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         int i;
  32         struct prod_data *p  = arg;
  33         struct test_perf *t = p->t;
  34         struct evt_options *opt = t->opt;
  35         const uint8_t dev_id = p->dev_id;
  36         const uint8_t port = p->port_id;
  37         struct rte_mempool *pool = t->pool;
  38         const uint64_t nb_pkts = t->nb_pkts;
  39         const uint32_t nb_flows = t->nb_flows;
  40         uint32_t flow_counter = 0;
  41         uint64_t count = 0;
  42         struct perf_elt *m[BURST_SIZE + 1] = {NULL};
  43         struct rte_event ev;
  44
  45         if (opt->verbose_level > 1)
  46                 printf("%s(): lcore %d dev_id %d port=%d queue %d\n", __func__,
  47                                 rte_lcore_id(), dev_id, port, p->queue_id);
  48
  49         ev.event = 0;
  50         ev.op = RTE_EVENT_OP_NEW;
  51         ev.queue_id = p->queue_id;
  52         ev.sched_type = t->opt->sched_type_list[0];
  53         ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
  54         ev.event_type =  RTE_EVENT_TYPE_CPU;
  55         ev.sub_event_type = 0; /* stage 0 */
  56
  57         while (count < nb_pkts && t->done == false) {
  58                 if (rte_mempool_get_bulk(pool, (void **)m, BURST_SIZE) < 0)
  59                         continue;
  60                 for (i = 0; i < BURST_SIZE; i++) {
  61                         ev.flow_id = flow_counter++ % nb_flows;
  62                         ev.event_ptr = m[i];
  63                         m[i]->timestamp = rte_get_timer_cycles();
  64                         while (rte_event_enqueue_burst(dev_id,
  65                                                        port, &ev, 1) != 1) {
  66                                 if (t->done)
  67                                         break;
  68                                 rte_pause();
  69                                 m[i]->timestamp = rte_get_timer_cycles();
  70                         }
  71                 }
  72                 count += BURST_SIZE;
  73         }
  74
  75         return 0;
  76 }
  77
  78 static inline int
  79 perf_event_timer_producer(void *arg)
  80 {
  81         int i;
  82         struct prod_data *p  = arg;
  83         struct test_perf *t = p->t;
  84         struct evt_options *opt = t->opt;
  85         uint32_t flow_counter = 0;
  86         uint64_t count = 0;
  87         uint64_t arm_latency = 0;
  88         const uint8_t nb_timer_adptrs = opt->nb_timer_adptrs;
  89         const uint32_t nb_flows = t->nb_flows;
  90         const uint64_t nb_timers = opt->nb_timers;
  91         struct rte_mempool *pool = t->pool;
  92         struct perf_elt *m[BURST_SIZE + 1] = {NULL};
  93         struct rte_event_timer_adapter **adptr = t->timer_adptr;
  94         struct rte_event_timer tim;
  95         uint64_t timeout_ticks = opt->expiry_nsec / opt->timer_tick_nsec;
  96
  97         memset(&tim, 0, sizeof(struct rte_event_timer));
  98         timeout_ticks = opt->optm_timer_tick_nsec ?
  99                         (timeout_ticks * opt->timer_tick_nsec)
 100                         / opt->optm_timer_tick_nsec : timeout_ticks;
 101         timeout_ticks += timeout_ticks ? 0 : 1;
 102         tim.ev.event_type =  RTE_EVENT_TYPE_TIMER;
 103         tim.ev.op = RTE_EVENT_OP_NEW;
 104         tim.ev.sched_type = t->opt->sched_type_list[0];
 105         tim.ev.queue_id = p->queue_id;
 106         tim.ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
 107         tim.state = RTE_EVENT_TIMER_NOT_ARMED;
 108         tim.timeout_ticks = timeout_ticks;
 109
 110         if (opt->verbose_level > 1)
 111                 printf("%s(): lcore %d\n", __func__, rte_lcore_id());
 112
 113         while (count < nb_timers && t->done == false) {
 114                 if (rte_mempool_get_bulk(pool, (void **)m, BURST_SIZE) < 0)
 115                         continue;
 116                 for (i = 0; i < BURST_SIZE; i++) {
 117                         rte_prefetch0(m[i + 1]);
 118                         m[i]->tim = tim;
 119                         m[i]->tim.ev.flow_id = flow_counter++ % nb_flows;
 120                         m[i]->tim.ev.event_ptr = m[i];
 121                         m[i]->timestamp = rte_get_timer_cycles();
 122                         while (rte_event_timer_arm_burst(
 123                                adptr[flow_counter % nb_timer_adptrs],
 124                                (struct rte_event_timer **)&m[i], 1) != 1) {
 125                                 if (t->done)
 126                                         break;
 127                                 m[i]->timestamp = rte_get_timer_cycles();
 128                         }
 129                         arm_latency += rte_get_timer_cycles() - m[i]->timestamp;
 130                 }
 131                 count += BURST_SIZE;
 132         }
 133         fflush(stdout);
 134         rte_delay_ms(1000);
 135         printf("%s(): lcore %d Average event timer arm latency = %.3f us\n",
 136                         __func__, rte_lcore_id(), (float)(arm_latency / count) /
 137                         (rte_get_timer_hz() / 1000000));
 138         return 0;
 139 }
 140
 141 static inline int
 142 perf_event_timer_producer_burst(void *arg)
 143 {
 144         int i;
 145         struct prod_data *p  = arg;
 146         struct test_perf *t = p->t;
 147         struct evt_options *opt = t->opt;
 148         uint32_t flow_counter = 0;
 149         uint64_t count = 0;
 150         uint64_t arm_latency = 0;
 151         const uint8_t nb_timer_adptrs = opt->nb_timer_adptrs;
 152         const uint32_t nb_flows = t->nb_flows;
 153         const uint64_t nb_timers = opt->nb_timers;
 154         struct rte_mempool *pool = t->pool;
 155         struct perf_elt *m[BURST_SIZE + 1] = {NULL};
 156         struct rte_event_timer_adapter **adptr = t->timer_adptr;
 157         struct rte_event_timer tim;
 158         uint64_t timeout_ticks = opt->expiry_nsec / opt->timer_tick_nsec;
 159
 160         memset(&tim, 0, sizeof(struct rte_event_timer));
 161         timeout_ticks = opt->optm_timer_tick_nsec ?
 162                         (timeout_ticks * opt->timer_tick_nsec)
 163                         / opt->optm_timer_tick_nsec : timeout_ticks;
 164         timeout_ticks += timeout_ticks ? 0 : 1;
 165         tim.ev.event_type =  RTE_EVENT_TYPE_TIMER;
 166         tim.ev.op = RTE_EVENT_OP_NEW;
 167         tim.ev.sched_type = t->opt->sched_type_list[0];
 168         tim.ev.queue_id = p->queue_id;
 169         tim.ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
 170         tim.state = RTE_EVENT_TIMER_NOT_ARMED;
 171         tim.timeout_ticks = timeout_ticks;
 172
 173         if (opt->verbose_level > 1)
 174                 printf("%s(): lcore %d\n", __func__, rte_lcore_id());
 175
 176         while (count < nb_timers && t->done == false) {
 177                 if (rte_mempool_get_bulk(pool, (void **)m, BURST_SIZE) < 0)
 178                         continue;
 179                 for (i = 0; i < BURST_SIZE; i++) {
 180                         rte_prefetch0(m[i + 1]);
 181                         m[i]->tim = tim;
 182                         m[i]->tim.ev.flow_id = flow_counter++ % nb_flows;
 183                         m[i]->tim.ev.event_ptr = m[i];
 184                         m[i]->timestamp = rte_get_timer_cycles();
 185                 }
 186                 rte_event_timer_arm_tmo_tick_burst(
 187                                 adptr[flow_counter % nb_timer_adptrs],
 188                                 (struct rte_event_timer **)m,
 189                                 tim.timeout_ticks,
 190                                 BURST_SIZE);
 191                 arm_latency += rte_get_timer_cycles() - m[i - 1]->timestamp;
 192                 count += BURST_SIZE;
 193         }
 194         fflush(stdout);
 195         rte_delay_ms(1000);
 196         printf("%s(): lcore %d Average event timer arm latency = %.3f us\n",
 197                         __func__, rte_lcore_id(), (float)(arm_latency / count) /
 198                         (rte_get_timer_hz() / 1000000));
 199         return 0;
 200 }
 201
 202 static int
 203 perf_producer_wrapper(void *arg)
 204 {
 205         struct prod_data *p  = arg;
 206         struct test_perf *t = p->t;
 207         /* Launch the producer function only in case of synthetic producer. */
 208         if (t->opt->prod_type == EVT_PROD_TYPE_SYNT)
 209                 return perf_producer(arg);
 210         else if (t->opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR &&
 211                         !t->opt->timdev_use_burst)
 212                 return perf_event_timer_producer(arg);
 213         else if (t->opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR &&
 214                         t->opt->timdev_use_burst)
 215                 return perf_event_timer_producer_burst(arg);
 216         return 0;
 217 }
 218
 219 static inline uint64_t
 220 processed_pkts(struct test_perf *t)
 221 {
 222         uint8_t i;
 223         uint64_t total = 0;
 224
 225         rte_smp_rmb();
 226         for (i = 0; i < t->nb_workers; i++)
 227                 total += t->worker[i].processed_pkts;
 228
 229         return total;
 230 }
 231
 232 static inline uint64_t
 233 total_latency(struct test_perf *t)
 234 {
 235         uint8_t i;
 236         uint64_t total = 0;
 237
 238         rte_smp_rmb();
 239         for (i = 0; i < t->nb_workers; i++)
 240                 total += t->worker[i].latency;
 241
 242         return total;
 243 }
 244
 245
 246 int
 247 perf_launch_lcores(struct evt_test *test, struct evt_options *opt,
 248                 int (*worker)(void *))
 249 {
 250         int ret, lcore_id;
 251         struct test_perf *t = evt_test_priv(test);
 252
 253         int port_idx = 0;
 254         /* launch workers */
 255         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 256                 if (!(opt->wlcores[lcore_id]))
 257                         continue;
 258
 259                 ret = rte_eal_remote_launch(worker,
 260                                  &t->worker[port_idx], lcore_id);
 261                 if (ret) {
 262                         evt_err("failed to launch worker %d", lcore_id);
 263                         return ret;
 264                 }
 265                 port_idx++;
 266         }
 267
 268         /* launch producers */
 269         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 270                 if (!(opt->plcores[lcore_id]))
 271                         continue;
 272
 273                 ret = rte_eal_remote_launch(perf_producer_wrapper,
 274                                 &t->prod[port_idx], lcore_id);
 275                 if (ret) {
 276                         evt_err("failed to launch perf_producer %d", lcore_id);
 277                         return ret;
 278                 }
 279                 port_idx++;
 280         }
 281
 282         const uint64_t total_pkts = t->outstand_pkts;
 283
 284         uint64_t dead_lock_cycles = rte_get_timer_cycles();
 285         int64_t dead_lock_remaining  =  total_pkts;
 286         const uint64_t dead_lock_sample = rte_get_timer_hz() * 5;
 287
 288         uint64_t perf_cycles = rte_get_timer_cycles();
 289         int64_t perf_remaining  = total_pkts;
 290         const uint64_t perf_sample = rte_get_timer_hz();
 291
 292         static float total_mpps;
 293         static uint64_t samples;
 294
 295         const uint64_t freq_mhz = rte_get_timer_hz() / 1000000;
 296         int64_t remaining = t->outstand_pkts - processed_pkts(t);
 297
 298         while (t->done == false) {
 299                 const uint64_t new_cycles = rte_get_timer_cycles();
 300
 301                 if ((new_cycles - perf_cycles) > perf_sample) {
 302                         const uint64_t latency = total_latency(t);
 303                         const uint64_t pkts = processed_pkts(t);
 304
 305                         remaining = t->outstand_pkts - pkts;
 306                         float mpps = (float)(perf_remaining-remaining)/1000000;
 307
 308                         perf_remaining = remaining;
 309                         perf_cycles = new_cycles;
 310                         total_mpps += mpps;
 311                         ++samples;
 312                         if (opt->fwd_latency && pkts > 0) {
 313                                 printf(CLGRN"\r%.3f mpps avg %.3f mpps [avg fwd latency %.3f us] "CLNRM,
 314                                         mpps, total_mpps/samples,
 315                                         (float)(latency/pkts)/freq_mhz);
 316                         } else {
 317                                 printf(CLGRN"\r%.3f mpps avg %.3f mpps"CLNRM,
 318                                         mpps, total_mpps/samples);
 319                         }
 320                         fflush(stdout);
 321
 322                         if (remaining <= 0) {
 323                                 t->result = EVT_TEST_SUCCESS;
 324                                 if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 325                                         opt->prod_type ==
 326                                         EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 327                                         t->done = true;
 328                                         rte_smp_wmb();
 329                                         break;
 330                                 }
 331                         }
 332                 }
 333
 334                 if (new_cycles - dead_lock_cycles > dead_lock_sample &&
 335                     (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 336                      opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR)) {
 337                         remaining = t->outstand_pkts - processed_pkts(t);
 338                         if (dead_lock_remaining == remaining) {
 339                                 rte_event_dev_dump(opt->dev_id, stdout);
 340                                 evt_err("No schedules for seconds, deadlock");
 341                                 t->done = true;
 342                                 rte_smp_wmb();
 343                                 break;
 344                         }
 345                         dead_lock_remaining = remaining;
 346                         dead_lock_cycles = new_cycles;
 347                 }
 348         }
 349         printf("\n");
 350         return 0;
 351 }
 352
 353 static int
 354 perf_event_rx_adapter_setup(struct evt_options *opt, uint8_t stride,
 355                 struct rte_event_port_conf prod_conf)
 356 {
 357         int ret = 0;
 358         uint16_t prod;
 359         struct rte_event_eth_rx_adapter_queue_conf queue_conf;
 360
 361         memset(&queue_conf, 0,
 362                         sizeof(struct rte_event_eth_rx_adapter_queue_conf));
 363         queue_conf.ev.sched_type = opt->sched_type_list[0];
 364         RTE_ETH_FOREACH_DEV(prod) {
 365                 uint32_t cap;
 366
 367                 ret = rte_event_eth_rx_adapter_caps_get(opt->dev_id,
 368                                 prod, &cap);
 369                 if (ret) {
 370                         evt_err("failed to get event rx adapter[%d]"
 371                                         " capabilities",
 372                                         opt->dev_id);
 373                         return ret;
 374                 }
 375                 queue_conf.ev.queue_id = prod * stride;
 376                 ret = rte_event_eth_rx_adapter_create(prod, opt->dev_id,
 377                                 &prod_conf);
 378                 if (ret) {
 379                         evt_err("failed to create rx adapter[%d]", prod);
 380                         return ret;
 381                 }
 382                 ret = rte_event_eth_rx_adapter_queue_add(prod, prod, -1,
 383                                 &queue_conf);
 384                 if (ret) {
 385                         evt_err("failed to add rx queues to adapter[%d]", prod);
 386                         return ret;
 387                 }
 388
 389                 if (!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT)) {
 390                         uint32_t service_id;
 391
 392                         rte_event_eth_rx_adapter_service_id_get(prod,
 393                                         &service_id);
 394                         ret = evt_service_setup(service_id);
 395                         if (ret) {
 396                                 evt_err("Failed to setup service core"
 397                                                 " for Rx adapter\n");
 398                                 return ret;
 399                         }
 400                 }
 401         }
 402
 403         return ret;
 404 }
 405
 406 static int
 407 perf_event_timer_adapter_setup(struct test_perf *t)
 408 {
 409         int i;
 410         int ret;
 411         struct rte_event_timer_adapter_info adapter_info;
 412         struct rte_event_timer_adapter *wl;
 413         uint8_t nb_producers = evt_nr_active_lcores(t->opt->plcores);
 414         uint8_t flags = RTE_EVENT_TIMER_ADAPTER_F_ADJUST_RES;
 415
 416         if (nb_producers == 1)
 417                 flags |= RTE_EVENT_TIMER_ADAPTER_F_SP_PUT;
 418
 419         for (i = 0; i < t->opt->nb_timer_adptrs; i++) {
 420                 struct rte_event_timer_adapter_conf config = {
 421                         .event_dev_id = t->opt->dev_id,
 422                         .timer_adapter_id = i,
 423                         .timer_tick_ns = t->opt->timer_tick_nsec,
 424                         .max_tmo_ns = t->opt->max_tmo_nsec,
 425                         .nb_timers = t->opt->pool_sz,
 426                         .flags = flags,
 427                 };
 428
 429                 wl = rte_event_timer_adapter_create(&config);
 430                 if (wl == NULL) {
 431                         evt_err("failed to create event timer ring %d", i);
 432                         return rte_errno;
 433                 }
 434
 435                 memset(&adapter_info, 0,
 436                                 sizeof(struct rte_event_timer_adapter_info));
 437                 rte_event_timer_adapter_get_info(wl, &adapter_info);
 438                 t->opt->optm_timer_tick_nsec = adapter_info.min_resolution_ns;
 439
 440                 if (!(adapter_info.caps &
 441                                 RTE_EVENT_TIMER_ADAPTER_CAP_INTERNAL_PORT)) {
 442                         uint32_t service_id;
 443
 444                         rte_event_timer_adapter_service_id_get(wl,
 445                                         &service_id);
 446                         ret = evt_service_setup(service_id);
 447                         if (ret) {
 448                                 evt_err("Failed to setup service core"
 449                                                 " for timer adapter\n");
 450                                 return ret;
 451                         }
 452                         rte_service_runstate_set(service_id, 1);
 453                 }
 454                 t->timer_adptr[i] = wl;
 455         }
 456         return 0;
 457 }
 458
 459 int
 460 perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt,
 461                                 uint8_t stride, uint8_t nb_queues,
 462                                 const struct rte_event_port_conf *port_conf)
 463 {
 464         struct test_perf *t = evt_test_priv(test);
 465         uint16_t port, prod;
 466         int ret = -1;
 467
 468         /* setup one port per worker, linking to all queues */
 469         for (port = 0; port < evt_nr_active_lcores(opt->wlcores);
 470                                 port++) {
 471                 struct worker_data *w = &t->worker[port];
 472
 473                 w->dev_id = opt->dev_id;
 474                 w->port_id = port;
 475                 w->t = t;
 476                 w->processed_pkts = 0;
 477                 w->latency = 0;
 478
 479                 ret = rte_event_port_setup(opt->dev_id, port, port_conf);
 480                 if (ret) {
 481                         evt_err("failed to setup port %d", port);
 482                         return ret;
 483                 }
 484
 485                 ret = rte_event_port_link(opt->dev_id, port, NULL, NULL, 0);
 486                 if (ret != nb_queues) {
 487                         evt_err("failed to link all queues to port %d", port);
 488                         return -EINVAL;
 489                 }
 490         }
 491
 492         /* port for producers, no links */
 493         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 494                 for ( ; port < perf_nb_event_ports(opt); port++) {
 495                         struct prod_data *p = &t->prod[port];
 496                         p->t = t;
 497                 }
 498
 499                 ret = perf_event_rx_adapter_setup(opt, stride, *port_conf);
 500                 if (ret)
 501                         return ret;
 502         } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 503                 prod = 0;
 504                 for ( ; port < perf_nb_event_ports(opt); port++) {
 505                         struct prod_data *p = &t->prod[port];
 506                         p->queue_id = prod * stride;
 507                         p->t = t;
 508                         prod++;
 509                 }
 510
 511                 ret = perf_event_timer_adapter_setup(t);
 512                 if (ret)
 513                         return ret;
 514         } else {
 515                 prod = 0;
 516                 for ( ; port < perf_nb_event_ports(opt); port++) {
 517                         struct prod_data *p = &t->prod[port];
 518
 519                         p->dev_id = opt->dev_id;
 520                         p->port_id = port;
 521                         p->queue_id = prod * stride;
 522                         p->t = t;
 523
 524                         ret = rte_event_port_setup(opt->dev_id, port,
 525                                         port_conf);
 526                         if (ret) {
 527                                 evt_err("failed to setup port %d", port);
 528                                 return ret;
 529                         }
 530                         prod++;
 531                 }
 532         }
 533
 534         return ret;
 535 }
 536
 537 int
 538 perf_opt_check(struct evt_options *opt, uint64_t nb_queues)
 539 {
 540         unsigned int lcores;
 541
 542         /* N producer + N worker + 1 master when producer cores are used
 543          * Else N worker + 1 master when Rx adapter is used
 544          */
 545         lcores = opt->prod_type == EVT_PROD_TYPE_SYNT ? 3 : 2;
 546
 547         if (rte_lcore_count() < lcores) {
 548                 evt_err("test need minimum %d lcores", lcores);
 549                 return -1;
 550         }
 551
 552         /* Validate worker lcores */
 553         if (evt_lcores_has_overlap(opt->wlcores, rte_get_master_lcore())) {
 554                 evt_err("worker lcores overlaps with master lcore");
 555                 return -1;
 556         }
 557         if (evt_lcores_has_overlap_multi(opt->wlcores, opt->plcores)) {
 558                 evt_err("worker lcores overlaps producer lcores");
 559                 return -1;
 560         }
 561         if (evt_has_disabled_lcore(opt->wlcores)) {
 562                 evt_err("one or more workers lcores are not enabled");
 563                 return -1;
 564         }
 565         if (!evt_has_active_lcore(opt->wlcores)) {
 566                 evt_err("minimum one worker is required");
 567                 return -1;
 568         }
 569
 570         if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 571                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 572                 /* Validate producer lcores */
 573                 if (evt_lcores_has_overlap(opt->plcores,
 574                                         rte_get_master_lcore())) {
 575                         evt_err("producer lcores overlaps with master lcore");
 576                         return -1;
 577                 }
 578                 if (evt_has_disabled_lcore(opt->plcores)) {
 579                         evt_err("one or more producer lcores are not enabled");
 580                         return -1;
 581                 }
 582                 if (!evt_has_active_lcore(opt->plcores)) {
 583                         evt_err("minimum one producer is required");
 584                         return -1;
 585                 }
 586         }
 587
 588         if (evt_has_invalid_stage(opt))
 589                 return -1;
 590
 591         if (evt_has_invalid_sched_type(opt))
 592                 return -1;
 593
 594         if (nb_queues > EVT_MAX_QUEUES) {
 595                 evt_err("number of queues exceeds %d", EVT_MAX_QUEUES);
 596                 return -1;
 597         }
 598         if (perf_nb_event_ports(opt) > EVT_MAX_PORTS) {
 599                 evt_err("number of ports exceeds %d", EVT_MAX_PORTS);
 600                 return -1;
 601         }
 602
 603         /* Fixups */
 604         if ((opt->nb_stages == 1 &&
 605                         opt->prod_type != EVT_PROD_TYPE_EVENT_TIMER_ADPTR) &&
 606                         opt->fwd_latency) {
 607                 evt_info("fwd_latency is valid when nb_stages > 1, disabling");
 608                 opt->fwd_latency = 0;
 609         }
 610
 611         if (opt->fwd_latency && !opt->q_priority) {
 612                 evt_info("enabled queue priority for latency measurement");
 613                 opt->q_priority = 1;
 614         }
 615         if (opt->nb_pkts == 0)
 616                 opt->nb_pkts = INT64_MAX/evt_nr_active_lcores(opt->plcores);
 617
 618         return 0;
 619 }
 620
 621 void
 622 perf_opt_dump(struct evt_options *opt, uint8_t nb_queues)
 623 {
 624         evt_dump("nb_prod_lcores", "%d", evt_nr_active_lcores(opt->plcores));
 625         evt_dump_producer_lcores(opt);
 626         evt_dump("nb_worker_lcores", "%d", evt_nr_active_lcores(opt->wlcores));
 627         evt_dump_worker_lcores(opt);
 628         evt_dump_nb_stages(opt);
 629         evt_dump("nb_evdev_ports", "%d", perf_nb_event_ports(opt));
 630         evt_dump("nb_evdev_queues", "%d", nb_queues);
 631         evt_dump_queue_priority(opt);
 632         evt_dump_sched_type_list(opt);
 633         evt_dump_producer_type(opt);
 634 }
 635
 636 void
 637 perf_eventdev_destroy(struct evt_test *test, struct evt_options *opt)
 638 {
 639         int i;
 640         struct test_perf *t = evt_test_priv(test);
 641
 642         if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 643                 for (i = 0; i < opt->nb_timer_adptrs; i++)
 644                         rte_event_timer_adapter_stop(t->timer_adptr[i]);
 645         }
 646         rte_event_dev_stop(opt->dev_id);
 647         rte_event_dev_close(opt->dev_id);
 648 }
 649
 650 static inline void
 651 perf_elt_init(struct rte_mempool *mp, void *arg __rte_unused,
 652             void *obj, unsigned i __rte_unused)
 653 {
 654         memset(obj, 0, mp->elt_size);
 655 }
 656
 657 #define NB_RX_DESC                      128
 658 #define NB_TX_DESC                      512
 659 int
 660 perf_ethdev_setup(struct evt_test *test, struct evt_options *opt)
 661 {
 662         uint16_t i;
 663         int ret;
 664         struct test_perf *t = evt_test_priv(test);
 665         struct rte_eth_conf port_conf = {
 666                 .rxmode = {
 667                         .mq_mode = ETH_MQ_RX_RSS,
 668                         .max_rx_pkt_len = RTE_ETHER_MAX_LEN,
 669                         .split_hdr_size = 0,
 670                 },
 671                 .rx_adv_conf = {
 672                         .rss_conf = {
 673                                 .rss_key = NULL,
 674                                 .rss_hf = ETH_RSS_IP,
 675                         },
 676                 },
 677         };
 678
 679         if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 680                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR)
 681                 return 0;
 682
 683         if (!rte_eth_dev_count_avail()) {
 684                 evt_err("No ethernet ports found.");
 685                 return -ENODEV;
 686         }
 687
 688         RTE_ETH_FOREACH_DEV(i) {
 689                 struct rte_eth_dev_info dev_info;
 690                 struct rte_eth_conf local_port_conf = port_conf;
 691
 692                 ret = rte_eth_dev_info_get(i, &dev_info);
 693                 if (ret != 0) {
 694                         evt_err("Error during getting device (port %u) info: %s\n",
 695                                         i, strerror(-ret));
 696                         return ret;
 697                 }
 698
 699                 local_port_conf.rx_adv_conf.rss_conf.rss_hf &=
 700                         dev_info.flow_type_rss_offloads;
 701                 if (local_port_conf.rx_adv_conf.rss_conf.rss_hf !=
 702                                 port_conf.rx_adv_conf.rss_conf.rss_hf) {
 703                         evt_info("Port %u modified RSS hash function based on hardware support,"
 704                                 "requested:%#"PRIx64" configured:%#"PRIx64"\n",
 705                                 i,
 706                                 port_conf.rx_adv_conf.rss_conf.rss_hf,
 707                                 local_port_conf.rx_adv_conf.rss_conf.rss_hf);
 708                 }
 709
 710                 if (rte_eth_dev_configure(i, 1, 1, &local_port_conf) < 0) {
 711                         evt_err("Failed to configure eth port [%d]", i);
 712                         return -EINVAL;
 713                 }
 714
 715                 if (rte_eth_rx_queue_setup(i, 0, NB_RX_DESC,
 716                                 rte_socket_id(), NULL, t->pool) < 0) {
 717                         evt_err("Failed to setup eth port [%d] rx_queue: %d.",
 718                                         i, 0);
 719                         return -EINVAL;
 720                 }
 721
 722                 if (rte_eth_tx_queue_setup(i, 0, NB_TX_DESC,
 723                                         rte_socket_id(), NULL) < 0) {
 724                         evt_err("Failed to setup eth port [%d] tx_queue: %d.",
 725                                         i, 0);
 726                         return -EINVAL;
 727                 }
 728
 729                 rte_eth_promiscuous_enable(i);
 730         }
 731
 732         return 0;
 733 }
 734
 735 void perf_ethdev_destroy(struct evt_test *test, struct evt_options *opt)
 736 {
 737         uint16_t i;
 738         RTE_SET_USED(test);
 739
 740         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 741                 RTE_ETH_FOREACH_DEV(i) {
 742                         rte_event_eth_rx_adapter_stop(i);
 743                         rte_eth_dev_stop(i);
 744                 }
 745         }
 746 }
 747
 748 int
 749 perf_mempool_setup(struct evt_test *test, struct evt_options *opt)
 750 {
 751         struct test_perf *t = evt_test_priv(test);
 752
 753         if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 754                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 755                 t->pool = rte_mempool_create(test->name, /* mempool name */
 756                                 opt->pool_sz, /* number of elements*/
 757                                 sizeof(struct perf_elt), /* element size*/
 758                                 512, /* cache size*/
 759                                 0, NULL, NULL,
 760                                 perf_elt_init, /* obj constructor */
 761                                 NULL, opt->socket_id, 0); /* flags */
 762         } else {
 763                 t->pool = rte_pktmbuf_pool_create(test->name, /* mempool name */
 764                                 opt->pool_sz, /* number of elements*/
 765                                 512, /* cache size*/
 766                                 0,
 767                                 RTE_MBUF_DEFAULT_BUF_SIZE,
 768                                 opt->socket_id); /* flags */
 769
 770         }
 771
 772         if (t->pool == NULL) {
 773                 evt_err("failed to create mempool");
 774                 return -ENOMEM;
 775         }
 776
 777         return 0;
 778 }
 779
 780 void
 781 perf_mempool_destroy(struct evt_test *test, struct evt_options *opt)
 782 {
 783         RTE_SET_USED(opt);
 784         struct test_perf *t = evt_test_priv(test);
 785
 786         rte_mempool_free(t->pool);
 787 }
 788
 789 int
 790 perf_test_setup(struct evt_test *test, struct evt_options *opt)
 791 {
 792         void *test_perf;
 793
 794         test_perf = rte_zmalloc_socket(test->name, sizeof(struct test_perf),
 795                                 RTE_CACHE_LINE_SIZE, opt->socket_id);
 796         if (test_perf  == NULL) {
 797                 evt_err("failed to allocate test_perf memory");
 798                 goto nomem;
 799         }
 800         test->test_priv = test_perf;
 801
 802         struct test_perf *t = evt_test_priv(test);
 803
 804         if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 805                 t->outstand_pkts = opt->nb_timers *
 806                         evt_nr_active_lcores(opt->plcores);
 807                 t->nb_pkts = opt->nb_timers;
 808         } else {
 809                 t->outstand_pkts = opt->nb_pkts *
 810                         evt_nr_active_lcores(opt->plcores);
 811                 t->nb_pkts = opt->nb_pkts;
 812         }
 813
 814         t->nb_workers = evt_nr_active_lcores(opt->wlcores);
 815         t->done = false;
 816         t->nb_flows = opt->nb_flows;
 817         t->result = EVT_TEST_FAILED;
 818         t->opt = opt;
 819         memcpy(t->sched_type_list, opt->sched_type_list,
 820                         sizeof(opt->sched_type_list));
 821         return 0;
 822 nomem:
 823         return -ENOMEM;
 824 }
 825
 826 void
 827 perf_test_destroy(struct evt_test *test, struct evt_options *opt)
 828 {
 829         RTE_SET_USED(opt);
 830
 831         rte_free(test->test_priv);
 832 }