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 ret = rte_eth_dev_start(prod); 398 if (ret) { 399 evt_err("Ethernet dev [%d] failed to start." 400 " Using synthetic producer", prod); 401 return ret; 402 } 403 404 ret = rte_event_eth_rx_adapter_start(prod); 405 if (ret) { 406 evt_err("Rx adapter[%d] start failed", prod); 407 return ret; 408 } 409 printf("%s: Port[%d] using Rx adapter[%d] started\n", __func__, 410 prod, prod); 411 } 412 413 return ret; 414 } 415 416 static int 417 perf_event_timer_adapter_setup(struct test_perf *t) 418 { 419 int i; 420 int ret; 421 struct rte_event_timer_adapter_info adapter_info; 422 struct rte_event_timer_adapter *wl; 423 uint8_t nb_producers = evt_nr_active_lcores(t->opt->plcores); 424 uint8_t flags = RTE_EVENT_TIMER_ADAPTER_F_ADJUST_RES; 425 426 if (nb_producers == 1) 427 flags |= RTE_EVENT_TIMER_ADAPTER_F_SP_PUT; 428 429 for (i = 0; i < t->opt->nb_timer_adptrs; i++) { 430 struct rte_event_timer_adapter_conf config = { 431 .event_dev_id = t->opt->dev_id, 432 .timer_adapter_id = i, 433 .timer_tick_ns = t->opt->timer_tick_nsec, 434 .max_tmo_ns = t->opt->max_tmo_nsec, 435 .nb_timers = 2 * 1024 * 1024, 436 .flags = flags, 437 }; 438 439 wl = rte_event_timer_adapter_create(&config); 440 if (wl == NULL) { 441 evt_err("failed to create event timer ring %d", i); 442 return rte_errno; 443 } 444 445 memset(&adapter_info, 0, 446 sizeof(struct rte_event_timer_adapter_info)); 447 rte_event_timer_adapter_get_info(wl, &adapter_info); 448 t->opt->optm_timer_tick_nsec = adapter_info.min_resolution_ns; 449 450 if (!(adapter_info.caps & 451 RTE_EVENT_TIMER_ADAPTER_CAP_INTERNAL_PORT)) { 452 uint32_t service_id; 453 454 rte_event_timer_adapter_service_id_get(wl, 455 &service_id); 456 ret = evt_service_setup(service_id); 457 if (ret) { 458 evt_err("Failed to setup service core" 459 " for timer adapter\n"); 460 return ret; 461 } 462 rte_service_runstate_set(service_id, 1); 463 } 464 465 ret = rte_event_timer_adapter_start(wl); 466 if (ret) { 467 evt_err("failed to Start event timer adapter %d", i); 468 return ret; 469 } 470 t->timer_adptr[i] = wl; 471 } 472 return 0; 473 } 474 475 int 476 perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt, 477 uint8_t stride, uint8_t nb_queues, 478 const struct rte_event_port_conf *port_conf) 479 { 480 struct test_perf *t = evt_test_priv(test); 481 uint16_t port, prod; 482 int ret = -1; 483 484 /* setup one port per worker, linking to all queues */ 485 for (port = 0; port < evt_nr_active_lcores(opt->wlcores); 486 port++) { 487 struct worker_data *w = &t->worker[port]; 488 489 w->dev_id = opt->dev_id; 490 w->port_id = port; 491 w->t = t; 492 w->processed_pkts = 0; 493 w->latency = 0; 494 495 ret = rte_event_port_setup(opt->dev_id, port, port_conf); 496 if (ret) { 497 evt_err("failed to setup port %d", port); 498 return ret; 499 } 500 501 ret = rte_event_port_link(opt->dev_id, port, NULL, NULL, 0); 502 if (ret != nb_queues) { 503 evt_err("failed to link all queues to port %d", port); 504 return -EINVAL; 505 } 506 } 507 508 /* port for producers, no links */ 509 if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) { 510 for ( ; port < perf_nb_event_ports(opt); port++) { 511 struct prod_data *p = &t->prod[port]; 512 p->t = t; 513 } 514 515 ret = perf_event_rx_adapter_setup(opt, stride, *port_conf); 516 if (ret) 517 return ret; 518 } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) { 519 prod = 0; 520 for ( ; port < perf_nb_event_ports(opt); port++) { 521 struct prod_data *p = &t->prod[port]; 522 p->queue_id = prod * stride; 523 p->t = t; 524 prod++; 525 } 526 527 ret = perf_event_timer_adapter_setup(t); 528 if (ret) 529 return ret; 530 } else { 531 prod = 0; 532 for ( ; port < perf_nb_event_ports(opt); port++) { 533 struct prod_data *p = &t->prod[port]; 534 535 p->dev_id = opt->dev_id; 536 p->port_id = port; 537 p->queue_id = prod * stride; 538 p->t = t; 539 540 ret = rte_event_port_setup(opt->dev_id, port, 541 port_conf); 542 if (ret) { 543 evt_err("failed to setup port %d", port); 544 return ret; 545 } 546 prod++; 547 } 548 } 549 550 return ret; 551 } 552 553 int 554 perf_opt_check(struct evt_options *opt, uint64_t nb_queues) 555 { 556 unsigned int lcores; 557 558 /* N producer + N worker + 1 master when producer cores are used 559 * Else N worker + 1 master when Rx adapter is used 560 */ 561 lcores = opt->prod_type == EVT_PROD_TYPE_SYNT ? 3 : 2; 562 563 if (rte_lcore_count() < lcores) { 564 evt_err("test need minimum %d lcores", lcores); 565 return -1; 566 } 567 568 /* Validate worker lcores */ 569 if (evt_lcores_has_overlap(opt->wlcores, rte_get_master_lcore())) { 570 evt_err("worker lcores overlaps with master lcore"); 571 return -1; 572 } 573 if (evt_lcores_has_overlap_multi(opt->wlcores, opt->plcores)) { 574 evt_err("worker lcores overlaps producer lcores"); 575 return -1; 576 } 577 if (evt_has_disabled_lcore(opt->wlcores)) { 578 evt_err("one or more workers lcores are not enabled"); 579 return -1; 580 } 581 if (!evt_has_active_lcore(opt->wlcores)) { 582 evt_err("minimum one worker is required"); 583 return -1; 584 } 585 586 if (opt->prod_type == EVT_PROD_TYPE_SYNT) { 587 /* Validate producer lcores */ 588 if (evt_lcores_has_overlap(opt->plcores, 589 rte_get_master_lcore())) { 590 evt_err("producer lcores overlaps with master lcore"); 591 return -1; 592 } 593 if (evt_has_disabled_lcore(opt->plcores)) { 594 evt_err("one or more producer lcores are not enabled"); 595 return -1; 596 } 597 if (!evt_has_active_lcore(opt->plcores)) { 598 evt_err("minimum one producer is required"); 599 return -1; 600 } 601 } 602 603 if (evt_has_invalid_stage(opt)) 604 return -1; 605 606 if (evt_has_invalid_sched_type(opt)) 607 return -1; 608 609 if (nb_queues > EVT_MAX_QUEUES) { 610 evt_err("number of queues exceeds %d", EVT_MAX_QUEUES); 611 return -1; 612 } 613 if (perf_nb_event_ports(opt) > EVT_MAX_PORTS) { 614 evt_err("number of ports exceeds %d", EVT_MAX_PORTS); 615 return -1; 616 } 617 618 /* Fixups */ 619 if ((opt->nb_stages == 1 && 620 opt->prod_type != EVT_PROD_TYPE_EVENT_TIMER_ADPTR) && 621 opt->fwd_latency) { 622 evt_info("fwd_latency is valid when nb_stages > 1, disabling"); 623 opt->fwd_latency = 0; 624 } 625 626 if (opt->fwd_latency && !opt->q_priority) { 627 evt_info("enabled queue priority for latency measurement"); 628 opt->q_priority = 1; 629 } 630 if (opt->nb_pkts == 0) 631 opt->nb_pkts = INT64_MAX/evt_nr_active_lcores(opt->plcores); 632 633 return 0; 634 } 635 636 void 637 perf_opt_dump(struct evt_options *opt, uint8_t nb_queues) 638 { 639 evt_dump("nb_prod_lcores", "%d", evt_nr_active_lcores(opt->plcores)); 640 evt_dump_producer_lcores(opt); 641 evt_dump("nb_worker_lcores", "%d", evt_nr_active_lcores(opt->wlcores)); 642 evt_dump_worker_lcores(opt); 643 evt_dump_nb_stages(opt); 644 evt_dump("nb_evdev_ports", "%d", perf_nb_event_ports(opt)); 645 evt_dump("nb_evdev_queues", "%d", nb_queues); 646 evt_dump_queue_priority(opt); 647 evt_dump_sched_type_list(opt); 648 evt_dump_producer_type(opt); 649 } 650 651 void 652 perf_eventdev_destroy(struct evt_test *test, struct evt_options *opt) 653 { 654 int i; 655 struct test_perf *t = evt_test_priv(test); 656 657 if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) { 658 for (i = 0; i < opt->nb_timer_adptrs; i++) 659 rte_event_timer_adapter_stop(t->timer_adptr[i]); 660 } 661 rte_event_dev_stop(opt->dev_id); 662 rte_event_dev_close(opt->dev_id); 663 } 664 665 static inline void 666 perf_elt_init(struct rte_mempool *mp, void *arg __rte_unused, 667 void *obj, unsigned i __rte_unused) 668 { 669 memset(obj, 0, mp->elt_size); 670 } 671 672 #define NB_RX_DESC 128 673 #define NB_TX_DESC 512 674 int 675 perf_ethdev_setup(struct evt_test *test, struct evt_options *opt) 676 { 677 uint16_t i; 678 struct test_perf *t = evt_test_priv(test); 679 struct rte_eth_conf port_conf = { 680 .rxmode = { 681 .mq_mode = ETH_MQ_RX_RSS, 682 .max_rx_pkt_len = ETHER_MAX_LEN, 683 .split_hdr_size = 0, 684 }, 685 .rx_adv_conf = { 686 .rss_conf = { 687 .rss_key = NULL, 688 .rss_hf = ETH_RSS_IP, 689 }, 690 }, 691 }; 692 693 if (opt->prod_type == EVT_PROD_TYPE_SYNT || 694 opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) 695 return 0; 696 697 if (!rte_eth_dev_count_avail()) { 698 evt_err("No ethernet ports found."); 699 return -ENODEV; 700 } 701 702 RTE_ETH_FOREACH_DEV(i) { 703 struct rte_eth_dev_info dev_info; 704 struct rte_eth_conf local_port_conf = port_conf; 705 706 rte_eth_dev_info_get(i, &dev_info); 707 708 local_port_conf.rx_adv_conf.rss_conf.rss_hf &= 709 dev_info.flow_type_rss_offloads; 710 if (local_port_conf.rx_adv_conf.rss_conf.rss_hf != 711 port_conf.rx_adv_conf.rss_conf.rss_hf) { 712 evt_info("Port %u modified RSS hash function based on hardware support," 713 "requested:%#"PRIx64" configured:%#"PRIx64"\n", 714 i, 715 port_conf.rx_adv_conf.rss_conf.rss_hf, 716 local_port_conf.rx_adv_conf.rss_conf.rss_hf); 717 } 718 719 if (rte_eth_dev_configure(i, 1, 1, &local_port_conf) < 0) { 720 evt_err("Failed to configure eth port [%d]", i); 721 return -EINVAL; 722 } 723 724 if (rte_eth_rx_queue_setup(i, 0, NB_RX_DESC, 725 rte_socket_id(), NULL, t->pool) < 0) { 726 evt_err("Failed to setup eth port [%d] rx_queue: %d.", 727 i, 0); 728 return -EINVAL; 729 } 730 731 if (rte_eth_tx_queue_setup(i, 0, NB_TX_DESC, 732 rte_socket_id(), NULL) < 0) { 733 evt_err("Failed to setup eth port [%d] tx_queue: %d.", 734 i, 0); 735 return -EINVAL; 736 } 737 738 rte_eth_promiscuous_enable(i); 739 } 740 741 return 0; 742 } 743 744 void perf_ethdev_destroy(struct evt_test *test, struct evt_options *opt) 745 { 746 uint16_t i; 747 RTE_SET_USED(test); 748 749 if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) { 750 RTE_ETH_FOREACH_DEV(i) { 751 rte_event_eth_rx_adapter_stop(i); 752 rte_eth_dev_stop(i); 753 } 754 } 755 } 756 757 int 758 perf_mempool_setup(struct evt_test *test, struct evt_options *opt) 759 { 760 struct test_perf *t = evt_test_priv(test); 761 762 if (opt->prod_type == EVT_PROD_TYPE_SYNT || 763 opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) { 764 t->pool = rte_mempool_create(test->name, /* mempool name */ 765 opt->pool_sz, /* number of elements*/ 766 sizeof(struct perf_elt), /* element size*/ 767 512, /* cache size*/ 768 0, NULL, NULL, 769 perf_elt_init, /* obj constructor */ 770 NULL, opt->socket_id, 0); /* flags */ 771 } else { 772 t->pool = rte_pktmbuf_pool_create(test->name, /* mempool name */ 773 opt->pool_sz, /* number of elements*/ 774 512, /* cache size*/ 775 0, 776 RTE_MBUF_DEFAULT_BUF_SIZE, 777 opt->socket_id); /* flags */ 778 779 } 780 781 if (t->pool == NULL) { 782 evt_err("failed to create mempool"); 783 return -ENOMEM; 784 } 785 786 return 0; 787 } 788 789 void 790 perf_mempool_destroy(struct evt_test *test, struct evt_options *opt) 791 { 792 RTE_SET_USED(opt); 793 struct test_perf *t = evt_test_priv(test); 794 795 rte_mempool_free(t->pool); 796 } 797 798 int 799 perf_test_setup(struct evt_test *test, struct evt_options *opt) 800 { 801 void *test_perf; 802 803 test_perf = rte_zmalloc_socket(test->name, sizeof(struct test_perf), 804 RTE_CACHE_LINE_SIZE, opt->socket_id); 805 if (test_perf == NULL) { 806 evt_err("failed to allocate test_perf memory"); 807 goto nomem; 808 } 809 test->test_priv = test_perf; 810 811 struct test_perf *t = evt_test_priv(test); 812 813 if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) { 814 t->outstand_pkts = opt->nb_timers * 815 evt_nr_active_lcores(opt->plcores); 816 t->nb_pkts = opt->nb_timers; 817 } else { 818 t->outstand_pkts = opt->nb_pkts * 819 evt_nr_active_lcores(opt->plcores); 820 t->nb_pkts = opt->nb_pkts; 821 } 822 823 t->nb_workers = evt_nr_active_lcores(opt->wlcores); 824 t->done = false; 825 t->nb_flows = opt->nb_flows; 826 t->result = EVT_TEST_FAILED; 827 t->opt = opt; 828 memcpy(t->sched_type_list, opt->sched_type_list, 829 sizeof(opt->sched_type_list)); 830 return 0; 831 nomem: 832 return -ENOMEM; 833 } 834 835 void 836 perf_test_destroy(struct evt_test *test, struct evt_options *opt) 837 { 838 RTE_SET_USED(opt); 839 840 rte_free(test->test_priv); 841 } 842