1 /* 2 * SPDX-License-Identifier: BSD-3-Clause 3 * Copyright 2016 Intel Corporation. 4 * Copyright 2017 Cavium, Inc. 5 */ 6 7 #include "pipeline_common.h" 8 9 static __rte_always_inline int 10 worker_generic(void *arg) 11 { 12 struct rte_event ev; 13 14 struct worker_data *data = (struct worker_data *)arg; 15 uint8_t dev_id = data->dev_id; 16 uint8_t port_id = data->port_id; 17 size_t sent = 0, received = 0; 18 unsigned int lcore_id = rte_lcore_id(); 19 20 while (!fdata->done) { 21 22 if (fdata->cap.scheduler) 23 fdata->cap.scheduler(lcore_id); 24 25 if (!fdata->worker_core[lcore_id]) { 26 rte_pause(); 27 continue; 28 } 29 30 const uint16_t nb_rx = rte_event_dequeue_burst(dev_id, port_id, 31 &ev, 1, 0); 32 33 if (nb_rx == 0) { 34 rte_pause(); 35 continue; 36 } 37 received++; 38 39 /* The first worker stage does classification */ 40 if (ev.queue_id == cdata.qid[0]) 41 ev.flow_id = ev.mbuf->hash.rss 42 % cdata.num_fids; 43 44 ev.queue_id = cdata.next_qid[ev.queue_id]; 45 ev.op = RTE_EVENT_OP_FORWARD; 46 ev.sched_type = cdata.queue_type; 47 48 work(); 49 50 while (rte_event_enqueue_burst(dev_id, port_id, &ev, 1) != 1) 51 rte_pause(); 52 sent++; 53 } 54 55 if (!cdata.quiet) 56 printf(" worker %u thread done. RX=%zu TX=%zu\n", 57 rte_lcore_id(), received, sent); 58 59 return 0; 60 } 61 62 static int 63 worker_generic_burst(void *arg) 64 { 65 struct rte_event events[BATCH_SIZE]; 66 67 struct worker_data *data = (struct worker_data *)arg; 68 uint8_t dev_id = data->dev_id; 69 uint8_t port_id = data->port_id; 70 size_t sent = 0, received = 0; 71 unsigned int lcore_id = rte_lcore_id(); 72 73 while (!fdata->done) { 74 uint16_t i; 75 76 if (fdata->cap.scheduler) 77 fdata->cap.scheduler(lcore_id); 78 79 if (!fdata->worker_core[lcore_id]) { 80 rte_pause(); 81 continue; 82 } 83 84 const uint16_t nb_rx = rte_event_dequeue_burst(dev_id, port_id, 85 events, RTE_DIM(events), 0); 86 87 if (nb_rx == 0) { 88 rte_pause(); 89 continue; 90 } 91 received += nb_rx; 92 93 for (i = 0; i < nb_rx; i++) { 94 95 /* The first worker stage does classification */ 96 if (events[i].queue_id == cdata.qid[0]) 97 events[i].flow_id = events[i].mbuf->hash.rss 98 % cdata.num_fids; 99 100 events[i].queue_id = cdata.next_qid[events[i].queue_id]; 101 events[i].op = RTE_EVENT_OP_FORWARD; 102 events[i].sched_type = cdata.queue_type; 103 104 work(); 105 } 106 uint16_t nb_tx = rte_event_enqueue_burst(dev_id, port_id, 107 events, nb_rx); 108 while (nb_tx < nb_rx && !fdata->done) 109 nb_tx += rte_event_enqueue_burst(dev_id, port_id, 110 events + nb_tx, 111 nb_rx - nb_tx); 112 sent += nb_tx; 113 } 114 115 if (!cdata.quiet) 116 printf(" worker %u thread done. RX=%zu TX=%zu\n", 117 rte_lcore_id(), received, sent); 118 119 return 0; 120 } 121 122 static __rte_always_inline int 123 consumer(void) 124 { 125 const uint64_t freq_khz = rte_get_timer_hz() / 1000; 126 struct rte_event packet; 127 128 static uint64_t received; 129 static uint64_t last_pkts; 130 static uint64_t last_time; 131 static uint64_t start_time; 132 int i; 133 uint8_t dev_id = cons_data.dev_id; 134 uint8_t port_id = cons_data.port_id; 135 136 do { 137 uint16_t n = rte_event_dequeue_burst(dev_id, port_id, 138 &packet, 1, 0); 139 140 if (n == 0) { 141 RTE_ETH_FOREACH_DEV(i) 142 rte_eth_tx_buffer_flush(i, 0, fdata->tx_buf[i]); 143 return 0; 144 } 145 if (start_time == 0) 146 last_time = start_time = rte_get_timer_cycles(); 147 148 received++; 149 uint8_t outport = packet.mbuf->port; 150 151 exchange_mac(packet.mbuf); 152 rte_eth_tx_buffer(outport, 0, fdata->tx_buf[outport], 153 packet.mbuf); 154 155 if (cons_data.release) 156 rte_event_enqueue_burst(dev_id, port_id, 157 &packet, n); 158 159 /* Print out mpps every 1<22 packets */ 160 if (!cdata.quiet && received >= last_pkts + (1<<22)) { 161 const uint64_t now = rte_get_timer_cycles(); 162 const uint64_t total_ms = (now - start_time) / freq_khz; 163 const uint64_t delta_ms = (now - last_time) / freq_khz; 164 uint64_t delta_pkts = received - last_pkts; 165 166 printf("# %s RX=%"PRIu64", time %"PRIu64 "ms, " 167 "avg %.3f mpps [current %.3f mpps]\n", 168 __func__, 169 received, 170 total_ms, 171 received / (total_ms * 1000.0), 172 delta_pkts / (delta_ms * 1000.0)); 173 last_pkts = received; 174 last_time = now; 175 } 176 177 cdata.num_packets--; 178 if (cdata.num_packets <= 0) 179 fdata->done = 1; 180 /* Be stuck in this loop if single. */ 181 } while (!fdata->done && fdata->tx_single); 182 183 return 0; 184 } 185 186 static __rte_always_inline int 187 consumer_burst(void) 188 { 189 const uint64_t freq_khz = rte_get_timer_hz() / 1000; 190 struct rte_event packets[BATCH_SIZE]; 191 192 static uint64_t received; 193 static uint64_t last_pkts; 194 static uint64_t last_time; 195 static uint64_t start_time; 196 unsigned int i, j; 197 uint8_t dev_id = cons_data.dev_id; 198 uint8_t port_id = cons_data.port_id; 199 200 do { 201 uint16_t n = rte_event_dequeue_burst(dev_id, port_id, 202 packets, RTE_DIM(packets), 0); 203 204 if (n == 0) { 205 RTE_ETH_FOREACH_DEV(j) 206 rte_eth_tx_buffer_flush(j, 0, fdata->tx_buf[j]); 207 return 0; 208 } 209 if (start_time == 0) 210 last_time = start_time = rte_get_timer_cycles(); 211 212 received += n; 213 for (i = 0; i < n; i++) { 214 uint8_t outport = packets[i].mbuf->port; 215 216 exchange_mac(packets[i].mbuf); 217 rte_eth_tx_buffer(outport, 0, fdata->tx_buf[outport], 218 packets[i].mbuf); 219 220 packets[i].op = RTE_EVENT_OP_RELEASE; 221 } 222 223 if (cons_data.release) { 224 uint16_t nb_tx; 225 226 nb_tx = rte_event_enqueue_burst(dev_id, port_id, 227 packets, n); 228 while (nb_tx < n) 229 nb_tx += rte_event_enqueue_burst(dev_id, 230 port_id, packets + nb_tx, 231 n - nb_tx); 232 } 233 234 /* Print out mpps every 1<22 packets */ 235 if (!cdata.quiet && received >= last_pkts + (1<<22)) { 236 const uint64_t now = rte_get_timer_cycles(); 237 const uint64_t total_ms = (now - start_time) / freq_khz; 238 const uint64_t delta_ms = (now - last_time) / freq_khz; 239 uint64_t delta_pkts = received - last_pkts; 240 241 printf("# consumer RX=%"PRIu64", time %"PRIu64 "ms, " 242 "avg %.3f mpps [current %.3f mpps]\n", 243 received, 244 total_ms, 245 received / (total_ms * 1000.0), 246 delta_pkts / (delta_ms * 1000.0)); 247 last_pkts = received; 248 last_time = now; 249 } 250 251 cdata.num_packets -= n; 252 if (cdata.num_packets <= 0) 253 fdata->done = 1; 254 /* Be stuck in this loop if single. */ 255 } while (!fdata->done && fdata->tx_single); 256 257 return 0; 258 } 259 260 static int 261 setup_eventdev_generic(struct cons_data *cons_data, 262 struct worker_data *worker_data) 263 { 264 const uint8_t dev_id = 0; 265 /* +1 stages is for a SINGLE_LINK TX stage */ 266 const uint8_t nb_queues = cdata.num_stages + 1; 267 /* + 1 is one port for consumer */ 268 const uint8_t nb_ports = cdata.num_workers + 1; 269 struct rte_event_dev_config config = { 270 .nb_event_queues = nb_queues, 271 .nb_event_ports = nb_ports, 272 .nb_events_limit = 4096, 273 .nb_event_queue_flows = 1024, 274 .nb_event_port_dequeue_depth = 128, 275 .nb_event_port_enqueue_depth = 128, 276 }; 277 struct rte_event_port_conf wkr_p_conf = { 278 .dequeue_depth = cdata.worker_cq_depth, 279 .enqueue_depth = 64, 280 .new_event_threshold = 4096, 281 }; 282 struct rte_event_queue_conf wkr_q_conf = { 283 .schedule_type = cdata.queue_type, 284 .priority = RTE_EVENT_DEV_PRIORITY_NORMAL, 285 .nb_atomic_flows = 1024, 286 .nb_atomic_order_sequences = 1024, 287 }; 288 struct rte_event_port_conf tx_p_conf = { 289 .dequeue_depth = 128, 290 .enqueue_depth = 128, 291 .new_event_threshold = 4096, 292 }; 293 struct rte_event_queue_conf tx_q_conf = { 294 .priority = RTE_EVENT_DEV_PRIORITY_HIGHEST, 295 .event_queue_cfg = RTE_EVENT_QUEUE_CFG_SINGLE_LINK, 296 }; 297 298 struct port_link worker_queues[MAX_NUM_STAGES]; 299 uint8_t disable_implicit_release; 300 struct port_link tx_queue; 301 unsigned int i; 302 303 int ret, ndev = rte_event_dev_count(); 304 if (ndev < 1) { 305 printf("%d: No Eventdev Devices Found\n", __LINE__); 306 return -1; 307 } 308 309 struct rte_event_dev_info dev_info; 310 ret = rte_event_dev_info_get(dev_id, &dev_info); 311 printf("\tEventdev %d: %s\n", dev_id, dev_info.driver_name); 312 313 disable_implicit_release = (dev_info.event_dev_cap & 314 RTE_EVENT_DEV_CAP_IMPLICIT_RELEASE_DISABLE); 315 316 wkr_p_conf.disable_implicit_release = disable_implicit_release; 317 tx_p_conf.disable_implicit_release = disable_implicit_release; 318 319 if (dev_info.max_event_port_dequeue_depth < 320 config.nb_event_port_dequeue_depth) 321 config.nb_event_port_dequeue_depth = 322 dev_info.max_event_port_dequeue_depth; 323 if (dev_info.max_event_port_enqueue_depth < 324 config.nb_event_port_enqueue_depth) 325 config.nb_event_port_enqueue_depth = 326 dev_info.max_event_port_enqueue_depth; 327 328 ret = rte_event_dev_configure(dev_id, &config); 329 if (ret < 0) { 330 printf("%d: Error configuring device\n", __LINE__); 331 return -1; 332 } 333 334 /* Q creation - one load balanced per pipeline stage*/ 335 printf(" Stages:\n"); 336 for (i = 0; i < cdata.num_stages; i++) { 337 if (rte_event_queue_setup(dev_id, i, &wkr_q_conf) < 0) { 338 printf("%d: error creating qid %d\n", __LINE__, i); 339 return -1; 340 } 341 cdata.qid[i] = i; 342 cdata.next_qid[i] = i+1; 343 worker_queues[i].queue_id = i; 344 if (cdata.enable_queue_priorities) { 345 /* calculate priority stepping for each stage, leaving 346 * headroom of 1 for the SINGLE_LINK TX below 347 */ 348 const uint32_t prio_delta = 349 (RTE_EVENT_DEV_PRIORITY_LOWEST-1) / nb_queues; 350 351 /* higher priority for queues closer to tx */ 352 wkr_q_conf.priority = 353 RTE_EVENT_DEV_PRIORITY_LOWEST - prio_delta * i; 354 } 355 356 const char *type_str = "Atomic"; 357 switch (wkr_q_conf.schedule_type) { 358 case RTE_SCHED_TYPE_ORDERED: 359 type_str = "Ordered"; 360 break; 361 case RTE_SCHED_TYPE_PARALLEL: 362 type_str = "Parallel"; 363 break; 364 } 365 printf("\tStage %d, Type %s\tPriority = %d\n", i, type_str, 366 wkr_q_conf.priority); 367 } 368 printf("\n"); 369 370 /* final queue for sending to TX core */ 371 if (rte_event_queue_setup(dev_id, i, &tx_q_conf) < 0) { 372 printf("%d: error creating qid %d\n", __LINE__, i); 373 return -1; 374 } 375 tx_queue.queue_id = i; 376 tx_queue.priority = RTE_EVENT_DEV_PRIORITY_HIGHEST; 377 378 if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth) 379 wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth; 380 if (wkr_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth) 381 wkr_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth; 382 383 /* set up one port per worker, linking to all stage queues */ 384 for (i = 0; i < cdata.num_workers; i++) { 385 struct worker_data *w = &worker_data[i]; 386 w->dev_id = dev_id; 387 if (rte_event_port_setup(dev_id, i, &wkr_p_conf) < 0) { 388 printf("Error setting up port %d\n", i); 389 return -1; 390 } 391 392 uint32_t s; 393 for (s = 0; s < cdata.num_stages; s++) { 394 if (rte_event_port_link(dev_id, i, 395 &worker_queues[s].queue_id, 396 &worker_queues[s].priority, 397 1) != 1) { 398 printf("%d: error creating link for port %d\n", 399 __LINE__, i); 400 return -1; 401 } 402 } 403 w->port_id = i; 404 } 405 406 if (tx_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth) 407 tx_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth; 408 if (tx_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth) 409 tx_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth; 410 411 /* port for consumer, linked to TX queue */ 412 if (rte_event_port_setup(dev_id, i, &tx_p_conf) < 0) { 413 printf("Error setting up port %d\n", i); 414 return -1; 415 } 416 if (rte_event_port_link(dev_id, i, &tx_queue.queue_id, 417 &tx_queue.priority, 1) != 1) { 418 printf("%d: error creating link for port %d\n", 419 __LINE__, i); 420 return -1; 421 } 422 *cons_data = (struct cons_data){.dev_id = dev_id, 423 .port_id = i, 424 .release = disable_implicit_release }; 425 426 ret = rte_event_dev_service_id_get(dev_id, 427 &fdata->evdev_service_id); 428 if (ret != -ESRCH && ret != 0) { 429 printf("Error getting the service ID for sw eventdev\n"); 430 return -1; 431 } 432 rte_service_runstate_set(fdata->evdev_service_id, 1); 433 rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0); 434 if (rte_event_dev_start(dev_id) < 0) { 435 printf("Error starting eventdev\n"); 436 return -1; 437 } 438 439 return dev_id; 440 } 441 442 static void 443 init_rx_adapter(uint16_t nb_ports) 444 { 445 int i; 446 int ret; 447 uint8_t evdev_id = 0; 448 struct rte_event_dev_info dev_info; 449 450 ret = rte_event_dev_info_get(evdev_id, &dev_info); 451 452 struct rte_event_port_conf rx_p_conf = { 453 .dequeue_depth = 8, 454 .enqueue_depth = 8, 455 .new_event_threshold = 1200, 456 }; 457 458 if (rx_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth) 459 rx_p_conf.dequeue_depth = dev_info.max_event_port_dequeue_depth; 460 if (rx_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth) 461 rx_p_conf.enqueue_depth = dev_info.max_event_port_enqueue_depth; 462 463 /* Create one adapter for all the ethernet ports. */ 464 ret = rte_event_eth_rx_adapter_create(cdata.rx_adapter_id, evdev_id, 465 &rx_p_conf); 466 if (ret) 467 rte_exit(EXIT_FAILURE, "failed to create rx adapter[%d]", 468 cdata.rx_adapter_id); 469 470 struct rte_event_eth_rx_adapter_queue_conf queue_conf; 471 memset(&queue_conf, 0, sizeof(queue_conf)); 472 queue_conf.ev.sched_type = cdata.queue_type; 473 queue_conf.ev.queue_id = cdata.qid[0]; 474 475 for (i = 0; i < nb_ports; i++) { 476 uint32_t cap; 477 478 ret = rte_event_eth_rx_adapter_caps_get(evdev_id, i, &cap); 479 if (ret) 480 rte_exit(EXIT_FAILURE, 481 "failed to get event rx adapter " 482 "capabilities"); 483 484 ret = rte_event_eth_rx_adapter_queue_add(cdata.rx_adapter_id, i, 485 -1, &queue_conf); 486 if (ret) 487 rte_exit(EXIT_FAILURE, 488 "Failed to add queues to Rx adapter"); 489 } 490 491 ret = rte_event_eth_rx_adapter_service_id_get(cdata.rx_adapter_id, 492 &fdata->rxadptr_service_id); 493 if (ret != -ESRCH && ret != 0) { 494 rte_exit(EXIT_FAILURE, 495 "Error getting the service ID for sw eventdev\n"); 496 } 497 rte_service_runstate_set(fdata->rxadptr_service_id, 1); 498 rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id, 0); 499 500 ret = rte_event_eth_rx_adapter_start(cdata.rx_adapter_id); 501 if (ret) 502 rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed", 503 cdata.rx_adapter_id); 504 } 505 506 static void 507 generic_opt_check(void) 508 { 509 int i; 510 int ret; 511 uint32_t cap = 0; 512 uint8_t rx_needed = 0; 513 struct rte_event_dev_info eventdev_info; 514 515 memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info)); 516 rte_event_dev_info_get(0, &eventdev_info); 517 518 if (cdata.all_type_queues && !(eventdev_info.event_dev_cap & 519 RTE_EVENT_DEV_CAP_QUEUE_ALL_TYPES)) 520 rte_exit(EXIT_FAILURE, 521 "Event dev doesn't support all type queues\n"); 522 523 RTE_ETH_FOREACH_DEV(i) { 524 ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap); 525 if (ret) 526 rte_exit(EXIT_FAILURE, 527 "failed to get event rx adapter capabilities"); 528 rx_needed |= 529 !(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT); 530 } 531 532 if (cdata.worker_lcore_mask == 0 || 533 (rx_needed && cdata.rx_lcore_mask == 0) || 534 cdata.tx_lcore_mask == 0 || (cdata.sched_lcore_mask == 0 535 && !(eventdev_info.event_dev_cap & 536 RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED))) { 537 printf("Core part of pipeline was not assigned any cores. " 538 "This will stall the pipeline, please check core masks " 539 "(use -h for details on setting core masks):\n" 540 "\trx: %"PRIu64"\n\ttx: %"PRIu64"\n\tsched: %"PRIu64 541 "\n\tworkers: %"PRIu64"\n", 542 cdata.rx_lcore_mask, cdata.tx_lcore_mask, 543 cdata.sched_lcore_mask, 544 cdata.worker_lcore_mask); 545 rte_exit(-1, "Fix core masks\n"); 546 } 547 548 if (eventdev_info.event_dev_cap & RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED) 549 memset(fdata->sched_core, 0, 550 sizeof(unsigned int) * MAX_NUM_CORE); 551 } 552 553 void 554 set_worker_generic_setup_data(struct setup_data *caps, bool burst) 555 { 556 if (burst) { 557 caps->consumer = consumer_burst; 558 caps->worker = worker_generic_burst; 559 } else { 560 caps->consumer = consumer; 561 caps->worker = worker_generic; 562 } 563 564 caps->adptr_setup = init_rx_adapter; 565 caps->scheduler = schedule_devices; 566 caps->evdev_setup = setup_eventdev_generic; 567 caps->check_opt = generic_opt_check; 568 } 569