examples/l2fwd-jobstats/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <locale.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <stdint.h>
   9 #include <ctype.h>
  10 #include <getopt.h>
  11
  12 #include <rte_common.h>
  13 #include <rte_log.h>
  14 #include <rte_malloc.h>
  15 #include <rte_memory.h>
  16 #include <rte_memcpy.h>
  17 #include <rte_eal.h>
  18 #include <rte_launch.h>
  19 #include <rte_atomic.h>
  20 #include <rte_cycles.h>
  21 #include <rte_prefetch.h>
  22 #include <rte_lcore.h>
  23 #include <rte_per_lcore.h>
  24 #include <rte_branch_prediction.h>
  25 #include <rte_interrupts.h>
  26 #include <rte_debug.h>
  27 #include <rte_ether.h>
  28 #include <rte_ethdev.h>
  29 #include <rte_mempool.h>
  30 #include <rte_mbuf.h>
  31 #include <rte_spinlock.h>
  32
  33 #include <rte_errno.h>
  34 #include <rte_jobstats.h>
  35 #include <rte_timer.h>
  36 #include <rte_alarm.h>
  37 #include <rte_pause.h>
  38
  39 #define RTE_LOGTYPE_L2FWD RTE_LOGTYPE_USER1
  40
  41 #define NB_MBUF   8192
  42
  43 #define MAX_PKT_BURST 32
  44 #define BURST_TX_DRAIN_US 100 /* TX drain every ~100us */
  45
  46 /*
  47  * Configurable number of RX/TX ring descriptors
  48  */
  49 #define RTE_TEST_RX_DESC_DEFAULT 128
  50 #define RTE_TEST_TX_DESC_DEFAULT 512
  51 static uint16_t nb_rxd = RTE_TEST_RX_DESC_DEFAULT;
  52 static uint16_t nb_txd = RTE_TEST_TX_DESC_DEFAULT;
  53
  54 /* ethernet addresses of ports */
  55 static struct ether_addr l2fwd_ports_eth_addr[RTE_MAX_ETHPORTS];
  56
  57 /* mask of enabled ports */
  58 static uint32_t l2fwd_enabled_port_mask;
  59
  60 /* list of enabled ports */
  61 static uint32_t l2fwd_dst_ports[RTE_MAX_ETHPORTS];
  62
  63 #define UPDATE_STEP_UP 1
  64 #define UPDATE_STEP_DOWN 32
  65
  66 static unsigned int l2fwd_rx_queue_per_lcore = 1;
  67
  68 #define MAX_RX_QUEUE_PER_LCORE 16
  69 #define MAX_TX_QUEUE_PER_PORT 16
  70 struct lcore_queue_conf {
  71         unsigned n_rx_port;
  72         unsigned rx_port_list[MAX_RX_QUEUE_PER_LCORE];
  73         uint64_t next_flush_time[RTE_MAX_ETHPORTS];
  74
  75         struct rte_timer rx_timers[MAX_RX_QUEUE_PER_LCORE];
  76         struct rte_jobstats port_fwd_jobs[MAX_RX_QUEUE_PER_LCORE];
  77
  78         struct rte_timer flush_timer;
  79         struct rte_jobstats flush_job;
  80         struct rte_jobstats idle_job;
  81         struct rte_jobstats_context jobs_context;
  82
  83         rte_atomic16_t stats_read_pending;
  84         rte_spinlock_t lock;
  85 } __rte_cache_aligned;
  86 struct lcore_queue_conf lcore_queue_conf[RTE_MAX_LCORE];
  87
  88 struct rte_eth_dev_tx_buffer *tx_buffer[RTE_MAX_ETHPORTS];
  89
  90 static const struct rte_eth_conf port_conf = {
  91         .rxmode = {
  92                 .split_hdr_size = 0,
  93                 .header_split   = 0, /**< Header Split disabled */
  94                 .hw_ip_checksum = 0, /**< IP checksum offload disabled */
  95                 .hw_vlan_filter = 0, /**< VLAN filtering disabled */
  96                 .jumbo_frame    = 0, /**< Jumbo Frame Support disabled */
  97                 .hw_strip_crc   = 1, /**< CRC stripped by hardware */
  98         },
  99         .txmode = {
 100                 .mq_mode = ETH_MQ_TX_NONE,
 101         },
 102 };
 103
 104 struct rte_mempool *l2fwd_pktmbuf_pool = NULL;
 105
 106 /* Per-port statistics struct */
 107 struct l2fwd_port_statistics {
 108         uint64_t tx;
 109         uint64_t rx;
 110         uint64_t dropped;
 111 } __rte_cache_aligned;
 112 struct l2fwd_port_statistics port_statistics[RTE_MAX_ETHPORTS];
 113
 114 /* 1 day max */
 115 #define MAX_TIMER_PERIOD 86400
 116 /* default period is 10 seconds */
 117 static int64_t timer_period = 10;
 118 /* default timer frequency */
 119 static double hz;
 120 /* BURST_TX_DRAIN_US converted to cycles */
 121 uint64_t drain_tsc;
 122 /* Convert cycles to ns */
 123 static inline double
 124 cycles_to_ns(uint64_t cycles)
 125 {
 126         double t = cycles;
 127
 128         t *= (double)NS_PER_S;
 129         t /= hz;
 130         return t;
 131 }
 132
 133 static void
 134 show_lcore_stats(unsigned lcore_id)
 135 {
 136         struct lcore_queue_conf *qconf = &lcore_queue_conf[lcore_id];
 137         struct rte_jobstats_context *ctx = &qconf->jobs_context;
 138         struct rte_jobstats *job;
 139         uint8_t i;
 140
 141         /* LCore statistics. */
 142         uint64_t stats_period, loop_count;
 143         uint64_t exec, exec_min, exec_max;
 144         uint64_t management, management_min, management_max;
 145         uint64_t busy, busy_min, busy_max;
 146
 147         /* Jobs statistics. */
 148         const uint16_t port_cnt = qconf->n_rx_port;
 149         uint64_t jobs_exec_cnt[port_cnt], jobs_period[port_cnt];
 150         uint64_t jobs_exec[port_cnt], jobs_exec_min[port_cnt],
 151                                 jobs_exec_max[port_cnt];
 152
 153         uint64_t flush_exec_cnt, flush_period;
 154         uint64_t flush_exec, flush_exec_min, flush_exec_max;
 155
 156         uint64_t idle_exec_cnt;
 157         uint64_t idle_exec, idle_exec_min, idle_exec_max;
 158         uint64_t collection_time = rte_get_timer_cycles();
 159
 160         /* Ask forwarding thread to give us stats. */
 161         rte_atomic16_set(&qconf->stats_read_pending, 1);
 162         rte_spinlock_lock(&qconf->lock);
 163         rte_atomic16_set(&qconf->stats_read_pending, 0);
 164
 165         /* Collect context statistics. */
 166         stats_period = ctx->state_time - ctx->start_time;
 167         loop_count = ctx->loop_cnt;
 168
 169         exec = ctx->exec_time;
 170         exec_min = ctx->min_exec_time;
 171         exec_max = ctx->max_exec_time;
 172
 173         management = ctx->management_time;
 174         management_min = ctx->min_management_time;
 175         management_max = ctx->max_management_time;
 176
 177         rte_jobstats_context_reset(ctx);
 178
 179         for (i = 0; i < port_cnt; i++) {
 180                 job = &qconf->port_fwd_jobs[i];
 181
 182                 jobs_exec_cnt[i] = job->exec_cnt;
 183                 jobs_period[i] = job->period;
 184
 185                 jobs_exec[i] = job->exec_time;
 186                 jobs_exec_min[i] = job->min_exec_time;
 187                 jobs_exec_max[i] = job->max_exec_time;
 188
 189                 rte_jobstats_reset(job);
 190         }
 191
 192         flush_exec_cnt = qconf->flush_job.exec_cnt;
 193         flush_period = qconf->flush_job.period;
 194         flush_exec = qconf->flush_job.exec_time;
 195         flush_exec_min = qconf->flush_job.min_exec_time;
 196         flush_exec_max = qconf->flush_job.max_exec_time;
 197         rte_jobstats_reset(&qconf->flush_job);
 198
 199         idle_exec_cnt = qconf->idle_job.exec_cnt;
 200         idle_exec = qconf->idle_job.exec_time;
 201         idle_exec_min = qconf->idle_job.min_exec_time;
 202         idle_exec_max = qconf->idle_job.max_exec_time;
 203         rte_jobstats_reset(&qconf->idle_job);
 204
 205         rte_spinlock_unlock(&qconf->lock);
 206
 207         exec -= idle_exec;
 208         busy = exec + management;
 209         busy_min = exec_min + management_min;
 210         busy_max = exec_max + management_max;
 211
 212
 213         collection_time = rte_get_timer_cycles() - collection_time;
 214
 215 #define STAT_FMT "\n%-18s %'14.0f %6.1f%% %'10.0f %'10.0f %'10.0f"
 216
 217         printf("\n----------------"
 218                         "\nLCore %3u: statistics (time in ns, collected in %'9.0f)"
 219                         "\n%-18s %14s %7s %10s %10s %10s "
 220                         "\n%-18s %'14.0f"
 221                         "\n%-18s %'14" PRIu64
 222                         STAT_FMT /* Exec */
 223                         STAT_FMT /* Management */
 224                         STAT_FMT /* Busy */
 225                         STAT_FMT, /* Idle  */
 226                         lcore_id, cycles_to_ns(collection_time),
 227                         "Stat type", "total", "%total", "avg", "min", "max",
 228                         "Stats duration:", cycles_to_ns(stats_period),
 229                         "Loop count:", loop_count,
 230                         "Exec time",
 231                         cycles_to_ns(exec), exec * 100.0 / stats_period,
 232                         cycles_to_ns(loop_count  ? exec / loop_count : 0),
 233                         cycles_to_ns(exec_min),
 234                         cycles_to_ns(exec_max),
 235                         "Management time",
 236                         cycles_to_ns(management), management * 100.0 / stats_period,
 237                         cycles_to_ns(loop_count  ? management / loop_count : 0),
 238                         cycles_to_ns(management_min),
 239                         cycles_to_ns(management_max),
 240                         "Exec + management",
 241                         cycles_to_ns(busy),  busy * 100.0 / stats_period,
 242                         cycles_to_ns(loop_count ? busy / loop_count : 0),
 243                         cycles_to_ns(busy_min),
 244                         cycles_to_ns(busy_max),
 245                         "Idle (job)",
 246                         cycles_to_ns(idle_exec), idle_exec * 100.0 / stats_period,
 247                         cycles_to_ns(idle_exec_cnt ? idle_exec / idle_exec_cnt : 0),
 248                         cycles_to_ns(idle_exec_min),
 249                         cycles_to_ns(idle_exec_max));
 250
 251         for (i = 0; i < qconf->n_rx_port; i++) {
 252                 job = &qconf->port_fwd_jobs[i];
 253                 printf("\n\nJob %" PRIu32 ": %-20s "
 254                                 "\n%-18s %'14" PRIu64
 255                                 "\n%-18s %'14.0f"
 256                                 STAT_FMT,
 257                                 i, job->name,
 258                                 "Exec count:", jobs_exec_cnt[i],
 259                                 "Exec period: ", cycles_to_ns(jobs_period[i]),
 260                                 "Exec time",
 261                                 cycles_to_ns(jobs_exec[i]), jobs_exec[i] * 100.0 / stats_period,
 262                                 cycles_to_ns(jobs_exec_cnt[i] ? jobs_exec[i] / jobs_exec_cnt[i]
 263                                                 : 0),
 264                                 cycles_to_ns(jobs_exec_min[i]),
 265                                 cycles_to_ns(jobs_exec_max[i]));
 266         }
 267
 268         if (qconf->n_rx_port > 0) {
 269                 job = &qconf->flush_job;
 270                 printf("\n\nJob %" PRIu32 ": %-20s "
 271                                 "\n%-18s %'14" PRIu64
 272                                 "\n%-18s %'14.0f"
 273                                 STAT_FMT,
 274                                 i, job->name,
 275                                 "Exec count:", flush_exec_cnt,
 276                                 "Exec period: ", cycles_to_ns(flush_period),
 277                                 "Exec time",
 278                                 cycles_to_ns(flush_exec), flush_exec * 100.0 / stats_period,
 279                                 cycles_to_ns(flush_exec_cnt ? flush_exec / flush_exec_cnt : 0),
 280                                 cycles_to_ns(flush_exec_min),
 281                                 cycles_to_ns(flush_exec_max));
 282         }
 283 }
 284
 285 /* Print out statistics on packets dropped */
 286 static void
 287 show_stats_cb(__rte_unused void *param)
 288 {
 289         uint64_t total_packets_dropped, total_packets_tx, total_packets_rx;
 290         unsigned portid, lcore_id;
 291
 292         total_packets_dropped = 0;
 293         total_packets_tx = 0;
 294         total_packets_rx = 0;
 295
 296         const char clr[] = { 27, '[', '2', 'J', '\0' };
 297         const char topLeft[] = { 27, '[', '1', ';', '1', 'H', '\0' };
 298
 299         /* Clear screen and move to top left */
 300         printf("%s%s"
 301                         "\nPort statistics ===================================",
 302                         clr, topLeft);
 303
 304         for (portid = 0; portid < RTE_MAX_ETHPORTS; portid++) {
 305                 /* skip disabled ports */
 306                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
 307                         continue;
 308                 printf("\nStatistics for port %u ------------------------------"
 309                                 "\nPackets sent: %24"PRIu64
 310                                 "\nPackets received: %20"PRIu64
 311                                 "\nPackets dropped: %21"PRIu64,
 312                                 portid,
 313                                 port_statistics[portid].tx,
 314                                 port_statistics[portid].rx,
 315                                 port_statistics[portid].dropped);
 316
 317                 total_packets_dropped += port_statistics[portid].dropped;
 318                 total_packets_tx += port_statistics[portid].tx;
 319                 total_packets_rx += port_statistics[portid].rx;
 320         }
 321
 322         printf("\nAggregate statistics ==============================="
 323                         "\nTotal packets sent: %18"PRIu64
 324                         "\nTotal packets received: %14"PRIu64
 325                         "\nTotal packets dropped: %15"PRIu64
 326                         "\n====================================================",
 327                         total_packets_tx,
 328                         total_packets_rx,
 329                         total_packets_dropped);
 330
 331         RTE_LCORE_FOREACH(lcore_id) {
 332                 if (lcore_queue_conf[lcore_id].n_rx_port > 0)
 333                         show_lcore_stats(lcore_id);
 334         }
 335
 336         printf("\n====================================================\n");
 337         rte_eal_alarm_set(timer_period * US_PER_S, show_stats_cb, NULL);
 338 }
 339
 340 static void
 341 l2fwd_simple_forward(struct rte_mbuf *m, unsigned portid)
 342 {
 343         struct ether_hdr *eth;
 344         void *tmp;
 345         int sent;
 346         unsigned dst_port;
 347         struct rte_eth_dev_tx_buffer *buffer;
 348
 349         dst_port = l2fwd_dst_ports[portid];
 350         eth = rte_pktmbuf_mtod(m, struct ether_hdr *);
 351
 352         /* 02:00:00:00:00:xx */
 353         tmp = &eth->d_addr.addr_bytes[0];
 354         *((uint64_t *)tmp) = 0x000000000002 + ((uint64_t)dst_port << 40);
 355
 356         /* src addr */
 357         ether_addr_copy(&l2fwd_ports_eth_addr[dst_port], &eth->s_addr);
 358
 359         buffer = tx_buffer[dst_port];
 360         sent = rte_eth_tx_buffer(dst_port, 0, buffer, m);
 361         if (sent)
 362                 port_statistics[dst_port].tx += sent;
 363 }
 364
 365 static void
 366 l2fwd_job_update_cb(struct rte_jobstats *job, int64_t result)
 367 {
 368         int64_t err = job->target - result;
 369         int64_t histeresis = job->target / 8;
 370
 371         if (err < -histeresis) {
 372                 if (job->min_period + UPDATE_STEP_DOWN < job->period)
 373                         job->period -= UPDATE_STEP_DOWN;
 374         } else if (err > histeresis) {
 375                 if (job->period + UPDATE_STEP_UP < job->max_period)
 376                         job->period += UPDATE_STEP_UP;
 377         }
 378 }
 379
 380 static void
 381 l2fwd_fwd_job(__rte_unused struct rte_timer *timer, void *arg)
 382 {
 383         struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
 384         struct rte_mbuf *m;
 385
 386         const uint16_t port_idx = (uintptr_t) arg;
 387         const unsigned lcore_id = rte_lcore_id();
 388         struct lcore_queue_conf *qconf = &lcore_queue_conf[lcore_id];
 389         struct rte_jobstats *job = &qconf->port_fwd_jobs[port_idx];
 390         const uint16_t portid = qconf->rx_port_list[port_idx];
 391
 392         uint8_t j;
 393         uint16_t total_nb_rx;
 394
 395         rte_jobstats_start(&qconf->jobs_context, job);
 396
 397         /* Call rx burst 2 times. This allow rte_jobstats logic to see if this
 398          * function must be called more frequently. */
 399
 400         total_nb_rx = rte_eth_rx_burst(portid, 0, pkts_burst,
 401                         MAX_PKT_BURST);
 402
 403         for (j = 0; j < total_nb_rx; j++) {
 404                 m = pkts_burst[j];
 405                 rte_prefetch0(rte_pktmbuf_mtod(m, void *));
 406                 l2fwd_simple_forward(m, portid);
 407         }
 408
 409         if (total_nb_rx == MAX_PKT_BURST) {
 410                 const uint16_t nb_rx = rte_eth_rx_burst(portid, 0, pkts_burst,
 411                                 MAX_PKT_BURST);
 412
 413                 total_nb_rx += nb_rx;
 414                 for (j = 0; j < nb_rx; j++) {
 415                         m = pkts_burst[j];
 416                         rte_prefetch0(rte_pktmbuf_mtod(m, void *));
 417                         l2fwd_simple_forward(m, portid);
 418                 }
 419         }
 420
 421         port_statistics[portid].rx += total_nb_rx;
 422
 423         /* Adjust period time in which we are running here. */
 424         if (rte_jobstats_finish(job, total_nb_rx) != 0) {
 425                 rte_timer_reset(&qconf->rx_timers[port_idx], job->period, PERIODICAL,
 426                                 lcore_id, l2fwd_fwd_job, arg);
 427         }
 428 }
 429
 430 static void
 431 l2fwd_flush_job(__rte_unused struct rte_timer *timer, __rte_unused void *arg)
 432 {
 433         uint64_t now;
 434         unsigned lcore_id;
 435         struct lcore_queue_conf *qconf;
 436         uint16_t portid;
 437         unsigned i;
 438         uint32_t sent;
 439         struct rte_eth_dev_tx_buffer *buffer;
 440
 441         lcore_id = rte_lcore_id();
 442         qconf = &lcore_queue_conf[lcore_id];
 443
 444         rte_jobstats_start(&qconf->jobs_context, &qconf->flush_job);
 445
 446         now = rte_get_timer_cycles();
 447         lcore_id = rte_lcore_id();
 448         qconf = &lcore_queue_conf[lcore_id];
 449
 450         for (i = 0; i < qconf->n_rx_port; i++) {
 451                 portid = l2fwd_dst_ports[qconf->rx_port_list[i]];
 452
 453                 if (qconf->next_flush_time[portid] <= now)
 454                         continue;
 455
 456                 buffer = tx_buffer[portid];
 457                 sent = rte_eth_tx_buffer_flush(portid, 0, buffer);
 458                 if (sent)
 459                         port_statistics[portid].tx += sent;
 460
 461                 qconf->next_flush_time[portid] = rte_get_timer_cycles() + drain_tsc;
 462         }
 463
 464         /* Pass target to indicate that this job is happy of time interwal
 465          * in which it was called. */
 466         rte_jobstats_finish(&qconf->flush_job, qconf->flush_job.target);
 467 }
 468
 469 /* main processing loop */
 470 static void
 471 l2fwd_main_loop(void)
 472 {
 473         unsigned lcore_id;
 474         unsigned i, portid;
 475         struct lcore_queue_conf *qconf;
 476         uint8_t stats_read_pending = 0;
 477         uint8_t need_manage;
 478
 479         lcore_id = rte_lcore_id();
 480         qconf = &lcore_queue_conf[lcore_id];
 481
 482         if (qconf->n_rx_port == 0) {
 483                 RTE_LOG(INFO, L2FWD, "lcore %u has nothing to do\n", lcore_id);
 484                 return;
 485         }
 486
 487         RTE_LOG(INFO, L2FWD, "entering main loop on lcore %u\n", lcore_id);
 488
 489         for (i = 0; i < qconf->n_rx_port; i++) {
 490
 491                 portid = qconf->rx_port_list[i];
 492                 RTE_LOG(INFO, L2FWD, " -- lcoreid=%u portid=%u\n", lcore_id,
 493                         portid);
 494         }
 495
 496         rte_jobstats_init(&qconf->idle_job, "idle", 0, 0, 0, 0);
 497
 498         for (;;) {
 499                 rte_spinlock_lock(&qconf->lock);
 500
 501                 do {
 502                         rte_jobstats_context_start(&qconf->jobs_context);
 503
 504                         /* Do the Idle job:
 505                          * - Read stats_read_pending flag
 506                          * - check if some real job need to be executed
 507                          */
 508                         rte_jobstats_start(&qconf->jobs_context, &qconf->idle_job);
 509
 510                         uint64_t repeats = 0;
 511
 512                         do {
 513                                 uint8_t i;
 514                                 uint64_t now = rte_get_timer_cycles();
 515
 516                                 repeats++;
 517                                 need_manage = qconf->flush_timer.expire < now;
 518                                 /* Check if we was esked to give a stats. */
 519                                 stats_read_pending =
 520                                                 rte_atomic16_read(&qconf->stats_read_pending);
 521                                 need_manage |= stats_read_pending;
 522
 523                                 for (i = 0; i < qconf->n_rx_port && !need_manage; i++)
 524                                         need_manage = qconf->rx_timers[i].expire < now;
 525
 526                         } while (!need_manage);
 527
 528                         if (likely(repeats != 1))
 529                                 rte_jobstats_finish(&qconf->idle_job, qconf->idle_job.target);
 530                         else
 531                                 rte_jobstats_abort(&qconf->idle_job);
 532
 533                         rte_timer_manage();
 534                         rte_jobstats_context_finish(&qconf->jobs_context);
 535                 } while (likely(stats_read_pending == 0));
 536
 537                 rte_spinlock_unlock(&qconf->lock);
 538                 rte_pause();
 539         }
 540 }
 541
 542 static int
 543 l2fwd_launch_one_lcore(__attribute__((unused)) void *dummy)
 544 {
 545         l2fwd_main_loop();
 546         return 0;
 547 }
 548
 549 /* display usage */
 550 static void
 551 l2fwd_usage(const char *prgname)
 552 {
 553         printf("%s [EAL options] -- -p PORTMASK [-q NQ]\n"
 554                "  -p PORTMASK: hexadecimal bitmask of ports to configure\n"
 555                "  -q NQ: number of queue (=ports) per lcore (default is 1)\n"
 556                    "  -T PERIOD: statistics will be refreshed each PERIOD seconds (0 to disable, 10 default, 86400 maximum)\n"
 557                    "  -l set system default locale instead of default (\"C\" locale) for thousands separator in stats.",
 558                prgname);
 559 }
 560
 561 static int
 562 l2fwd_parse_portmask(const char *portmask)
 563 {
 564         char *end = NULL;
 565         unsigned long pm;
 566
 567         /* parse hexadecimal string */
 568         pm = strtoul(portmask, &end, 16);
 569         if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0'))
 570                 return -1;
 571
 572         if (pm == 0)
 573                 return -1;
 574
 575         return pm;
 576 }
 577
 578 static unsigned int
 579 l2fwd_parse_nqueue(const char *q_arg)
 580 {
 581         char *end = NULL;
 582         unsigned long n;
 583
 584         /* parse hexadecimal string */
 585         n = strtoul(q_arg, &end, 10);
 586         if ((q_arg[0] == '\0') || (end == NULL) || (*end != '\0'))
 587                 return 0;
 588         if (n == 0)
 589                 return 0;
 590         if (n >= MAX_RX_QUEUE_PER_LCORE)
 591                 return 0;
 592
 593         return n;
 594 }
 595
 596 static int
 597 l2fwd_parse_timer_period(const char *q_arg)
 598 {
 599         char *end = NULL;
 600         int n;
 601
 602         /* parse number string */
 603         n = strtol(q_arg, &end, 10);
 604         if ((q_arg[0] == '\0') || (end == NULL) || (*end != '\0'))
 605                 return -1;
 606         if (n >= MAX_TIMER_PERIOD)
 607                 return -1;
 608
 609         return n;
 610 }
 611
 612 /* Parse the argument given in the command line of the application */
 613 static int
 614 l2fwd_parse_args(int argc, char **argv)
 615 {
 616         int opt, ret;
 617         char **argvopt;
 618         int option_index;
 619         char *prgname = argv[0];
 620         static struct option lgopts[] = {
 621                 {NULL, 0, 0, 0}
 622         };
 623
 624         argvopt = argv;
 625
 626         while ((opt = getopt_long(argc, argvopt, "p:q:T:l",
 627                                   lgopts, &option_index)) != EOF) {
 628
 629                 switch (opt) {
 630                 /* portmask */
 631                 case 'p':
 632                         l2fwd_enabled_port_mask = l2fwd_parse_portmask(optarg);
 633                         if (l2fwd_enabled_port_mask == 0) {
 634                                 printf("invalid portmask\n");
 635                                 l2fwd_usage(prgname);
 636                                 return -1;
 637                         }
 638                         break;
 639
 640                 /* nqueue */
 641                 case 'q':
 642                         l2fwd_rx_queue_per_lcore = l2fwd_parse_nqueue(optarg);
 643                         if (l2fwd_rx_queue_per_lcore == 0) {
 644                                 printf("invalid queue number\n");
 645                                 l2fwd_usage(prgname);
 646                                 return -1;
 647                         }
 648                         break;
 649
 650                 /* timer period */
 651                 case 'T':
 652                         timer_period = l2fwd_parse_timer_period(optarg);
 653                         if (timer_period < 0) {
 654                                 printf("invalid timer period\n");
 655                                 l2fwd_usage(prgname);
 656                                 return -1;
 657                         }
 658                         break;
 659
 660                 /* For thousands separator in printf. */
 661                 case 'l':
 662                         setlocale(LC_ALL, "");
 663                         break;
 664
 665                 /* long options */
 666                 case 0:
 667                         l2fwd_usage(prgname);
 668                         return -1;
 669
 670                 default:
 671                         l2fwd_usage(prgname);
 672                         return -1;
 673                 }
 674         }
 675
 676         if (optind >= 0)
 677                 argv[optind-1] = prgname;
 678
 679         ret = optind-1;
 680         optind = 1; /* reset getopt lib */
 681         return ret;
 682 }
 683
 684 /* Check the link status of all ports in up to 9s, and print them finally */
 685 static void
 686 check_all_ports_link_status(uint16_t port_num, uint32_t port_mask)
 687 {
 688 #define CHECK_INTERVAL 100 /* 100ms */
 689 #define MAX_CHECK_TIME 90 /* 9s (90 * 100ms) in total */
 690         uint16_t portid;
 691         uint8_t count, all_ports_up, print_flag = 0;
 692         struct rte_eth_link link;
 693
 694         printf("\nChecking link status");
 695         fflush(stdout);
 696         for (count = 0; count <= MAX_CHECK_TIME; count++) {
 697                 all_ports_up = 1;
 698                 for (portid = 0; portid < port_num; portid++) {
 699                         if ((port_mask & (1 << portid)) == 0)
 700                                 continue;
 701                         memset(&link, 0, sizeof(link));
 702                         rte_eth_link_get_nowait(portid, &link);
 703                         /* print link status if flag set */
 704                         if (print_flag == 1) {
 705                                 if (link.link_status)
 706                                         printf(
 707                                         "Port%d Link Up. Speed %u Mbps - %s\n",
 708                                                 portid, link.link_speed,
 709                                 (link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
 710                                         ("full-duplex") : ("half-duplex\n"));
 711                                 else
 712                                         printf("Port %d Link Down\n", portid);
 713                                 continue;
 714                         }
 715                         /* clear all_ports_up flag if any link down */
 716                         if (link.link_status == ETH_LINK_DOWN) {
 717                                 all_ports_up = 0;
 718                                 break;
 719                         }
 720                 }
 721                 /* after finally printing all link status, get out */
 722                 if (print_flag == 1)
 723                         break;
 724
 725                 if (all_ports_up == 0) {
 726                         printf(".");
 727                         fflush(stdout);
 728                         rte_delay_ms(CHECK_INTERVAL);
 729                 }
 730
 731                 /* set the print_flag if all ports up or timeout */
 732                 if (all_ports_up == 1 || count == (MAX_CHECK_TIME - 1)) {
 733                         print_flag = 1;
 734                         printf("done\n");
 735                 }
 736         }
 737 }
 738
 739 int
 740 main(int argc, char **argv)
 741 {
 742         struct lcore_queue_conf *qconf;
 743         unsigned lcore_id, rx_lcore_id;
 744         unsigned nb_ports_in_mask = 0;
 745         int ret;
 746         char name[RTE_JOBSTATS_NAMESIZE];
 747         uint16_t nb_ports;
 748         uint16_t nb_ports_available;
 749         uint16_t portid, last_port;
 750         uint8_t i;
 751
 752         /* init EAL */
 753         ret = rte_eal_init(argc, argv);
 754         if (ret < 0)
 755                 rte_exit(EXIT_FAILURE, "Invalid EAL arguments\n");
 756         argc -= ret;
 757         argv += ret;
 758
 759         /* parse application arguments (after the EAL ones) */
 760         ret = l2fwd_parse_args(argc, argv);
 761         if (ret < 0)
 762                 rte_exit(EXIT_FAILURE, "Invalid L2FWD arguments\n");
 763
 764         rte_timer_subsystem_init();
 765
 766         /* fetch default timer frequency. */
 767         hz = rte_get_timer_hz();
 768
 769         /* create the mbuf pool */
 770         l2fwd_pktmbuf_pool =
 771                 rte_pktmbuf_pool_create("mbuf_pool", NB_MBUF, 32,
 772                         0, RTE_MBUF_DEFAULT_BUF_SIZE, rte_socket_id());
 773         if (l2fwd_pktmbuf_pool == NULL)
 774                 rte_exit(EXIT_FAILURE, "Cannot init mbuf pool\n");
 775
 776         nb_ports = rte_eth_dev_count();
 777         if (nb_ports == 0)
 778                 rte_exit(EXIT_FAILURE, "No Ethernet ports - bye\n");
 779
 780         /* reset l2fwd_dst_ports */
 781         for (portid = 0; portid < RTE_MAX_ETHPORTS; portid++)
 782                 l2fwd_dst_ports[portid] = 0;
 783         last_port = 0;
 784
 785         /*
 786          * Each logical core is assigned a dedicated TX queue on each port.
 787          */
 788         for (portid = 0; portid < nb_ports; portid++) {
 789                 /* skip ports that are not enabled */
 790                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
 791                         continue;
 792
 793                 if (nb_ports_in_mask % 2) {
 794                         l2fwd_dst_ports[portid] = last_port;
 795                         l2fwd_dst_ports[last_port] = portid;
 796                 } else
 797                         last_port = portid;
 798
 799                 nb_ports_in_mask++;
 800         }
 801         if (nb_ports_in_mask % 2) {
 802                 printf("Notice: odd number of ports in portmask.\n");
 803                 l2fwd_dst_ports[last_port] = last_port;
 804         }
 805
 806         rx_lcore_id = 0;
 807         qconf = NULL;
 808
 809         /* Initialize the port/queue configuration of each logical core */
 810         for (portid = 0; portid < nb_ports; portid++) {
 811                 /* skip ports that are not enabled */
 812                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
 813                         continue;
 814
 815                 /* get the lcore_id for this port */
 816                 while (rte_lcore_is_enabled(rx_lcore_id) == 0 ||
 817                        lcore_queue_conf[rx_lcore_id].n_rx_port ==
 818                        l2fwd_rx_queue_per_lcore) {
 819                         rx_lcore_id++;
 820                         if (rx_lcore_id >= RTE_MAX_LCORE)
 821                                 rte_exit(EXIT_FAILURE, "Not enough cores\n");
 822                 }
 823
 824                 if (qconf != &lcore_queue_conf[rx_lcore_id])
 825                         /* Assigned a new logical core in the loop above. */
 826                         qconf = &lcore_queue_conf[rx_lcore_id];
 827
 828                 qconf->rx_port_list[qconf->n_rx_port] = portid;
 829                 qconf->n_rx_port++;
 830                 printf("Lcore %u: RX port %u\n", rx_lcore_id, portid);
 831         }
 832
 833         nb_ports_available = nb_ports;
 834
 835         /* Initialise each port */
 836         for (portid = 0; portid < nb_ports; portid++) {
 837                 /* skip ports that are not enabled */
 838                 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0) {
 839                         printf("Skipping disabled port %u\n", portid);
 840                         nb_ports_available--;
 841                         continue;
 842                 }
 843                 /* init port */
 844                 printf("Initializing port %u... ", portid);
 845                 fflush(stdout);
 846                 ret = rte_eth_dev_configure(portid, 1, 1, &port_conf);
 847                 if (ret < 0)
 848                         rte_exit(EXIT_FAILURE, "Cannot configure device: err=%d, port=%u\n",
 849                                   ret, portid);
 850
 851                 ret = rte_eth_dev_adjust_nb_rx_tx_desc(portid, &nb_rxd,
 852                                                        &nb_txd);
 853                 if (ret < 0)
 854                         rte_exit(EXIT_FAILURE,
 855                                  "Cannot adjust number of descriptors: err=%d, port=%u\n",
 856                                  ret, portid);
 857
 858                 rte_eth_macaddr_get(portid, &l2fwd_ports_eth_addr[portid]);
 859
 860                 /* init one RX queue */
 861                 fflush(stdout);
 862                 ret = rte_eth_rx_queue_setup(portid, 0, nb_rxd,
 863                                              rte_eth_dev_socket_id(portid),
 864                                              NULL,
 865                                              l2fwd_pktmbuf_pool);
 866                 if (ret < 0)
 867                         rte_exit(EXIT_FAILURE, "rte_eth_rx_queue_setup:err=%d, port=%u\n",
 868                                   ret, portid);
 869
 870                 /* init one TX queue on each port */
 871                 fflush(stdout);
 872                 ret = rte_eth_tx_queue_setup(portid, 0, nb_txd,
 873                                 rte_eth_dev_socket_id(portid),
 874                                 NULL);
 875                 if (ret < 0)
 876                         rte_exit(EXIT_FAILURE,
 877                         "rte_eth_tx_queue_setup:err=%d, port=%u\n",
 878                                 ret, portid);
 879
 880                 /* Initialize TX buffers */
 881                 tx_buffer[portid] = rte_zmalloc_socket("tx_buffer",
 882                                 RTE_ETH_TX_BUFFER_SIZE(MAX_PKT_BURST), 0,
 883                                 rte_eth_dev_socket_id(portid));
 884                 if (tx_buffer[portid] == NULL)
 885                         rte_exit(EXIT_FAILURE, "Cannot allocate buffer for tx on port %u\n",
 886                                         portid);
 887
 888                 rte_eth_tx_buffer_init(tx_buffer[portid], MAX_PKT_BURST);
 889
 890                 ret = rte_eth_tx_buffer_set_err_callback(tx_buffer[portid],
 891                                 rte_eth_tx_buffer_count_callback,
 892                                 &port_statistics[portid].dropped);
 893                 if (ret < 0)
 894                         rte_exit(EXIT_FAILURE,
 895                         "Cannot set error callback for tx buffer on port %u\n",
 896                                  portid);
 897
 898                 /* Start device */
 899                 ret = rte_eth_dev_start(portid);
 900                 if (ret < 0)
 901                         rte_exit(EXIT_FAILURE, "rte_eth_dev_start:err=%d, port=%u\n",
 902                                   ret, portid);
 903
 904                 printf("done:\n");
 905
 906                 rte_eth_promiscuous_enable(portid);
 907
 908                 printf("Port %u, MAC address: %02X:%02X:%02X:%02X:%02X:%02X\n\n",
 909                                 portid,
 910                                 l2fwd_ports_eth_addr[portid].addr_bytes[0],
 911                                 l2fwd_ports_eth_addr[portid].addr_bytes[1],
 912                                 l2fwd_ports_eth_addr[portid].addr_bytes[2],
 913                                 l2fwd_ports_eth_addr[portid].addr_bytes[3],
 914                                 l2fwd_ports_eth_addr[portid].addr_bytes[4],
 915                                 l2fwd_ports_eth_addr[portid].addr_bytes[5]);
 916
 917                 /* initialize port stats */
 918                 memset(&port_statistics, 0, sizeof(port_statistics));
 919         }
 920
 921         if (!nb_ports_available) {
 922                 rte_exit(EXIT_FAILURE,
 923                         "All available ports are disabled. Please set portmask.\n");
 924         }
 925
 926         check_all_ports_link_status(nb_ports, l2fwd_enabled_port_mask);
 927
 928         drain_tsc = (hz + US_PER_S - 1) / US_PER_S * BURST_TX_DRAIN_US;
 929
 930         RTE_LCORE_FOREACH(lcore_id) {
 931                 qconf = &lcore_queue_conf[lcore_id];
 932
 933                 rte_spinlock_init(&qconf->lock);
 934
 935                 if (rte_jobstats_context_init(&qconf->jobs_context) != 0)
 936                         rte_panic("Jobs stats context for core %u init failed\n", lcore_id);
 937
 938                 if (qconf->n_rx_port == 0) {
 939                         RTE_LOG(INFO, L2FWD,
 940                                 "lcore %u: no ports so no jobs stats context initialization\n",
 941                                 lcore_id);
 942                         continue;
 943                 }
 944                 /* Add flush job.
 945                  * Set fixed period by setting min = max = initial period. Set target to
 946                  * zero as it is irrelevant for this job. */
 947                 rte_jobstats_init(&qconf->flush_job, "flush", drain_tsc, drain_tsc,
 948                                 drain_tsc, 0);
 949
 950                 rte_timer_init(&qconf->flush_timer);
 951                 ret = rte_timer_reset(&qconf->flush_timer, drain_tsc, PERIODICAL,
 952                                 lcore_id, &l2fwd_flush_job, NULL);
 953
 954                 if (ret < 0) {
 955                         rte_exit(1, "Failed to reset flush job timer for lcore %u: %s",
 956                                         lcore_id, rte_strerror(-ret));
 957                 }
 958
 959                 for (i = 0; i < qconf->n_rx_port; i++) {
 960                         struct rte_jobstats *job = &qconf->port_fwd_jobs[i];
 961
 962                         portid = qconf->rx_port_list[i];
 963                         printf("Setting forward job for port %u\n", portid);
 964
 965                         snprintf(name, RTE_DIM(name), "port %u fwd", portid);
 966                         /* Setup forward job.
 967                          * Set min, max and initial period. Set target to MAX_PKT_BURST as
 968                          * this is desired optimal RX/TX burst size. */
 969                         rte_jobstats_init(job, name, 0, drain_tsc, 0, MAX_PKT_BURST);
 970                         rte_jobstats_set_update_period_function(job, l2fwd_job_update_cb);
 971
 972                         rte_timer_init(&qconf->rx_timers[i]);
 973                         ret = rte_timer_reset(&qconf->rx_timers[i], 0, PERIODICAL, lcore_id,
 974                                         &l2fwd_fwd_job, (void *)(uintptr_t)i);
 975
 976                         if (ret < 0) {
 977                                 rte_exit(1, "Failed to reset lcore %u port %u job timer: %s",
 978                                                 lcore_id, qconf->rx_port_list[i], rte_strerror(-ret));
 979                         }
 980                 }
 981         }
 982
 983         if (timer_period)
 984                 rte_eal_alarm_set(timer_period * MS_PER_S, show_stats_cb, NULL);
 985         else
 986                 RTE_LOG(INFO, L2FWD, "Stats display disabled\n");
 987
 988         /* launch per-lcore init on every lcore */
 989         rte_eal_mp_remote_launch(l2fwd_launch_one_lcore, NULL, CALL_MASTER);
 990         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 991                 if (rte_eal_wait_lcore(lcore_id) < 0)
 992                         return -1;
 993         }
 994
 995         return 0;
 996 }