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(),
 137                         count ? (float)(arm_latency / count) /
 138                         (rte_get_timer_hz() / 1000000) : 0);
 139         return 0;
 140 }
 141
 142 static inline int
 143 perf_event_timer_producer_burst(void *arg)
 144 {
 145         int i;
 146         struct prod_data *p  = arg;
 147         struct test_perf *t = p->t;
 148         struct evt_options *opt = t->opt;
 149         uint32_t flow_counter = 0;
 150         uint64_t count = 0;
 151         uint64_t arm_latency = 0;
 152         const uint8_t nb_timer_adptrs = opt->nb_timer_adptrs;
 153         const uint32_t nb_flows = t->nb_flows;
 154         const uint64_t nb_timers = opt->nb_timers;
 155         struct rte_mempool *pool = t->pool;
 156         struct perf_elt *m[BURST_SIZE + 1] = {NULL};
 157         struct rte_event_timer_adapter **adptr = t->timer_adptr;
 158         struct rte_event_timer tim;
 159         uint64_t timeout_ticks = opt->expiry_nsec / opt->timer_tick_nsec;
 160
 161         memset(&tim, 0, sizeof(struct rte_event_timer));
 162         timeout_ticks = opt->optm_timer_tick_nsec ?
 163                         (timeout_ticks * opt->timer_tick_nsec)
 164                         / opt->optm_timer_tick_nsec : timeout_ticks;
 165         timeout_ticks += timeout_ticks ? 0 : 1;
 166         tim.ev.event_type =  RTE_EVENT_TYPE_TIMER;
 167         tim.ev.op = RTE_EVENT_OP_NEW;
 168         tim.ev.sched_type = t->opt->sched_type_list[0];
 169         tim.ev.queue_id = p->queue_id;
 170         tim.ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
 171         tim.state = RTE_EVENT_TIMER_NOT_ARMED;
 172         tim.timeout_ticks = timeout_ticks;
 173
 174         if (opt->verbose_level > 1)
 175                 printf("%s(): lcore %d\n", __func__, rte_lcore_id());
 176
 177         while (count < nb_timers && t->done == false) {
 178                 if (rte_mempool_get_bulk(pool, (void **)m, BURST_SIZE) < 0)
 179                         continue;
 180                 for (i = 0; i < BURST_SIZE; i++) {
 181                         rte_prefetch0(m[i + 1]);
 182                         m[i]->tim = tim;
 183                         m[i]->tim.ev.flow_id = flow_counter++ % nb_flows;
 184                         m[i]->tim.ev.event_ptr = m[i];
 185                         m[i]->timestamp = rte_get_timer_cycles();
 186                 }
 187                 rte_event_timer_arm_tmo_tick_burst(
 188                                 adptr[flow_counter % nb_timer_adptrs],
 189                                 (struct rte_event_timer **)m,
 190                                 tim.timeout_ticks,
 191                                 BURST_SIZE);
 192                 arm_latency += rte_get_timer_cycles() - m[i - 1]->timestamp;
 193                 count += BURST_SIZE;
 194         }
 195         fflush(stdout);
 196         rte_delay_ms(1000);
 197         printf("%s(): lcore %d Average event timer arm latency = %.3f us\n",
 198                         __func__, rte_lcore_id(),
 199                         count ? (float)(arm_latency / count) /
 200                         (rte_get_timer_hz() / 1000000) : 0);
 201         return 0;
 202 }
 203
 204 static int
 205 perf_producer_wrapper(void *arg)
 206 {
 207         struct prod_data *p  = arg;
 208         struct test_perf *t = p->t;
 209         /* Launch the producer function only in case of synthetic producer. */
 210         if (t->opt->prod_type == EVT_PROD_TYPE_SYNT)
 211                 return perf_producer(arg);
 212         else if (t->opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR &&
 213                         !t->opt->timdev_use_burst)
 214                 return perf_event_timer_producer(arg);
 215         else if (t->opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR &&
 216                         t->opt->timdev_use_burst)
 217                 return perf_event_timer_producer_burst(arg);
 218         return 0;
 219 }
 220
 221 static inline uint64_t
 222 processed_pkts(struct test_perf *t)
 223 {
 224         uint8_t i;
 225         uint64_t total = 0;
 226
 227         for (i = 0; i < t->nb_workers; i++)
 228                 total += t->worker[i].processed_pkts;
 229
 230         return total;
 231 }
 232
 233 static inline uint64_t
 234 total_latency(struct test_perf *t)
 235 {
 236         uint8_t i;
 237         uint64_t total = 0;
 238
 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_WORKER(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_WORKER(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                                         break;
 329                                 }
 330                         }
 331                 }
 332
 333                 if (new_cycles - dead_lock_cycles > dead_lock_sample &&
 334                     (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 335                      opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR)) {
 336                         remaining = t->outstand_pkts - processed_pkts(t);
 337                         if (dead_lock_remaining == remaining) {
 338                                 rte_event_dev_dump(opt->dev_id, stdout);
 339                                 evt_err("No schedules for seconds, deadlock");
 340                                 t->done = true;
 341                                 break;
 342                         }
 343                         dead_lock_remaining = remaining;
 344                         dead_lock_cycles = new_cycles;
 345                 }
 346         }
 347         printf("\n");
 348         return 0;
 349 }
 350
 351 static int
 352 perf_event_rx_adapter_setup(struct evt_options *opt, uint8_t stride,
 353                 struct rte_event_port_conf prod_conf)
 354 {
 355         int ret = 0;
 356         uint16_t prod;
 357         struct rte_event_eth_rx_adapter_queue_conf queue_conf;
 358
 359         memset(&queue_conf, 0,
 360                         sizeof(struct rte_event_eth_rx_adapter_queue_conf));
 361         queue_conf.ev.sched_type = opt->sched_type_list[0];
 362         RTE_ETH_FOREACH_DEV(prod) {
 363                 uint32_t cap;
 364
 365                 ret = rte_event_eth_rx_adapter_caps_get(opt->dev_id,
 366                                 prod, &cap);
 367                 if (ret) {
 368                         evt_err("failed to get event rx adapter[%d]"
 369                                         " capabilities",
 370                                         opt->dev_id);
 371                         return ret;
 372                 }
 373                 queue_conf.ev.queue_id = prod * stride;
 374                 ret = rte_event_eth_rx_adapter_create(prod, opt->dev_id,
 375                                 &prod_conf);
 376                 if (ret) {
 377                         evt_err("failed to create rx adapter[%d]", prod);
 378                         return ret;
 379                 }
 380                 ret = rte_event_eth_rx_adapter_queue_add(prod, prod, -1,
 381                                 &queue_conf);
 382                 if (ret) {
 383                         evt_err("failed to add rx queues to adapter[%d]", prod);
 384                         return ret;
 385                 }
 386
 387                 if (!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT)) {
 388                         uint32_t service_id;
 389
 390                         rte_event_eth_rx_adapter_service_id_get(prod,
 391                                         &service_id);
 392                         ret = evt_service_setup(service_id);
 393                         if (ret) {
 394                                 evt_err("Failed to setup service core"
 395                                                 " for Rx adapter\n");
 396                                 return ret;
 397                         }
 398                 }
 399         }
 400
 401         return ret;
 402 }
 403
 404 static int
 405 perf_event_timer_adapter_setup(struct test_perf *t)
 406 {
 407         int i;
 408         int ret;
 409         struct rte_event_timer_adapter_info adapter_info;
 410         struct rte_event_timer_adapter *wl;
 411         uint8_t nb_producers = evt_nr_active_lcores(t->opt->plcores);
 412         uint8_t flags = RTE_EVENT_TIMER_ADAPTER_F_ADJUST_RES;
 413
 414         if (nb_producers == 1)
 415                 flags |= RTE_EVENT_TIMER_ADAPTER_F_SP_PUT;
 416
 417         for (i = 0; i < t->opt->nb_timer_adptrs; i++) {
 418                 struct rte_event_timer_adapter_conf config = {
 419                         .event_dev_id = t->opt->dev_id,
 420                         .timer_adapter_id = i,
 421                         .timer_tick_ns = t->opt->timer_tick_nsec,
 422                         .max_tmo_ns = t->opt->max_tmo_nsec,
 423                         .nb_timers = t->opt->pool_sz,
 424                         .flags = flags,
 425                 };
 426
 427                 wl = rte_event_timer_adapter_create(&config);
 428                 if (wl == NULL) {
 429                         evt_err("failed to create event timer ring %d", i);
 430                         return rte_errno;
 431                 }
 432
 433                 memset(&adapter_info, 0,
 434                                 sizeof(struct rte_event_timer_adapter_info));
 435                 rte_event_timer_adapter_get_info(wl, &adapter_info);
 436                 t->opt->optm_timer_tick_nsec = adapter_info.min_resolution_ns;
 437
 438                 if (!(adapter_info.caps &
 439                                 RTE_EVENT_TIMER_ADAPTER_CAP_INTERNAL_PORT)) {
 440                         uint32_t service_id = -1U;
 441
 442                         rte_event_timer_adapter_service_id_get(wl,
 443                                         &service_id);
 444                         ret = evt_service_setup(service_id);
 445                         if (ret) {
 446                                 evt_err("Failed to setup service core"
 447                                                 " for timer adapter\n");
 448                                 return ret;
 449                         }
 450                         rte_service_runstate_set(service_id, 1);
 451                 }
 452                 t->timer_adptr[i] = wl;
 453         }
 454         return 0;
 455 }
 456
 457 int
 458 perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt,
 459                                 uint8_t stride, uint8_t nb_queues,
 460                                 const struct rte_event_port_conf *port_conf)
 461 {
 462         struct test_perf *t = evt_test_priv(test);
 463         uint16_t port, prod;
 464         int ret = -1;
 465
 466         /* setup one port per worker, linking to all queues */
 467         for (port = 0; port < evt_nr_active_lcores(opt->wlcores);
 468                                 port++) {
 469                 struct worker_data *w = &t->worker[port];
 470
 471                 w->dev_id = opt->dev_id;
 472                 w->port_id = port;
 473                 w->t = t;
 474                 w->processed_pkts = 0;
 475                 w->latency = 0;
 476
 477                 ret = rte_event_port_setup(opt->dev_id, port, port_conf);
 478                 if (ret) {
 479                         evt_err("failed to setup port %d", port);
 480                         return ret;
 481                 }
 482
 483                 ret = rte_event_port_link(opt->dev_id, port, NULL, NULL, 0);
 484                 if (ret != nb_queues) {
 485                         evt_err("failed to link all queues to port %d", port);
 486                         return -EINVAL;
 487                 }
 488         }
 489
 490         /* port for producers, no links */
 491         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 492                 for ( ; port < perf_nb_event_ports(opt); port++) {
 493                         struct prod_data *p = &t->prod[port];
 494                         p->t = t;
 495                 }
 496
 497                 ret = perf_event_rx_adapter_setup(opt, stride, *port_conf);
 498                 if (ret)
 499                         return ret;
 500         } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 501                 prod = 0;
 502                 for ( ; port < perf_nb_event_ports(opt); port++) {
 503                         struct prod_data *p = &t->prod[port];
 504                         p->queue_id = prod * stride;
 505                         p->t = t;
 506                         prod++;
 507                 }
 508
 509                 ret = perf_event_timer_adapter_setup(t);
 510                 if (ret)
 511                         return ret;
 512         } else {
 513                 prod = 0;
 514                 for ( ; port < perf_nb_event_ports(opt); port++) {
 515                         struct prod_data *p = &t->prod[port];
 516
 517                         p->dev_id = opt->dev_id;
 518                         p->port_id = port;
 519                         p->queue_id = prod * stride;
 520                         p->t = t;
 521
 522                         ret = rte_event_port_setup(opt->dev_id, port,
 523                                         port_conf);
 524                         if (ret) {
 525                                 evt_err("failed to setup port %d", port);
 526                                 return ret;
 527                         }
 528                         prod++;
 529                 }
 530         }
 531
 532         return ret;
 533 }
 534
 535 int
 536 perf_opt_check(struct evt_options *opt, uint64_t nb_queues)
 537 {
 538         unsigned int lcores;
 539
 540         /* N producer + N worker + main when producer cores are used
 541          * Else N worker + main when Rx adapter is used
 542          */
 543         lcores = opt->prod_type == EVT_PROD_TYPE_SYNT ? 3 : 2;
 544
 545         if (rte_lcore_count() < lcores) {
 546                 evt_err("test need minimum %d lcores", lcores);
 547                 return -1;
 548         }
 549
 550         /* Validate worker lcores */
 551         if (evt_lcores_has_overlap(opt->wlcores, rte_get_main_lcore())) {
 552                 evt_err("worker lcores overlaps with main lcore");
 553                 return -1;
 554         }
 555         if (evt_lcores_has_overlap_multi(opt->wlcores, opt->plcores)) {
 556                 evt_err("worker lcores overlaps producer lcores");
 557                 return -1;
 558         }
 559         if (evt_has_disabled_lcore(opt->wlcores)) {
 560                 evt_err("one or more workers lcores are not enabled");
 561                 return -1;
 562         }
 563         if (!evt_has_active_lcore(opt->wlcores)) {
 564                 evt_err("minimum one worker is required");
 565                 return -1;
 566         }
 567
 568         if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 569                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 570                 /* Validate producer lcores */
 571                 if (evt_lcores_has_overlap(opt->plcores,
 572                                         rte_get_main_lcore())) {
 573                         evt_err("producer lcores overlaps with main lcore");
 574                         return -1;
 575                 }
 576                 if (evt_has_disabled_lcore(opt->plcores)) {
 577                         evt_err("one or more producer lcores are not enabled");
 578                         return -1;
 579                 }
 580                 if (!evt_has_active_lcore(opt->plcores)) {
 581                         evt_err("minimum one producer is required");
 582                         return -1;
 583                 }
 584         }
 585
 586         if (evt_has_invalid_stage(opt))
 587                 return -1;
 588
 589         if (evt_has_invalid_sched_type(opt))
 590                 return -1;
 591
 592         if (nb_queues > EVT_MAX_QUEUES) {
 593                 evt_err("number of queues exceeds %d", EVT_MAX_QUEUES);
 594                 return -1;
 595         }
 596         if (perf_nb_event_ports(opt) > EVT_MAX_PORTS) {
 597                 evt_err("number of ports exceeds %d", EVT_MAX_PORTS);
 598                 return -1;
 599         }
 600
 601         /* Fixups */
 602         if ((opt->nb_stages == 1 &&
 603                         opt->prod_type != EVT_PROD_TYPE_EVENT_TIMER_ADPTR) &&
 604                         opt->fwd_latency) {
 605                 evt_info("fwd_latency is valid when nb_stages > 1, disabling");
 606                 opt->fwd_latency = 0;
 607         }
 608
 609         if (opt->fwd_latency && !opt->q_priority) {
 610                 evt_info("enabled queue priority for latency measurement");
 611                 opt->q_priority = 1;
 612         }
 613         if (opt->nb_pkts == 0)
 614                 opt->nb_pkts = INT64_MAX/evt_nr_active_lcores(opt->plcores);
 615
 616         return 0;
 617 }
 618
 619 void
 620 perf_opt_dump(struct evt_options *opt, uint8_t nb_queues)
 621 {
 622         evt_dump("nb_prod_lcores", "%d", evt_nr_active_lcores(opt->plcores));
 623         evt_dump_producer_lcores(opt);
 624         evt_dump("nb_worker_lcores", "%d", evt_nr_active_lcores(opt->wlcores));
 625         evt_dump_worker_lcores(opt);
 626         evt_dump_nb_stages(opt);
 627         evt_dump("nb_evdev_ports", "%d", perf_nb_event_ports(opt));
 628         evt_dump("nb_evdev_queues", "%d", nb_queues);
 629         evt_dump_queue_priority(opt);
 630         evt_dump_sched_type_list(opt);
 631         evt_dump_producer_type(opt);
 632 }
 633
 634 void
 635 perf_eventdev_destroy(struct evt_test *test, struct evt_options *opt)
 636 {
 637         int i;
 638         struct test_perf *t = evt_test_priv(test);
 639
 640         if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 641                 for (i = 0; i < opt->nb_timer_adptrs; i++)
 642                         rte_event_timer_adapter_stop(t->timer_adptr[i]);
 643         }
 644         rte_event_dev_stop(opt->dev_id);
 645         rte_event_dev_close(opt->dev_id);
 646 }
 647
 648 static inline void
 649 perf_elt_init(struct rte_mempool *mp, void *arg __rte_unused,
 650             void *obj, unsigned i __rte_unused)
 651 {
 652         memset(obj, 0, mp->elt_size);
 653 }
 654
 655 #define NB_RX_DESC                      128
 656 #define NB_TX_DESC                      512
 657 int
 658 perf_ethdev_setup(struct evt_test *test, struct evt_options *opt)
 659 {
 660         uint16_t i;
 661         int ret;
 662         struct test_perf *t = evt_test_priv(test);
 663         struct rte_eth_conf port_conf = {
 664                 .rxmode = {
 665                         .mq_mode = ETH_MQ_RX_RSS,
 666                         .max_rx_pkt_len = RTE_ETHER_MAX_LEN,
 667                         .split_hdr_size = 0,
 668                 },
 669                 .rx_adv_conf = {
 670                         .rss_conf = {
 671                                 .rss_key = NULL,
 672                                 .rss_hf = ETH_RSS_IP,
 673                         },
 674                 },
 675         };
 676
 677         if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 678                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR)
 679                 return 0;
 680
 681         if (!rte_eth_dev_count_avail()) {
 682                 evt_err("No ethernet ports found.");
 683                 return -ENODEV;
 684         }
 685
 686         RTE_ETH_FOREACH_DEV(i) {
 687                 struct rte_eth_dev_info dev_info;
 688                 struct rte_eth_conf local_port_conf = port_conf;
 689
 690                 ret = rte_eth_dev_info_get(i, &dev_info);
 691                 if (ret != 0) {
 692                         evt_err("Error during getting device (port %u) info: %s\n",
 693                                         i, strerror(-ret));
 694                         return ret;
 695                 }
 696
 697                 local_port_conf.rx_adv_conf.rss_conf.rss_hf &=
 698                         dev_info.flow_type_rss_offloads;
 699                 if (local_port_conf.rx_adv_conf.rss_conf.rss_hf !=
 700                                 port_conf.rx_adv_conf.rss_conf.rss_hf) {
 701                         evt_info("Port %u modified RSS hash function based on hardware support,"
 702                                 "requested:%#"PRIx64" configured:%#"PRIx64"\n",
 703                                 i,
 704                                 port_conf.rx_adv_conf.rss_conf.rss_hf,
 705                                 local_port_conf.rx_adv_conf.rss_conf.rss_hf);
 706                 }
 707
 708                 if (rte_eth_dev_configure(i, 1, 1, &local_port_conf) < 0) {
 709                         evt_err("Failed to configure eth port [%d]", i);
 710                         return -EINVAL;
 711                 }
 712
 713                 if (rte_eth_rx_queue_setup(i, 0, NB_RX_DESC,
 714                                 rte_socket_id(), NULL, t->pool) < 0) {
 715                         evt_err("Failed to setup eth port [%d] rx_queue: %d.",
 716                                         i, 0);
 717                         return -EINVAL;
 718                 }
 719
 720                 if (rte_eth_tx_queue_setup(i, 0, NB_TX_DESC,
 721                                         rte_socket_id(), NULL) < 0) {
 722                         evt_err("Failed to setup eth port [%d] tx_queue: %d.",
 723                                         i, 0);
 724                         return -EINVAL;
 725                 }
 726
 727                 ret = rte_eth_promiscuous_enable(i);
 728                 if (ret != 0) {
 729                         evt_err("Failed to enable promiscuous mode for eth port [%d]: %s",
 730                                 i, rte_strerror(-ret));
 731                         return ret;
 732                 }
 733         }
 734
 735         return 0;
 736 }
 737
 738 void perf_ethdev_destroy(struct evt_test *test, struct evt_options *opt)
 739 {
 740         uint16_t i;
 741         RTE_SET_USED(test);
 742
 743         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 744                 RTE_ETH_FOREACH_DEV(i) {
 745                         rte_event_eth_rx_adapter_stop(i);
 746                         rte_eth_dev_stop(i);
 747                 }
 748         }
 749 }
 750
 751 int
 752 perf_mempool_setup(struct evt_test *test, struct evt_options *opt)
 753 {
 754         struct test_perf *t = evt_test_priv(test);
 755
 756         if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
 757                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 758                 t->pool = rte_mempool_create(test->name, /* mempool name */
 759                                 opt->pool_sz, /* number of elements*/
 760                                 sizeof(struct perf_elt), /* element size*/
 761                                 512, /* cache size*/
 762                                 0, NULL, NULL,
 763                                 perf_elt_init, /* obj constructor */
 764                                 NULL, opt->socket_id, 0); /* flags */
 765         } else {
 766                 t->pool = rte_pktmbuf_pool_create(test->name, /* mempool name */
 767                                 opt->pool_sz, /* number of elements*/
 768                                 512, /* cache size*/
 769                                 0,
 770                                 RTE_MBUF_DEFAULT_BUF_SIZE,
 771                                 opt->socket_id); /* flags */
 772
 773         }
 774
 775         if (t->pool == NULL) {
 776                 evt_err("failed to create mempool");
 777                 return -ENOMEM;
 778         }
 779
 780         return 0;
 781 }
 782
 783 void
 784 perf_mempool_destroy(struct evt_test *test, struct evt_options *opt)
 785 {
 786         RTE_SET_USED(opt);
 787         struct test_perf *t = evt_test_priv(test);
 788
 789         rte_mempool_free(t->pool);
 790 }
 791
 792 int
 793 perf_test_setup(struct evt_test *test, struct evt_options *opt)
 794 {
 795         void *test_perf;
 796
 797         test_perf = rte_zmalloc_socket(test->name, sizeof(struct test_perf),
 798                                 RTE_CACHE_LINE_SIZE, opt->socket_id);
 799         if (test_perf  == NULL) {
 800                 evt_err("failed to allocate test_perf memory");
 801                 goto nomem;
 802         }
 803         test->test_priv = test_perf;
 804
 805         struct test_perf *t = evt_test_priv(test);
 806
 807         if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 808                 t->outstand_pkts = opt->nb_timers *
 809                         evt_nr_active_lcores(opt->plcores);
 810                 t->nb_pkts = opt->nb_timers;
 811         } else {
 812                 t->outstand_pkts = opt->nb_pkts *
 813                         evt_nr_active_lcores(opt->plcores);
 814                 t->nb_pkts = opt->nb_pkts;
 815         }
 816
 817         t->nb_workers = evt_nr_active_lcores(opt->wlcores);
 818         t->done = false;
 819         t->nb_flows = opt->nb_flows;
 820         t->result = EVT_TEST_FAILED;
 821         t->opt = opt;
 822         memcpy(t->sched_type_list, opt->sched_type_list,
 823                         sizeof(opt->sched_type_list));
 824         return 0;
 825 nomem:
 826         return -ENOMEM;
 827 }
 828
 829 void
 830 perf_test_destroy(struct evt_test *test, struct evt_options *opt)
 831 {
 832         RTE_SET_USED(opt);
 833
 834         rte_free(test->test_priv);
 835 }