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 int 76 perf_producer_wrapper(void *arg) 77 { 78 struct prod_data *p = arg; 79 struct test_perf *t = p->t; 80 /* Launch the producer function only in case of synthetic producer. */ 81 if (t->opt->prod_type == EVT_PROD_TYPE_SYNT) 82 return perf_producer(arg); 83 return 0; 84 } 85 86 static inline uint64_t 87 processed_pkts(struct test_perf *t) 88 { 89 uint8_t i; 90 uint64_t total = 0; 91 92 rte_smp_rmb(); 93 for (i = 0; i < t->nb_workers; i++) 94 total += t->worker[i].processed_pkts; 95 96 return total; 97 } 98 99 static inline uint64_t 100 total_latency(struct test_perf *t) 101 { 102 uint8_t i; 103 uint64_t total = 0; 104 105 rte_smp_rmb(); 106 for (i = 0; i < t->nb_workers; i++) 107 total += t->worker[i].latency; 108 109 return total; 110 } 111 112 113 int 114 perf_launch_lcores(struct evt_test *test, struct evt_options *opt, 115 int (*worker)(void *)) 116 { 117 int ret, lcore_id; 118 struct test_perf *t = evt_test_priv(test); 119 120 int port_idx = 0; 121 /* launch workers */ 122 RTE_LCORE_FOREACH_SLAVE(lcore_id) { 123 if (!(opt->wlcores[lcore_id])) 124 continue; 125 126 ret = rte_eal_remote_launch(worker, 127 &t->worker[port_idx], lcore_id); 128 if (ret) { 129 evt_err("failed to launch worker %d", lcore_id); 130 return ret; 131 } 132 port_idx++; 133 } 134 135 /* launch producers */ 136 RTE_LCORE_FOREACH_SLAVE(lcore_id) { 137 if (!(opt->plcores[lcore_id])) 138 continue; 139 140 ret = rte_eal_remote_launch(perf_producer_wrapper, 141 &t->prod[port_idx], lcore_id); 142 if (ret) { 143 evt_err("failed to launch perf_producer %d", lcore_id); 144 return ret; 145 } 146 port_idx++; 147 } 148 149 const uint64_t total_pkts = opt->nb_pkts * 150 evt_nr_active_lcores(opt->plcores); 151 152 uint64_t dead_lock_cycles = rte_get_timer_cycles(); 153 int64_t dead_lock_remaining = total_pkts; 154 const uint64_t dead_lock_sample = rte_get_timer_hz() * 5; 155 156 uint64_t perf_cycles = rte_get_timer_cycles(); 157 int64_t perf_remaining = total_pkts; 158 const uint64_t perf_sample = rte_get_timer_hz(); 159 160 static float total_mpps; 161 static uint64_t samples; 162 163 const uint64_t freq_mhz = rte_get_timer_hz() / 1000000; 164 int64_t remaining = t->outstand_pkts - processed_pkts(t); 165 166 while (t->done == false) { 167 const uint64_t new_cycles = rte_get_timer_cycles(); 168 169 if ((new_cycles - perf_cycles) > perf_sample) { 170 const uint64_t latency = total_latency(t); 171 const uint64_t pkts = processed_pkts(t); 172 173 remaining = t->outstand_pkts - pkts; 174 float mpps = (float)(perf_remaining-remaining)/1000000; 175 176 perf_remaining = remaining; 177 perf_cycles = new_cycles; 178 total_mpps += mpps; 179 ++samples; 180 if (opt->fwd_latency && pkts > 0) { 181 printf(CLGRN"\r%.3f mpps avg %.3f mpps [avg fwd latency %.3f us] "CLNRM, 182 mpps, total_mpps/samples, 183 (float)(latency/pkts)/freq_mhz); 184 } else { 185 printf(CLGRN"\r%.3f mpps avg %.3f mpps"CLNRM, 186 mpps, total_mpps/samples); 187 } 188 fflush(stdout); 189 190 if (remaining <= 0) { 191 t->result = EVT_TEST_SUCCESS; 192 if (opt->prod_type == EVT_PROD_TYPE_SYNT) { 193 t->done = true; 194 rte_smp_wmb(); 195 break; 196 } 197 } 198 } 199 200 if (new_cycles - dead_lock_cycles > dead_lock_sample && 201 opt->prod_type == EVT_PROD_TYPE_SYNT) { 202 remaining = t->outstand_pkts - processed_pkts(t); 203 if (dead_lock_remaining == remaining) { 204 rte_event_dev_dump(opt->dev_id, stdout); 205 evt_err("No schedules for seconds, deadlock"); 206 t->done = true; 207 rte_smp_wmb(); 208 break; 209 } 210 dead_lock_remaining = remaining; 211 dead_lock_cycles = new_cycles; 212 } 213 } 214 printf("\n"); 215 return 0; 216 } 217 218 static int 219 perf_event_rx_adapter_setup(struct evt_options *opt, uint8_t stride, 220 struct rte_event_port_conf prod_conf) 221 { 222 int ret = 0; 223 uint16_t prod; 224 struct rte_event_eth_rx_adapter_queue_conf queue_conf; 225 226 memset(&queue_conf, 0, 227 sizeof(struct rte_event_eth_rx_adapter_queue_conf)); 228 queue_conf.ev.sched_type = opt->sched_type_list[0]; 229 for (prod = 0; prod < rte_eth_dev_count(); prod++) { 230 uint32_t cap; 231 232 ret = rte_event_eth_rx_adapter_caps_get(opt->dev_id, 233 prod, &cap); 234 if (ret) { 235 evt_err("failed to get event rx adapter[%d]" 236 " capabilities", 237 opt->dev_id); 238 return ret; 239 } 240 queue_conf.ev.queue_id = prod * stride; 241 ret = rte_event_eth_rx_adapter_create(prod, opt->dev_id, 242 &prod_conf); 243 if (ret) { 244 evt_err("failed to create rx adapter[%d]", prod); 245 return ret; 246 } 247 ret = rte_event_eth_rx_adapter_queue_add(prod, prod, -1, 248 &queue_conf); 249 if (ret) { 250 evt_err("failed to add rx queues to adapter[%d]", prod); 251 return ret; 252 } 253 254 if (!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT)) { 255 uint32_t service_id; 256 257 rte_event_eth_rx_adapter_service_id_get(prod, 258 &service_id); 259 ret = evt_service_setup(service_id); 260 if (ret) { 261 evt_err("Failed to setup service core" 262 " for Rx adapter\n"); 263 return ret; 264 } 265 } 266 267 ret = rte_eth_dev_start(prod); 268 if (ret) { 269 evt_err("Ethernet dev [%d] failed to start." 270 " Using synthetic producer", prod); 271 return ret; 272 } 273 274 ret = rte_event_eth_rx_adapter_start(prod); 275 if (ret) { 276 evt_err("Rx adapter[%d] start failed", prod); 277 return ret; 278 } 279 printf("%s: Port[%d] using Rx adapter[%d] started\n", __func__, 280 prod, prod); 281 } 282 283 return ret; 284 } 285 286 int 287 perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt, 288 uint8_t stride, uint8_t nb_queues, 289 const struct rte_event_port_conf *port_conf) 290 { 291 struct test_perf *t = evt_test_priv(test); 292 uint16_t port, prod; 293 int ret = -1; 294 295 /* setup one port per worker, linking to all queues */ 296 for (port = 0; port < evt_nr_active_lcores(opt->wlcores); 297 port++) { 298 struct worker_data *w = &t->worker[port]; 299 300 w->dev_id = opt->dev_id; 301 w->port_id = port; 302 w->t = t; 303 w->processed_pkts = 0; 304 w->latency = 0; 305 306 ret = rte_event_port_setup(opt->dev_id, port, port_conf); 307 if (ret) { 308 evt_err("failed to setup port %d", port); 309 return ret; 310 } 311 312 ret = rte_event_port_link(opt->dev_id, port, NULL, NULL, 0); 313 if (ret != nb_queues) { 314 evt_err("failed to link all queues to port %d", port); 315 return -EINVAL; 316 } 317 } 318 319 /* port for producers, no links */ 320 if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) { 321 for ( ; port < perf_nb_event_ports(opt); port++) { 322 struct prod_data *p = &t->prod[port]; 323 p->t = t; 324 } 325 326 ret = perf_event_rx_adapter_setup(opt, stride, *port_conf); 327 if (ret) 328 return ret; 329 } else { 330 prod = 0; 331 for ( ; port < perf_nb_event_ports(opt); port++) { 332 struct prod_data *p = &t->prod[port]; 333 334 p->dev_id = opt->dev_id; 335 p->port_id = port; 336 p->queue_id = prod * stride; 337 p->t = t; 338 339 ret = rte_event_port_setup(opt->dev_id, port, 340 port_conf); 341 if (ret) { 342 evt_err("failed to setup port %d", port); 343 return ret; 344 } 345 prod++; 346 } 347 } 348 349 return ret; 350 } 351 352 int 353 perf_opt_check(struct evt_options *opt, uint64_t nb_queues) 354 { 355 unsigned int lcores; 356 357 /* N producer + N worker + 1 master when producer cores are used 358 * Else N worker + 1 master when Rx adapter is used 359 */ 360 lcores = opt->prod_type == EVT_PROD_TYPE_SYNT ? 3 : 2; 361 362 if (rte_lcore_count() < lcores) { 363 evt_err("test need minimum %d lcores", lcores); 364 return -1; 365 } 366 367 /* Validate worker lcores */ 368 if (evt_lcores_has_overlap(opt->wlcores, rte_get_master_lcore())) { 369 evt_err("worker lcores overlaps with master lcore"); 370 return -1; 371 } 372 if (evt_lcores_has_overlap_multi(opt->wlcores, opt->plcores)) { 373 evt_err("worker lcores overlaps producer lcores"); 374 return -1; 375 } 376 if (evt_has_disabled_lcore(opt->wlcores)) { 377 evt_err("one or more workers lcores are not enabled"); 378 return -1; 379 } 380 if (!evt_has_active_lcore(opt->wlcores)) { 381 evt_err("minimum one worker is required"); 382 return -1; 383 } 384 385 if (opt->prod_type == EVT_PROD_TYPE_SYNT) { 386 /* Validate producer lcores */ 387 if (evt_lcores_has_overlap(opt->plcores, 388 rte_get_master_lcore())) { 389 evt_err("producer lcores overlaps with master lcore"); 390 return -1; 391 } 392 if (evt_has_disabled_lcore(opt->plcores)) { 393 evt_err("one or more producer lcores are not enabled"); 394 return -1; 395 } 396 if (!evt_has_active_lcore(opt->plcores)) { 397 evt_err("minimum one producer is required"); 398 return -1; 399 } 400 } 401 402 if (evt_has_invalid_stage(opt)) 403 return -1; 404 405 if (evt_has_invalid_sched_type(opt)) 406 return -1; 407 408 if (nb_queues > EVT_MAX_QUEUES) { 409 evt_err("number of queues exceeds %d", EVT_MAX_QUEUES); 410 return -1; 411 } 412 if (perf_nb_event_ports(opt) > EVT_MAX_PORTS) { 413 evt_err("number of ports exceeds %d", EVT_MAX_PORTS); 414 return -1; 415 } 416 417 /* Fixups */ 418 if (opt->nb_stages == 1 && opt->fwd_latency) { 419 evt_info("fwd_latency is valid when nb_stages > 1, disabling"); 420 opt->fwd_latency = 0; 421 } 422 if (opt->fwd_latency && !opt->q_priority) { 423 evt_info("enabled queue priority for latency measurement"); 424 opt->q_priority = 1; 425 } 426 if (opt->nb_pkts == 0) 427 opt->nb_pkts = INT64_MAX/evt_nr_active_lcores(opt->plcores); 428 429 return 0; 430 } 431 432 void 433 perf_opt_dump(struct evt_options *opt, uint8_t nb_queues) 434 { 435 evt_dump("nb_prod_lcores", "%d", evt_nr_active_lcores(opt->plcores)); 436 evt_dump_producer_lcores(opt); 437 evt_dump("nb_worker_lcores", "%d", evt_nr_active_lcores(opt->wlcores)); 438 evt_dump_worker_lcores(opt); 439 evt_dump_nb_stages(opt); 440 evt_dump("nb_evdev_ports", "%d", perf_nb_event_ports(opt)); 441 evt_dump("nb_evdev_queues", "%d", nb_queues); 442 evt_dump_queue_priority(opt); 443 evt_dump_sched_type_list(opt); 444 evt_dump_producer_type(opt); 445 } 446 447 void 448 perf_eventdev_destroy(struct evt_test *test, struct evt_options *opt) 449 { 450 RTE_SET_USED(test); 451 452 rte_event_dev_stop(opt->dev_id); 453 rte_event_dev_close(opt->dev_id); 454 } 455 456 static inline void 457 perf_elt_init(struct rte_mempool *mp, void *arg __rte_unused, 458 void *obj, unsigned i __rte_unused) 459 { 460 memset(obj, 0, mp->elt_size); 461 } 462 463 #define NB_RX_DESC 128 464 #define NB_TX_DESC 512 465 int 466 perf_ethdev_setup(struct evt_test *test, struct evt_options *opt) 467 { 468 int i; 469 struct test_perf *t = evt_test_priv(test); 470 struct rte_eth_conf port_conf = { 471 .rxmode = { 472 .mq_mode = ETH_MQ_RX_RSS, 473 .max_rx_pkt_len = ETHER_MAX_LEN, 474 .split_hdr_size = 0, 475 .header_split = 0, 476 .hw_ip_checksum = 0, 477 .hw_vlan_filter = 0, 478 .hw_vlan_strip = 0, 479 .hw_vlan_extend = 0, 480 .jumbo_frame = 0, 481 .hw_strip_crc = 1, 482 }, 483 .rx_adv_conf = { 484 .rss_conf = { 485 .rss_key = NULL, 486 .rss_hf = ETH_RSS_IP, 487 }, 488 }, 489 }; 490 491 if (opt->prod_type == EVT_PROD_TYPE_SYNT) 492 return 0; 493 494 if (!rte_eth_dev_count()) { 495 evt_err("No ethernet ports found."); 496 return -ENODEV; 497 } 498 499 for (i = 0; i < rte_eth_dev_count(); i++) { 500 501 if (rte_eth_dev_configure(i, 1, 1, 502 &port_conf) 503 < 0) { 504 evt_err("Failed to configure eth port [%d]", i); 505 return -EINVAL; 506 } 507 508 if (rte_eth_rx_queue_setup(i, 0, NB_RX_DESC, 509 rte_socket_id(), NULL, t->pool) < 0) { 510 evt_err("Failed to setup eth port [%d] rx_queue: %d.", 511 i, 0); 512 return -EINVAL; 513 } 514 515 if (rte_eth_tx_queue_setup(i, 0, NB_TX_DESC, 516 rte_socket_id(), NULL) < 0) { 517 evt_err("Failed to setup eth port [%d] tx_queue: %d.", 518 i, 0); 519 return -EINVAL; 520 } 521 522 rte_eth_promiscuous_enable(i); 523 } 524 525 return 0; 526 } 527 528 void perf_ethdev_destroy(struct evt_test *test, struct evt_options *opt) 529 { 530 int i; 531 RTE_SET_USED(test); 532 533 if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) { 534 for (i = 0; i < rte_eth_dev_count(); i++) { 535 rte_event_eth_rx_adapter_stop(i); 536 rte_eth_dev_stop(i); 537 rte_eth_dev_close(i); 538 } 539 } 540 } 541 542 int 543 perf_mempool_setup(struct evt_test *test, struct evt_options *opt) 544 { 545 struct test_perf *t = evt_test_priv(test); 546 547 if (opt->prod_type == EVT_PROD_TYPE_SYNT) { 548 t->pool = rte_mempool_create(test->name, /* mempool name */ 549 opt->pool_sz, /* number of elements*/ 550 sizeof(struct perf_elt), /* element size*/ 551 512, /* cache size*/ 552 0, NULL, NULL, 553 perf_elt_init, /* obj constructor */ 554 NULL, opt->socket_id, 0); /* flags */ 555 } else { 556 t->pool = rte_pktmbuf_pool_create(test->name, /* mempool name */ 557 opt->pool_sz, /* number of elements*/ 558 512, /* cache size*/ 559 0, 560 RTE_MBUF_DEFAULT_BUF_SIZE, 561 opt->socket_id); /* flags */ 562 563 } 564 565 if (t->pool == NULL) { 566 evt_err("failed to create mempool"); 567 return -ENOMEM; 568 } 569 570 return 0; 571 } 572 573 void 574 perf_mempool_destroy(struct evt_test *test, struct evt_options *opt) 575 { 576 RTE_SET_USED(opt); 577 struct test_perf *t = evt_test_priv(test); 578 579 rte_mempool_free(t->pool); 580 } 581 582 int 583 perf_test_setup(struct evt_test *test, struct evt_options *opt) 584 { 585 void *test_perf; 586 587 test_perf = rte_zmalloc_socket(test->name, sizeof(struct test_perf), 588 RTE_CACHE_LINE_SIZE, opt->socket_id); 589 if (test_perf == NULL) { 590 evt_err("failed to allocate test_perf memory"); 591 goto nomem; 592 } 593 test->test_priv = test_perf; 594 595 struct test_perf *t = evt_test_priv(test); 596 597 t->outstand_pkts = opt->nb_pkts * evt_nr_active_lcores(opt->plcores); 598 t->nb_workers = evt_nr_active_lcores(opt->wlcores); 599 t->done = false; 600 t->nb_pkts = opt->nb_pkts; 601 t->nb_flows = opt->nb_flows; 602 t->result = EVT_TEST_FAILED; 603 t->opt = opt; 604 memcpy(t->sched_type_list, opt->sched_type_list, 605 sizeof(opt->sched_type_list)); 606 return 0; 607 nomem: 608 return -ENOMEM; 609 } 610 611 void 612 perf_test_destroy(struct evt_test *test, struct evt_options *opt) 613 { 614 RTE_SET_USED(opt); 615 616 rte_free(test->test_priv); 617 } 618