app/test-eventdev/test_perf_common.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2017 Cavium, Inc
   3  */
   4
   5 #include "test_perf_common.h"
   6
   7 int
   8 perf_test_result(struct evt_test *test, struct evt_options *opt)
   9 {
  10         RTE_SET_USED(opt);
  11         int i;
  12         uint64_t total = 0;
  13         struct test_perf *t = evt_test_priv(test);
  14
  15         printf("Packet distribution across worker cores :\n");
  16         for (i = 0; i < t->nb_workers; i++)
  17                 total += t->worker[i].processed_pkts;
  18         for (i = 0; i < t->nb_workers; i++)
  19                 printf("Worker %d packets: "CLGRN"%"PRIx64" "CLNRM"percentage:"
  20                                 CLGRN" %3.2f\n"CLNRM, i,
  21                                 t->worker[i].processed_pkts,
  22                                 (((double)t->worker[i].processed_pkts)/total)
  23                                 * 100);
  24
  25         return t->result;
  26 }
  27
  28 static inline int
  29 perf_producer(void *arg)
  30 {
  31         struct prod_data *p  = arg;
  32         struct test_perf *t = p->t;
  33         struct evt_options *opt = t->opt;
  34         const uint8_t dev_id = p->dev_id;
  35         const uint8_t port = p->port_id;
  36         struct rte_mempool *pool = t->pool;
  37         const uint64_t nb_pkts = t->nb_pkts;
  38         const uint32_t nb_flows = t->nb_flows;
  39         uint32_t flow_counter = 0;
  40         uint64_t count = 0;
  41         struct perf_elt *m;
  42         struct rte_event ev;
  43
  44         if (opt->verbose_level > 1)
  45                 printf("%s(): lcore %d dev_id %d port=%d queue %d\n", __func__,
  46                                 rte_lcore_id(), dev_id, port, p->queue_id);
  47
  48         ev.event = 0;
  49         ev.op = RTE_EVENT_OP_NEW;
  50         ev.queue_id = p->queue_id;
  51         ev.sched_type = t->opt->sched_type_list[0];
  52         ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
  53         ev.event_type =  RTE_EVENT_TYPE_CPU;
  54         ev.sub_event_type = 0; /* stage 0 */
  55
  56         while (count < nb_pkts && t->done == false) {
  57                 if (rte_mempool_get(pool, (void **)&m) < 0)
  58                         continue;
  59
  60                 ev.flow_id = flow_counter++ % nb_flows;
  61                 ev.event_ptr = m;
  62                 m->timestamp = rte_get_timer_cycles();
  63                 while (rte_event_enqueue_burst(dev_id, port, &ev, 1) != 1) {
  64                         if (t->done)
  65                                 break;
  66                         rte_pause();
  67                         m->timestamp = rte_get_timer_cycles();
  68                 }
  69                 count++;
  70         }
  71
  72         return 0;
  73 }
  74
  75 static inline int
  76 perf_event_timer_producer(void *arg)
  77 {
  78         struct prod_data *p  = arg;
  79         struct test_perf *t = p->t;
  80         struct evt_options *opt = t->opt;
  81         uint32_t flow_counter = 0;
  82         uint64_t count = 0;
  83         uint64_t arm_latency = 0;
  84         const uint8_t nb_timer_adptrs = opt->nb_timer_adptrs;
  85         const uint32_t nb_flows = t->nb_flows;
  86         const uint64_t nb_timers = opt->nb_timers;
  87         struct rte_mempool *pool = t->pool;
  88         struct perf_elt *m;
  89         struct rte_event_timer_adapter **adptr = t->timer_adptr;
  90         struct rte_event_timer tim;
  91         uint64_t timeout_ticks = opt->expiry_nsec / opt->timer_tick_nsec;
  92
  93         memset(&tim, 0, sizeof(struct rte_event_timer));
  94         timeout_ticks = opt->optm_timer_tick_nsec ?
  95                         (timeout_ticks * opt->timer_tick_nsec)
  96                         / opt->optm_timer_tick_nsec : timeout_ticks;
  97         timeout_ticks += timeout_ticks ? 0 : 1;
  98         tim.ev.event_type =  RTE_EVENT_TYPE_TIMER;
  99         tim.ev.op = RTE_EVENT_OP_NEW;
 100         tim.ev.sched_type = t->opt->sched_type_list[0];
 101         tim.ev.queue_id = p->queue_id;
 102         tim.ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
 103         tim.state = RTE_EVENT_TIMER_NOT_ARMED;
 104         tim.timeout_ticks = timeout_ticks;
 105
 106         if (opt->verbose_level > 1)
 107                 printf("%s(): lcore %d\n", __func__, rte_lcore_id());
 108
 109         while (count < nb_timers && t->done == false) {
 110                 if (rte_mempool_get(pool, (void **)&m) < 0)
 111                         continue;
 112
 113                 m->tim = tim;
 114                 m->tim.ev.flow_id = flow_counter++ % nb_flows;
 115                 m->tim.ev.event_ptr = m;
 116                 m->timestamp = rte_get_timer_cycles();
 117                 while (rte_event_timer_arm_burst(
 118                                 adptr[flow_counter % nb_timer_adptrs],
 119                                 (struct rte_event_timer **)&m, 1) != 1) {
 120                         if (t->done)
 121                                 break;
 122                         rte_pause();
 123                         m->timestamp = rte_get_timer_cycles();
 124                 }
 125                 arm_latency += rte_get_timer_cycles() - m->timestamp;
 126                 count++;
 127         }
 128         fflush(stdout);
 129         rte_delay_ms(1000);
 130         printf("%s(): lcore %d Average event timer arm latency = %.3f us\n",
 131                         __func__, rte_lcore_id(), (float)(arm_latency / count) /
 132                         (rte_get_timer_hz() / 1000000));
 133         return 0;
 134 }
 135
 136 static inline int
 137 perf_event_timer_producer_burst(void *arg)
 138 {
 139         int i;
 140         struct prod_data *p  = arg;
 141         struct test_perf *t = p->t;
 142         struct evt_options *opt = t->opt;
 143         uint32_t flow_counter = 0;
 144         uint64_t count = 0;
 145         uint64_t arm_latency = 0;
 146         const uint8_t nb_timer_adptrs = opt->nb_timer_adptrs;
 147         const uint32_t nb_flows = t->nb_flows;
 148         const uint64_t nb_timers = opt->nb_timers;
 149         struct rte_mempool *pool = t->pool;
 150         struct perf_elt *m[BURST_SIZE + 1] = {NULL};
 151         struct rte_event_timer_adapter **adptr = t->timer_adptr;
 152         struct rte_event_timer tim;
 153         uint64_t timeout_ticks = opt->expiry_nsec / opt->timer_tick_nsec;
 154
 155         memset(&tim, 0, sizeof(struct rte_event_timer));
 156         timeout_ticks = opt->optm_timer_tick_nsec ?
 157                         (timeout_ticks * opt->timer_tick_nsec)
 158                         / opt->optm_timer_tick_nsec : timeout_ticks;
 159         timeout_ticks += timeout_ticks ? 0 : 1;
 160         tim.ev.event_type =  RTE_EVENT_TYPE_TIMER;
 161         tim.ev.op = RTE_EVENT_OP_NEW;
 162         tim.ev.sched_type = t->opt->sched_type_list[0];
 163         tim.ev.queue_id = p->queue_id;
 164         tim.ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
 165         tim.state = RTE_EVENT_TIMER_NOT_ARMED;
 166         tim.timeout_ticks = timeout_ticks;
 167
 168         if (opt->verbose_level > 1)
 169                 printf("%s(): lcore %d\n", __func__, rte_lcore_id());
 170
 171         while (count < nb_timers && t->done == false) {
 172                 if (rte_mempool_get_bulk(pool, (void **)m, BURST_SIZE) < 0)
 173                         continue;
 174                 for (i = 0; i < BURST_SIZE; i++) {
 175                         rte_prefetch0(m[i + 1]);
 176                         m[i]->tim = tim;
 177                         m[i]->tim.ev.flow_id = flow_counter++ % nb_flows;
 178                         m[i]->tim.ev.event_ptr = m[i];
 179                         m[i]->timestamp = rte_get_timer_cycles();
 180                 }
 181                 rte_event_timer_arm_tmo_tick_burst(
 182                                 adptr[flow_counter % nb_timer_adptrs],
 183                                 (struct rte_event_timer **)m,
 184                                 tim.timeout_ticks,
 185                                 BURST_SIZE);
 186                 arm_latency += rte_get_timer_cycles() - m[i - 1]->timestamp;
 187                 count += BURST_SIZE;
 188         }
 189         fflush(stdout);
 190         rte_delay_ms(1000);
 191         printf("%s(): lcore %d Average event timer arm latency = %.3f us\n",
 192                         __func__, rte_lcore_id(), (float)(arm_latency / count) /
 193                         (rte_get_timer_hz() / 1000000));
 194         return 0;
 195 }
 196
 197 static int
 198 perf_producer_wrapper(void *arg)
 199 {
 200         struct prod_data *p  = arg;
 201         struct test_perf *t = p->t;
 202         /* Launch the producer function only in case of synthetic producer. */
 203         if (t->opt->prod_type == EVT_PROD_TYPE_SYNT)
 204                 return perf_producer(arg);
 205         else if (t->opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR &&
 206                         !t->opt->timdev_use_burst)
 207                 return perf_event_timer_producer(arg);
 208         else if (t->opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR &&
 209                         t->opt->timdev_use_burst)
 210                 return perf_event_timer_producer_burst(arg);
 211         return 0;
 212 }
 213
 214 static inline uint64_t
 215 processed_pkts(struct test_perf *t)
 216 {
 217         uint8_t i;
 218         uint64_t total = 0;
 219
 220         rte_smp_rmb();
 221         for (i = 0; i < t->nb_workers; i++)
 222                 total += t->worker[i].processed_pkts;
 223
 224         return total;
 225 }
 226
 227 static inline uint64_t
 228 total_latency(struct test_perf *t)
 229 {
 230         uint8_t i;
 231         uint64_t total = 0;
 232
 233         rte_smp_rmb();
 234         for (i = 0; i < t->nb_workers; i++)
 235                 total += t->worker[i].latency;
 236
 237         return total;
 238 }
 239
 240
 241 int
 242 perf_launch_lcores(struct evt_test *test, struct evt_options *opt,
 243                 int (*worker)(void *))
 244 {
 245         int ret, lcore_id;
 246         struct test_perf *t = evt_test_priv(test);
 247
 248         int port_idx = 0;
 249         /* launch workers */
 250         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 251                 if (!(opt->wlcores[lcore_id]))
 252                         continue;
 253
 254                 ret = rte_eal_remote_launch(worker,
 255                                  &t->worker[port_idx], lcore_id);
 256                 if (ret) {
 257                         evt_err("failed to launch worker %d", lcore_id);
 258                         return ret;
 259                 }
 260                 port_idx++;
 261         }
 262
 263         /* launch producers */
 264         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 265                 if (!(opt->plcores[lcore_id]))
 266                         continue;
 267
 268                 ret = rte_eal_remote_launch(perf_producer_wrapper,
 269                                 &t->prod[port_idx], lcore_id);
 270                 if (ret) {
 271                         evt_err("failed to launch perf_producer %d", lcore_id);
 272                         return ret;
 273                 }
 274                 port_idx++;
 275         }
 276
 277         const uint64_t total_pkts = t->outstand_pkts;
 278
 279         uint64_t dead_lock_cycles = rte_get_timer_cycles();
 280         int64_t dead_lock_remaining  =  total_pkts;
 281         const uint64_t dead_lock_sample = rte_get_timer_hz() * 5;
 282
 283         uint64_t perf_cycles = rte_get_timer_cycles();
 284         int64_t perf_remaining  = total_pkts;
 285         const uint64_t perf_sample = rte_get_timer_hz();
 286
 287         static float total_mpps;
 288         static uint64_t samples;
 289
 290         const uint64_t freq_mhz = rte_get_timer_hz() / 1000000;
 291         int64_t remaining = t->outstand_pkts - processed_pkts(t);
 292
 293         while (t->done == false) {
 294                 const uint64_t new_cycles = rte_get_timer_cycles();
 295
 296                 if ((new_cycles - perf_cycles) > perf_sample) {
 297                         const uint64_t latency = total_latency(t);
 298                         const uint64_t pkts = processed_pkts(t);
 299
 300                         remaining = t->outstand_pkts - pkts;
 301                         float mpps = (float)(perf_remaining-remaining)/1000000;
 302
 303                         perf_remaining = remaining;
 304                         perf_cycles = new_cycles;
 305                         total_mpps += mpps;
 306                         ++samples;
 307                         if (opt->fwd_latency && pkts > 0) {
 308                                 printf(CLGRN"\r%.3f mpps avg %.3f mpps [avg fwd latency %.3f us] "CLNRM,
 309                                         mpps, total_mpps/samples,
 310                                         (float)(latency/pkts)/freq_mhz);
 311                         } else {
 312                                 printf(CLGRN"\r%.3f mpps avg %.3f mpps"CLNRM,
 313                                         mpps, total_mpps/samples);
 314                         }
 315                         fflush(stdout);
 316
 317                         if (remaining <= 0) {
 318                                 t->result = EVT_TEST_SUCCESS;
 319                                 if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 320                                         opt->prod_type ==
 321                                         EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 322                                         t->done = true;
 323                                         rte_smp_wmb();
 324                                         break;
 325                                 }
 326                         }
 327                 }
 328
 329                 if (new_cycles - dead_lock_cycles > dead_lock_sample &&
 330                     (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 331                      opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR)) {
 332                         remaining = t->outstand_pkts - processed_pkts(t);
 333                         if (dead_lock_remaining == remaining) {
 334                                 rte_event_dev_dump(opt->dev_id, stdout);
 335                                 evt_err("No schedules for seconds, deadlock");
 336                                 t->done = true;
 337                                 rte_smp_wmb();
 338                                 break;
 339                         }
 340                         dead_lock_remaining = remaining;
 341                         dead_lock_cycles = new_cycles;
 342                 }
 343         }
 344         printf("\n");
 345         return 0;
 346 }
 347
 348 static int
 349 perf_event_rx_adapter_setup(struct evt_options *opt, uint8_t stride,
 350                 struct rte_event_port_conf prod_conf)
 351 {
 352         int ret = 0;
 353         uint16_t prod;
 354         struct rte_event_eth_rx_adapter_queue_conf queue_conf;
 355
 356         memset(&queue_conf, 0,
 357                         sizeof(struct rte_event_eth_rx_adapter_queue_conf));
 358         queue_conf.ev.sched_type = opt->sched_type_list[0];
 359         RTE_ETH_FOREACH_DEV(prod) {
 360                 uint32_t cap;
 361
 362                 ret = rte_event_eth_rx_adapter_caps_get(opt->dev_id,
 363                                 prod, &cap);
 364                 if (ret) {
 365                         evt_err("failed to get event rx adapter[%d]"
 366                                         " capabilities",
 367                                         opt->dev_id);
 368                         return ret;
 369                 }
 370                 queue_conf.ev.queue_id = prod * stride;
 371                 ret = rte_event_eth_rx_adapter_create(prod, opt->dev_id,
 372                                 &prod_conf);
 373                 if (ret) {
 374                         evt_err("failed to create rx adapter[%d]", prod);
 375                         return ret;
 376                 }
 377                 ret = rte_event_eth_rx_adapter_queue_add(prod, prod, -1,
 378                                 &queue_conf);
 379                 if (ret) {
 380                         evt_err("failed to add rx queues to adapter[%d]", prod);
 381                         return ret;
 382                 }
 383
 384                 if (!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT)) {
 385                         uint32_t service_id;
 386
 387                         rte_event_eth_rx_adapter_service_id_get(prod,
 388                                         &service_id);
 389                         ret = evt_service_setup(service_id);
 390                         if (ret) {
 391                                 evt_err("Failed to setup service core"
 392                                                 " for Rx adapter\n");
 393                                 return ret;
 394                         }
 395                 }
 396         }
 397
 398         return ret;
 399 }
 400
 401 static int
 402 perf_event_timer_adapter_setup(struct test_perf *t)
 403 {
 404         int i;
 405         int ret;
 406         struct rte_event_timer_adapter_info adapter_info;
 407         struct rte_event_timer_adapter *wl;
 408         uint8_t nb_producers = evt_nr_active_lcores(t->opt->plcores);
 409         uint8_t flags = RTE_EVENT_TIMER_ADAPTER_F_ADJUST_RES;
 410
 411         if (nb_producers == 1)
 412                 flags |= RTE_EVENT_TIMER_ADAPTER_F_SP_PUT;
 413
 414         for (i = 0; i < t->opt->nb_timer_adptrs; i++) {
 415                 struct rte_event_timer_adapter_conf config = {
 416                         .event_dev_id = t->opt->dev_id,
 417                         .timer_adapter_id = i,
 418                         .timer_tick_ns = t->opt->timer_tick_nsec,
 419                         .max_tmo_ns = t->opt->max_tmo_nsec,
 420                         .nb_timers = t->opt->pool_sz,
 421                         .flags = flags,
 422                 };
 423
 424                 wl = rte_event_timer_adapter_create(&config);
 425                 if (wl == NULL) {
 426                         evt_err("failed to create event timer ring %d", i);
 427                         return rte_errno;
 428                 }
 429
 430                 memset(&adapter_info, 0,
 431                                 sizeof(struct rte_event_timer_adapter_info));
 432                 rte_event_timer_adapter_get_info(wl, &adapter_info);
 433                 t->opt->optm_timer_tick_nsec = adapter_info.min_resolution_ns;
 434
 435                 if (!(adapter_info.caps &
 436                                 RTE_EVENT_TIMER_ADAPTER_CAP_INTERNAL_PORT)) {
 437                         uint32_t service_id;
 438
 439                         rte_event_timer_adapter_service_id_get(wl,
 440                                         &service_id);
 441                         ret = evt_service_setup(service_id);
 442                         if (ret) {
 443                                 evt_err("Failed to setup service core"
 444                                                 " for timer adapter\n");
 445                                 return ret;
 446                         }
 447                         rte_service_runstate_set(service_id, 1);
 448                 }
 449                 t->timer_adptr[i] = wl;
 450         }
 451         return 0;
 452 }
 453
 454 int
 455 perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt,
 456                                 uint8_t stride, uint8_t nb_queues,
 457                                 const struct rte_event_port_conf *port_conf)
 458 {
 459         struct test_perf *t = evt_test_priv(test);
 460         uint16_t port, prod;
 461         int ret = -1;
 462
 463         /* setup one port per worker, linking to all queues */
 464         for (port = 0; port < evt_nr_active_lcores(opt->wlcores);
 465                                 port++) {
 466                 struct worker_data *w = &t->worker[port];
 467
 468                 w->dev_id = opt->dev_id;
 469                 w->port_id = port;
 470                 w->t = t;
 471                 w->processed_pkts = 0;
 472                 w->latency = 0;
 473
 474                 ret = rte_event_port_setup(opt->dev_id, port, port_conf);
 475                 if (ret) {
 476                         evt_err("failed to setup port %d", port);
 477                         return ret;
 478                 }
 479
 480                 ret = rte_event_port_link(opt->dev_id, port, NULL, NULL, 0);
 481                 if (ret != nb_queues) {
 482                         evt_err("failed to link all queues to port %d", port);
 483                         return -EINVAL;
 484                 }
 485         }
 486
 487         /* port for producers, no links */
 488         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 489                 for ( ; port < perf_nb_event_ports(opt); port++) {
 490                         struct prod_data *p = &t->prod[port];
 491                         p->t = t;
 492                 }
 493
 494                 ret = perf_event_rx_adapter_setup(opt, stride, *port_conf);
 495                 if (ret)
 496                         return ret;
 497         } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 498                 prod = 0;
 499                 for ( ; port < perf_nb_event_ports(opt); port++) {
 500                         struct prod_data *p = &t->prod[port];
 501                         p->queue_id = prod * stride;
 502                         p->t = t;
 503                         prod++;
 504                 }
 505
 506                 ret = perf_event_timer_adapter_setup(t);
 507                 if (ret)
 508                         return ret;
 509         } else {
 510                 prod = 0;
 511                 for ( ; port < perf_nb_event_ports(opt); port++) {
 512                         struct prod_data *p = &t->prod[port];
 513
 514                         p->dev_id = opt->dev_id;
 515                         p->port_id = port;
 516                         p->queue_id = prod * stride;
 517                         p->t = t;
 518
 519                         ret = rte_event_port_setup(opt->dev_id, port,
 520                                         port_conf);
 521                         if (ret) {
 522                                 evt_err("failed to setup port %d", port);
 523                                 return ret;
 524                         }
 525                         prod++;
 526                 }
 527         }
 528
 529         return ret;
 530 }
 531
 532 int
 533 perf_opt_check(struct evt_options *opt, uint64_t nb_queues)
 534 {
 535         unsigned int lcores;
 536
 537         /* N producer + N worker + 1 master when producer cores are used
 538          * Else N worker + 1 master when Rx adapter is used
 539          */
 540         lcores = opt->prod_type == EVT_PROD_TYPE_SYNT ? 3 : 2;
 541
 542         if (rte_lcore_count() < lcores) {
 543                 evt_err("test need minimum %d lcores", lcores);
 544                 return -1;
 545         }
 546
 547         /* Validate worker lcores */
 548         if (evt_lcores_has_overlap(opt->wlcores, rte_get_master_lcore())) {
 549                 evt_err("worker lcores overlaps with master lcore");
 550                 return -1;
 551         }
 552         if (evt_lcores_has_overlap_multi(opt->wlcores, opt->plcores)) {
 553                 evt_err("worker lcores overlaps producer lcores");
 554                 return -1;
 555         }
 556         if (evt_has_disabled_lcore(opt->wlcores)) {
 557                 evt_err("one or more workers lcores are not enabled");
 558                 return -1;
 559         }
 560         if (!evt_has_active_lcore(opt->wlcores)) {
 561                 evt_err("minimum one worker is required");
 562                 return -1;
 563         }
 564
 565         if (opt->prod_type == EVT_PROD_TYPE_SYNT) {
 566                 /* Validate producer lcores */
 567                 if (evt_lcores_has_overlap(opt->plcores,
 568                                         rte_get_master_lcore())) {
 569                         evt_err("producer lcores overlaps with master lcore");
 570                         return -1;
 571                 }
 572                 if (evt_has_disabled_lcore(opt->plcores)) {
 573                         evt_err("one or more producer lcores are not enabled");
 574                         return -1;
 575                 }
 576                 if (!evt_has_active_lcore(opt->plcores)) {
 577                         evt_err("minimum one producer is required");
 578                         return -1;
 579                 }
 580         }
 581
 582         if (evt_has_invalid_stage(opt))
 583                 return -1;
 584
 585         if (evt_has_invalid_sched_type(opt))
 586                 return -1;
 587
 588         if (nb_queues > EVT_MAX_QUEUES) {
 589                 evt_err("number of queues exceeds %d", EVT_MAX_QUEUES);
 590                 return -1;
 591         }
 592         if (perf_nb_event_ports(opt) > EVT_MAX_PORTS) {
 593                 evt_err("number of ports exceeds %d", EVT_MAX_PORTS);
 594                 return -1;
 595         }
 596
 597         /* Fixups */
 598         if ((opt->nb_stages == 1 &&
 599                         opt->prod_type != EVT_PROD_TYPE_EVENT_TIMER_ADPTR) &&
 600                         opt->fwd_latency) {
 601                 evt_info("fwd_latency is valid when nb_stages > 1, disabling");
 602                 opt->fwd_latency = 0;
 603         }
 604
 605         if (opt->fwd_latency && !opt->q_priority) {
 606                 evt_info("enabled queue priority for latency measurement");
 607                 opt->q_priority = 1;
 608         }
 609         if (opt->nb_pkts == 0)
 610                 opt->nb_pkts = INT64_MAX/evt_nr_active_lcores(opt->plcores);
 611
 612         return 0;
 613 }
 614
 615 void
 616 perf_opt_dump(struct evt_options *opt, uint8_t nb_queues)
 617 {
 618         evt_dump("nb_prod_lcores", "%d", evt_nr_active_lcores(opt->plcores));
 619         evt_dump_producer_lcores(opt);
 620         evt_dump("nb_worker_lcores", "%d", evt_nr_active_lcores(opt->wlcores));
 621         evt_dump_worker_lcores(opt);
 622         evt_dump_nb_stages(opt);
 623         evt_dump("nb_evdev_ports", "%d", perf_nb_event_ports(opt));
 624         evt_dump("nb_evdev_queues", "%d", nb_queues);
 625         evt_dump_queue_priority(opt);
 626         evt_dump_sched_type_list(opt);
 627         evt_dump_producer_type(opt);
 628 }
 629
 630 void
 631 perf_eventdev_destroy(struct evt_test *test, struct evt_options *opt)
 632 {
 633         int i;
 634         struct test_perf *t = evt_test_priv(test);
 635
 636         if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 637                 for (i = 0; i < opt->nb_timer_adptrs; i++)
 638                         rte_event_timer_adapter_stop(t->timer_adptr[i]);
 639         }
 640         rte_event_dev_stop(opt->dev_id);
 641         rte_event_dev_close(opt->dev_id);
 642 }
 643
 644 static inline void
 645 perf_elt_init(struct rte_mempool *mp, void *arg __rte_unused,
 646             void *obj, unsigned i __rte_unused)
 647 {
 648         memset(obj, 0, mp->elt_size);
 649 }
 650
 651 #define NB_RX_DESC                      128
 652 #define NB_TX_DESC                      512
 653 int
 654 perf_ethdev_setup(struct evt_test *test, struct evt_options *opt)
 655 {
 656         uint16_t i;
 657         struct test_perf *t = evt_test_priv(test);
 658         struct rte_eth_conf port_conf = {
 659                 .rxmode = {
 660                         .mq_mode = ETH_MQ_RX_RSS,
 661                         .max_rx_pkt_len = RTE_ETHER_MAX_LEN,
 662                         .split_hdr_size = 0,
 663                 },
 664                 .rx_adv_conf = {
 665                         .rss_conf = {
 666                                 .rss_key = NULL,
 667                                 .rss_hf = ETH_RSS_IP,
 668                         },
 669                 },
 670         };
 671
 672         if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 673                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR)
 674                 return 0;
 675
 676         if (!rte_eth_dev_count_avail()) {
 677                 evt_err("No ethernet ports found.");
 678                 return -ENODEV;
 679         }
 680
 681         RTE_ETH_FOREACH_DEV(i) {
 682                 struct rte_eth_dev_info dev_info;
 683                 struct rte_eth_conf local_port_conf = port_conf;
 684
 685                 rte_eth_dev_info_get(i, &dev_info);
 686
 687                 local_port_conf.rx_adv_conf.rss_conf.rss_hf &=
 688                         dev_info.flow_type_rss_offloads;
 689                 if (local_port_conf.rx_adv_conf.rss_conf.rss_hf !=
 690                                 port_conf.rx_adv_conf.rss_conf.rss_hf) {
 691                         evt_info("Port %u modified RSS hash function based on hardware support,"
 692                                 "requested:%#"PRIx64" configured:%#"PRIx64"\n",
 693                                 i,
 694                                 port_conf.rx_adv_conf.rss_conf.rss_hf,
 695                                 local_port_conf.rx_adv_conf.rss_conf.rss_hf);
 696                 }
 697
 698                 if (rte_eth_dev_configure(i, 1, 1, &local_port_conf) < 0) {
 699                         evt_err("Failed to configure eth port [%d]", i);
 700                         return -EINVAL;
 701                 }
 702
 703                 if (rte_eth_rx_queue_setup(i, 0, NB_RX_DESC,
 704                                 rte_socket_id(), NULL, t->pool) < 0) {
 705                         evt_err("Failed to setup eth port [%d] rx_queue: %d.",
 706                                         i, 0);
 707                         return -EINVAL;
 708                 }
 709
 710                 if (rte_eth_tx_queue_setup(i, 0, NB_TX_DESC,
 711                                         rte_socket_id(), NULL) < 0) {
 712                         evt_err("Failed to setup eth port [%d] tx_queue: %d.",
 713                                         i, 0);
 714                         return -EINVAL;
 715                 }
 716
 717                 rte_eth_promiscuous_enable(i);
 718         }
 719
 720         return 0;
 721 }
 722
 723 void perf_ethdev_destroy(struct evt_test *test, struct evt_options *opt)
 724 {
 725         uint16_t i;
 726         RTE_SET_USED(test);
 727
 728         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 729                 RTE_ETH_FOREACH_DEV(i) {
 730                         rte_event_eth_rx_adapter_stop(i);
 731                         rte_eth_dev_stop(i);
 732                 }
 733         }
 734 }
 735
 736 int
 737 perf_mempool_setup(struct evt_test *test, struct evt_options *opt)
 738 {
 739         struct test_perf *t = evt_test_priv(test);
 740
 741         if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 742                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 743                 t->pool = rte_mempool_create(test->name, /* mempool name */
 744                                 opt->pool_sz, /* number of elements*/
 745                                 sizeof(struct perf_elt), /* element size*/
 746                                 512, /* cache size*/
 747                                 0, NULL, NULL,
 748                                 perf_elt_init, /* obj constructor */
 749                                 NULL, opt->socket_id, 0); /* flags */
 750         } else {
 751                 t->pool = rte_pktmbuf_pool_create(test->name, /* mempool name */
 752                                 opt->pool_sz, /* number of elements*/
 753                                 512, /* cache size*/
 754                                 0,
 755                                 RTE_MBUF_DEFAULT_BUF_SIZE,
 756                                 opt->socket_id); /* flags */
 757
 758         }
 759
 760         if (t->pool == NULL) {
 761                 evt_err("failed to create mempool");
 762                 return -ENOMEM;
 763         }
 764
 765         return 0;
 766 }
 767
 768 void
 769 perf_mempool_destroy(struct evt_test *test, struct evt_options *opt)
 770 {
 771         RTE_SET_USED(opt);
 772         struct test_perf *t = evt_test_priv(test);
 773
 774         rte_mempool_free(t->pool);
 775 }
 776
 777 int
 778 perf_test_setup(struct evt_test *test, struct evt_options *opt)
 779 {
 780         void *test_perf;
 781
 782         test_perf = rte_zmalloc_socket(test->name, sizeof(struct test_perf),
 783                                 RTE_CACHE_LINE_SIZE, opt->socket_id);
 784         if (test_perf  == NULL) {
 785                 evt_err("failed to allocate test_perf memory");
 786                 goto nomem;
 787         }
 788         test->test_priv = test_perf;
 789
 790         struct test_perf *t = evt_test_priv(test);
 791
 792         if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 793                 t->outstand_pkts = opt->nb_timers *
 794                         evt_nr_active_lcores(opt->plcores);
 795                 t->nb_pkts = opt->nb_timers;
 796         } else {
 797                 t->outstand_pkts = opt->nb_pkts *
 798                         evt_nr_active_lcores(opt->plcores);
 799                 t->nb_pkts = opt->nb_pkts;
 800         }
 801
 802         t->nb_workers = evt_nr_active_lcores(opt->wlcores);
 803         t->done = false;
 804         t->nb_flows = opt->nb_flows;
 805         t->result = EVT_TEST_FAILED;
 806         t->opt = opt;
 807         memcpy(t->sched_type_list, opt->sched_type_list,
 808                         sizeof(opt->sched_type_list));
 809         return 0;
 810 nomem:
 811         return -ENOMEM;
 812 }
 813
 814 void
 815 perf_test_destroy(struct evt_test *test, struct evt_options *opt)
 816 {
 817         RTE_SET_USED(opt);
 818
 819         rte_free(test->test_priv);
 820 }