 unsigned flcore_id;
 674
 675                 /* Change it to floating process, also change it's lcore_id */
 676                 clear_cpu_affinity();
 677                 RTE_PER_LCORE(_lcore_id) = 0;
 678                 /* Get a lcore_id */
 679                 if (flib_assign_lcore_id() < 0 ) {
 680                         printf("flib_assign_lcore_id failed\n");
 681                         return -1;
 682                 }
 683                 flcore_id = rte_lcore_id();
 684                 /* Set mapping id, so master can return it after slave exited */
 685                 mapping_id[lcore_id] = flcore_id;
 686                 printf("Org lcore_id = %u, cur lcore_id = %u\n",
 687                                 lcore_id, flcore_id);
 688                 remapping_slave_resource(lcore_id, flcore_id);
 689         }
 690
 691         l2fwd_main_loop();
 692
 693         /* return lcore_id before return */
 694         if (float_proc) {
 695                 flib_free_lcore_id(rte_lcore_id());
 696                 mapping_id[lcore_id] = INVALID_MAPPING_ID;
 697         }
 698         return 0;
 699 }
 700
 701 /* display usage */
 702 static void
 703 l2fwd_usage(const char *prgname)
 704 {
 705         printf("%s [EAL options] -- -p PORTMASK -s COREMASK [-q NQ] -f\n"
 706                "  -p PORTMASK: hexadecimal bitmask of ports to configure\n"
 707                "  -q NQ: number of queue (=ports) per lcore (default is 1)\n"
 708                "  -f use floating process which won't bind to any core to run\n"
 709                    "  -T PERIOD: statistics will be refreshed each PERIOD seconds (0 to disable, 10 default, 86400 maximum)\n",
 710                prgname);
 711 }
 712
 713 static int
 714 l2fwd_parse_portmask(const char *portmask)
 715 {
 716         char *end = NULL;
 717         unsigned long pm;
 718
 719         /* parse hexadecimal string */
 720         pm = strtoul(portmask, &end, 16);
 721         if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0'))
 722                 return -1;
 723
 724         if (pm == 0)
 725                 return -1;
 726
 727         return pm;
 728 }
 729
 730 static unsigned int
 731 l2fwd_parse_nqueue(const char *q_arg)
 732 {
 733         char *end = NULL;
 734         unsigned long n;
 735
 736         /* parse hexadecimal string */
 737         n = strtoul(q_arg, &end, 10);
 738         if ((q_arg[0] == '\0') || (end == NULL) || (*end != '\0'))
 739                 return 0;
 740         if (n == 0)
 741                 return 0;
 742         if (n >= MAX_RX_QUEUE_PER_LCORE)
 743                 return 0;
 744
 745         return n;
 746 }
 747
 748 static int
 749 l2fwd_parse_timer_period(const char *q_arg)
 750 {
 751         char *end = NULL;
 752         int n;
 753
 754         /* parse number string */
 755         n = strtol(q_arg, &end, 10);
 756         if ((q_arg[0] == '\0') || (end == NULL) || (*end != '\0'))
 757                 return -1;
 758         if (n >= MAX_TIMER_PERIOD)
 759                 return -1;
 760
 761         return n;
 762 }
 763
 764 /* Parse the argument given in the command line of the application */
 765 static int
 766 l2fwd_parse_args(int argc, char **argv)
 767 {
 768         int opt, ret;
 769         char **argvopt;
 770         int option_index;
 771         char *prgname = argv[0];
 772         static struct option lgopts[] = {
 773                 {NULL, 0, 0, 0}
 774         };
 775         int has_pmask = 0;
 776
 777         argvopt = argv;
 778
 779         while ((opt = getopt_long(argc, argvopt, "p:q:T:f",
 780                                   lgopts, &option_index)) != EOF) {
 781
 782                 switch (opt) {
 783                 /* portmask */
 784                 case 'p':
 785                         l2fwd_enabled_port_mask = l2fwd_parse_portmask(optarg);
 786                         if (l2fwd_enabled_port_mask == 0) {
 787                                 printf("invalid portmask\n");
 788                                 l2fwd_usage(prgname);
 789                                 return -1;
 790                         }
 791                         has_pmask = 1;
 792                         break;
 793
 794                 /* nqueue */
 795                 case 'q':
 796                         l2fwd_rx_queue_per_lcore = l2fwd_parse_nqueue(optarg);
 797                         if (l2fwd_rx_queue_per_lcore == 0) {
 798                                 printf("invalid queue number\n");
 799                                 l2fwd_usage(prgname);
 800                                 return -1;
 801                         }
 802                         break;
 803
 804                 /* timer period */
 805                 case 'T':
 806                         timer_period = l2fwd_parse_timer_period(optarg) * 1000 * TIMER_MILLISECOND;
 807                         if (timer_period < 0) {
 808                                 printf("invalid timer period\n");
 809                                 l2fwd_usage(prgname);
 810                                 return -1;
 811                         }
 812                         break;
 813
 814                 /* use floating process */
 815                 case 'f':
 816                         float_proc = 1;
 817                         break;
 818
 819                 /* long options */
 820                 case 0:
 821                         l2fwd_usage(prgname);
 822                         return -1;
 823
 824                 default:
 825                         l2fwd_usage(prgname);
 826                         return -1;
 827                 }
 828         }
 829
 830         if (optind >= 0)
 831                 argv[optind-1] = prgname;
 832
 833         if (!has_pmask) {
 834                 l2fwd_usage(prgname);
 835                 return -1;
 836         }
 837         ret = optind-1;
 838         optind = 1; /* reset getopt lib */
 839         return ret;
 840 }
 841
 842 /* Check the link status of all ports in up to 9s, and print them finally */
 843 static void
 844 check_all_ports_link_status(uint16_t port_num, uint32_t port_mask)
 845 {
 846 #define CHECK_INTERVAL 100 /* 100ms */
 847 #define MAX_CHECK_TIME 90 /* 9s (90 * 100ms) in total */
 848         uint16_t portid;
 849         uint8_t count, all_ports_up, print_flag = 0;
 850         struct rte_eth_link link;
 851
 852         printf("\nChecking link status");
 853         fflush(stdout);
 854         for (count = 0; count <= MAX_CHECK_TIME; count++) {
 855                 all_ports_up = 1;
 856                 for (portid = 0; portid < port_num; portid++) {
 857                         if ((port_mask & (1 << portid)) == 0)
 858                                 continue;
 859                         memset(&link, 0, sizeof(link));
 860                         rte_eth_link_get_nowait(portid, &link);
 861                         /* print link status if flag set */
 862                         if (print_flag == 1) {
 863                                 if (link.link_status)
 864                                         printf(
 865                                         "Port%d Link Up- speed %u Mbps- %s\n",
 866                                         portid, link.link_speed,
 867                                 (link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
 868                                         ("full-duplex") : ("half-duplex\n"));
 869                                 else
 870                                         printf("Port %d Link Down\n", portid);
 871                                 continue;
 872                         }
 873                         /* clear all_ports_up flag if any link down */
 874                         if (link.link_status == ETH_LINK_DOWN) {
 875                                 all_ports_up = 0;
 876                                 break;
 877                         }
 878                 }
 879                 /* after finally printing all link status, get out */
 880                 if (print_flag == 1)
 881                         break;
 882
 883                 if (all_ports_up == 0) {
 884                         printf(".");
 885                         fflush(stdout);
 886                         rte_delay_ms(CHECK_INTERVAL);
 887                 }
 888
 889                 /* set the print_flag if all ports up or timeout */
 890                 if (all_ports_up == 1 || count == (MAX_CHECK_TIME - 1)) {
 891                         print_flag = 1;
 892                         printf("done\n");
 893                 }
 894         }
 895 }
 896
 897 int
 898 main(int argc, char **argv)
 899 {
 900         struct lcore_queue_conf *qconf;
 901         struct rte_eth_dev_info dev_info;
 902         int ret;
 903         uint16_t nb_ports;
 904         uint16_t nb_ports_available;
 905         uint16_t portid, last_port;
 906         unsigned rx_lcore_id;
 907         unsigned nb_ports_in_mask = 0;
 908         unsigned i;
 909         uint64_t prev_tsc, diff_tsc, cur_tsc, timer_tsc;
 910
 911         /* Save cpu_affinity first, restore it in case it's floating process option */
 912         if (get_cpu_affinity() != 0)
 913                 rte_exit(EXIT_FAILURE, "get_cpu_affinity error\n");
 914
 915         /* Also tries to set cpu affinity to detect whether  it will fail in child process */
 916         if(clear_cpu_affinity() != 0)
 917                 rte_exit(EXIT_FAILURE, "clear_cpu_affinity error\n");
 918
 919         /* init EAL */
 920         ret = rte_eal_init(argc, argv);
 921         if (ret < 0)
 922                 rte_exit(EXIT_FAILURE, "Invalid EAL arguments\n");
 923         argc -= ret;
 924         argv += ret;
 925
 926         /* parse application arguments (after the EAL ones) */
 927         ret = l2fwd_parse_args(argc, argv);
 928         if (ret < 0)
 929                 rte_exit(EXIT_FAILURE, "Invalid L2FWD arguments\n");
 930
 931         /*flib init */
 932         if (flib_init() != 0)
 933                 rte_exit(EXIT_FAILURE, "flib init error");
 934
 935         /**
 936           * Allocated structures that slave lcore would change. For those that slaves are
 937           * read only, needn't use malloc to share and global or static variables is ok since
 938           * slave inherit all the knowledge that master initialized.
 939           **/
 940         if (l2fwd_malloc_shared_struct() != 0)
 941                 rte_exit(EXIT_FAILURE, "malloc mem failed\n");
 942
 943         /* Initialize lcore_resource structures */
 944         memset(lcore_resource, 0, sizeof(lcore_resource));
 945         for (i = 0; i < RTE_MAX_LCORE; i++)
 946                 lcore_resource[i].lcore_id = i;
 947
 948         nb_ports = rte_eth_dev_count();
 949         if (nb_ports == 0)
 950                 rte_exit(EXIT_FAILURE, "No Ethernet ports - bye\n");
 951
 952         /* create the mbuf pool */
 953         for (portid = 0; portid < nb_ports; portid++) {
 954                 /* skip ports that are not enabled */
 955                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
 956                         continue;
 957                 char buf_name[RTE_MEMPOOL_NAMESIZE];
 958                 snprintf(buf_name, RTE_MEMPOOL_NAMESIZE, MBUF_NAME, portid);
 959                 l2fwd_pktmbuf_pool[portid] =
 960                         rte_pktmbuf_pool_create(buf_name, NB_MBUF, 32,
 961                                 0, MBUF_DATA_SIZE, rte_socket_id());
 962                 if (l2fwd_pktmbuf_pool[portid] == NULL)
 963                         rte_exit(EXIT_FAILURE, "Cannot create mbuf pool\n");
 964
 965                 printf("Create mbuf %s\n", buf_name);
 966         }
 967
 968         /* reset l2fwd_dst_ports */
 969         for (portid = 0; portid < RTE_MAX_ETHPORTS; portid++)
 970                 l2fwd_dst_ports[portid] = 0;
 971         last_port = 0;
 972
 973         /*
 974          * Each logical core is assigned a dedicated TX queue on each port.
 975          */
 976         for (portid = 0; portid < nb_ports; portid++) {
 977                 /* skip ports that are not enabled */
 978                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
 979                         continue;
 980
 981                 if (nb_ports_in_mask % 2) {
 982                         l2fwd_dst_ports[portid] = last_port;
 983                         l2fwd_dst_ports[last_port] = portid;
 984                 }
 985                 else
 986                         last_port = portid;
 987
 988                 nb_ports_in_mask++;
 989
 990                 rte_eth_dev_info_get(portid, &dev_info);
 991         }
 992         if (nb_ports_in_mask % 2) {
 993                 printf("Notice: odd number of ports in portmask.\n");
 994                 l2fwd_dst_ports[last_port] = last_port;
 995         }
 996
 997         rx_lcore_id = 0;
 998         qconf = NULL;
 999
1000         /* Initialize the port/queue configuration of each logical core */
1001         for (portid = 0; portid < nb_ports; portid++) {
1002                 struct lcore_resource_struct *res;
1003                 /* skip ports that are not enabled */
1004                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
1005                         continue;
1006
1007                 /* get the lcore_id for this port */
1008                 /* skip master lcore */
1009                 while (rte_lcore_is_enabled(rx_lcore_id) == 0 ||
1010                            rte_get_master_lcore() == rx_lcore_id ||
1011                        lcore_queue_conf[rx_lcore_id].n_rx_port ==
1012                        l2fwd_rx_queue_per_lcore) {
1013
1014                         rx_lcore_id++;
1015                         if (rx_lcore_id >= RTE_MAX_LCORE)
1016                                 rte_exit(EXIT_FAILURE, "Not enough cores\n");
1017                 }
1018
1019                 if (qconf != &lcore_queue_conf[rx_lcore_id])
1020                         /* Assigned a new logical core in the loop above. */
1021                         qconf = &lcore_queue_conf[rx_lcore_id];
1022
1023                 qconf->rx_port_list[qconf->n_rx_port] = portid;
1024                 qconf->n_rx_port++;
1025
1026                 /* Save the port resource info into lcore_resource strucutres */
1027                 res = &lcore_resource[rx_lcore_id];
1028                 res->enabled = 1;
1029                 res->port[res->port_num++] = portid;
1030
1031                 printf("Lcore %u: RX port %u\n", rx_lcore_id, (unsigned) portid);
1032         }
1033
1034         nb_ports_available = nb_ports;
1035
1036         /* Initialise each port */
1037         for (portid = 0; portid < nb_ports; portid++) {
1038                 /* skip ports that are not enabled */
1039                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0) {
1040                         printf("Skipping disabled port %u\n", (unsigned) portid);
1041                         nb_ports_available--;
1042                         continue;
1043                 }
1044                 /* init port */
1045                 printf("Initializing port %u... ", (unsigned) portid);
1046                 fflush(stdout);
1047                 ret = rte_eth_dev_configure(portid, 1, 1, &port_conf);
1048                 if (ret < 0)
1049                         rte_exit(EXIT_FAILURE, "Cannot configure device: err=%d, port=%u\n",
1050                                   ret, (unsigned) portid);
1051
1052                 ret = rte_eth_dev_adjust_nb_rx_tx_desc(portid, &nb_rxd,
1053                                                        &nb_txd);
1054                 if (ret < 0)
1055                         rte_exit(EXIT_FAILURE,
1056                                  "rte_eth_dev_adjust_nb_rx_tx_desc: err=%d, port=%u\n",
1057                                  ret, (unsigned) portid);
1058
1059                 rte_eth_macaddr_get(portid,&l2fwd_ports_eth_addr[portid]);
1060
1061                 /* init one RX queue */
1062                 fflush(stdout);
1063                 ret = rte_eth_rx_queue_setup(portid, 0, nb_rxd,
1064                                              rte_eth_dev_socket_id(portid),
1065                                              NULL,
1066                                              l2fwd_pktmbuf_pool[portid]);
1067                 if (ret < 0)
1068                         rte_exit(EXIT_FAILURE, "rte_eth_rx_queue_setup:err=%d, port=%u\n",
1069                                   ret, (unsigned) portid);
1070
1071                 /* init one TX queue on each port */
1072                 fflush(stdout);
1073                 ret = rte_eth_tx_queue_setup(portid, 0, nb_txd,
1074                                 rte_eth_dev_socket_id(portid),
1075                                 NULL);
1076                 if (ret < 0)
1077                         rte_exit(EXIT_FAILURE, "rte_eth_tx_queue_setup:err=%d, port=%u\n",
1078                                 ret, (unsigned) portid);
1079
1080                 /* Initialize TX buffers */
1081                 tx_buffer[portid] = rte_zmalloc_socket("tx_buffer",
1082                                 RTE_ETH_TX_BUFFER_SIZE(MAX_PKT_BURST), 0,
1083                                 rte_eth_dev_socket_id(portid));
1084                 if (tx_buffer[portid] == NULL)
1085                         rte_exit(EXIT_FAILURE, "Cannot allocate buffer for tx on port %u\n",
1086                                         (unsigned) portid);
1087
1088                 rte_eth_tx_buffer_init(tx_buffer[portid], MAX_PKT_BURST);
1089
1090                 ret = rte_eth_tx_buffer_set_err_callback(tx_buffer[portid],
1091                                 rte_eth_tx_buffer_count_callback,
1092                                 &port_statistics[portid].dropped);
1093                 if (ret < 0)
1094                                 rte_exit(EXIT_FAILURE, "Cannot set error callback for "
1095                                                 "tx buffer on port %u\n", (unsigned) portid);
1096
1097                 /* Start device */
1098                 ret = rte_eth_dev_start(portid);
1099                 if (ret < 0)
1100                         rte_exit(EXIT_FAILURE, "rte_eth_dev_start:err=%d, port=%u\n",
1101                                   ret, (unsigned) portid);
1102
1103                 printf("done: \n");
1104
1105                 rte_eth_promiscuous_enable(portid);
1106
1107                 printf("Port %u, MAC address: %02X:%02X:%02X:%02X:%02X:%02X\n\n",
1108                                 (unsigned) portid,
1109                                 l2fwd_ports_eth_addr[portid].addr_bytes[0],
1110                                 l2fwd_ports_eth_addr[portid].addr_bytes[1],
1111                                 l2fwd_ports_eth_addr[portid].addr_bytes[2],
1112                                 l2fwd_ports_eth_addr[portid].addr_bytes[3],
1113                                 l2fwd_ports_eth_addr[portid].addr_bytes[4],
1114                                 l2fwd_ports_eth_addr[portid].addr_bytes[5]);
1115
1116                 /* initialize port stats */
1117                 //memset(&port_statistics, 0, sizeof(port_statistics));
1118         }
1119
1120         if (!nb_ports_available) {
1121                 rte_exit(EXIT_FAILURE,
1122                         "All available ports are disabled. Please set portmask.\n");
1123         }
1124
1125         check_all_ports_link_status(nb_ports, l2fwd_enabled_port_mask);
1126
1127         /* Record pair lcore */
1128         /**
1129          * Since l2fwd example would create pair between different neighbour port, that's
1130          * port 0 receive and forward to port 1, the same to port 1, these 2 ports will have
1131          * dependency. If one port stopped working (killed, for example), the port need to
1132          * be stopped/started again. During the time, another port need to wait until stop/start
1133          * procedure completed. So, record the pair relationship for those lcores working
1134          * on ports.
1135          **/
1136         for (portid = 0; portid < nb_ports; portid++) {
1137                 uint32_t pair_port;
1138                 unsigned lcore = 0, pair_lcore = 0;
1139                 unsigned j, find_lcore, find_pair_lcore;
1140                 /* skip ports that are not enabled */
1141                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
1142                         continue;
1143
1144                 /* Find pair ports' lcores */
1145                 find_lcore = find_pair_lcore = 0;
1146                 pair_port = l2fwd_dst_ports[portid];
1147                 for (i = 0; i < RTE_MAX_LCORE; i++) {
1148                         if (!rte_lcore_is_enabled(i))
1149                                 continue;
1150                         for (j = 0; j < lcore_queue_conf[i].n_rx_port;j++) {
1151                                 if (lcore_queue_conf[i].rx_port_list[j] == portid) {
1152                                         lcore = i;
1153                                         find_lcore = 1;
1154                                         break;
1155                                 }
1156                                 if (lcore_queue_conf[i].rx_port_list[j] == pair_port) {
1157                                         pair_lcore = i;
1158                                         find_pair_lcore = 1;
1159                                         break;
1160                                 }
1161                         }
1162                         if (find_lcore && find_pair_lcore)
1163                                 break;
1164                 }
1165                 if (!find_lcore || !find_pair_lcore)
1166                         rte_exit(EXIT_FAILURE, "Not find port=%d pair\n", portid);
1167
1168                 printf("lcore %u and %u paired\n", lcore, pair_lcore);
1169                 lcore_resource[lcore].pair_id = pair_lcore;
1170                 lcore_resource[pair_lcore].pair_id = lcore;
1171         }
1172
1173         /* Create message buffer for all master and slave */
1174         message_pool = rte_mempool_create("ms_msg_pool",
1175                            NB_CORE_MSGBUF * RTE_MAX_LCORE,
1176                            sizeof(enum l2fwd_cmd), NB_CORE_MSGBUF / 2,
1177                            0, NULL, NULL, NULL, NULL, rte_socket_id(), 0);
1178
1179         if (message_pool == NULL)
1180                 rte_exit(EXIT_FAILURE, "Create msg mempool failed\n");
1181
1182         /* Create ring for each master and slave pair, also register cb when slave leaves */
1183         for (i = 0; i < RTE_MAX_LCORE; i++) {
1184                 /**
1185                  * Only create ring and register slave_exit cb in case that core involved into
1186                  * packet forwarding
1187                  **/
1188                 if (lcore_resource[i].enabled) {
1189                         /* Create ring for master and slave communication */
1190                         ret = create_ms_ring(i);
1191                         if (ret != 0)
1192                                 rte_exit(EXIT_FAILURE, "Create ring for lcore=%u failed",
1193                                 i);
1194
1195                         if (flib_register_slave_exit_notify(i,
1196                                 slave_exit_cb) != 0)
1197                                 rte_exit(EXIT_FAILURE,
1198                                                 "Register master_trace_slave_exit failed");
1199                 }
1200         }
1201
1202         /* launch per-lcore init on every lcore except master */
1203         flib_mp_remote_launch(l2fwd_launch_one_lcore, NULL, SKIP_MASTER);
1204
1205         /* print statistics 10 second */
1206         prev_tsc = cur_tsc = rte_rdtsc();
1207         timer_tsc = 0;
1208         while (1) {
1209                 sleep(1);
1210                 cur_tsc = rte_rdtsc();
1211                 diff_tsc = cur_tsc - prev_tsc;
1212                 /* if timer is enabled */
1213                 if (timer_period > 0) {
1214
1215                         /* advance the timer */
1216                         timer_tsc += diff_tsc;
1217
1218                         /* if timer has reached its timeout */
1219                         if (unlikely(timer_tsc >= (uint64_t) timer_period)) {
1220
1221                                 print_stats();
1222                                 /* reset the timer */
1223                                 timer_tsc = 0;
1224                         }
1225                 }
1226
1227                 prev_tsc = cur_tsc;
1228
1229                 /* Check any slave need restart or recreate */
1230                 rte_spinlock_lock(&res_lock);
1231                 for (i = 0; i < RTE_MAX_LCORE; i++) {
1232                         struct lcore_resource_struct *res  = &lcore_resource[i];
1233                         struct lcore_resource_struct *pair = &lcore_resource[res->pair_id];
1234
1235                         /* If find slave exited, try to reset pair */
1236                         if (res->enabled && res->flags && pair->enabled) {
1237                                 if (!pair->flags) {
1238                                         master_sendcmd_with_ack(pair->lcore_id, CMD_STOP);
1239                                         rte_spinlock_unlock(&res_lock);
1240                                         sleep(1);
1241                                         rte_spinlock_lock(&res_lock);
1242                                         if (pair->flags)
1243                                                 continue;
1244                                 }
1245                                 if (reset_pair(res->lcore_id, pair->lcore_id) != 0)
1246                                         rte_exit(EXIT_FAILURE, "failed to reset slave");
1247                                 res->flags  = 0;
1248                                 pair->flags = 0;
1249                         }
1250                 }
1251                 rte_spinlock_unlock(&res_lock);
1252         }
1253
1254 }