examples/eventdev_pipeline_sw_pmd/pipeline_worker_generic.c

   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(ev.mbuf);
  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(events[i].mbuf);
 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                         for (i = 0; i < rte_eth_dev_count(); 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                 rte_eth_tx_buffer(outport, 0, fdata->tx_buf[outport],
 152                                 packet.mbuf);
 153
 154                 if (cons_data.release)
 155                         rte_event_enqueue_burst(dev_id, port_id,
 156                                                                 &packet, n);
 157
 158                 /* Print out mpps every 1<22 packets */
 159                 if (!cdata.quiet && received >= last_pkts + (1<<22)) {
 160                         const uint64_t now = rte_get_timer_cycles();
 161                         const uint64_t total_ms = (now - start_time) / freq_khz;
 162                         const uint64_t delta_ms = (now - last_time) / freq_khz;
 163                         uint64_t delta_pkts = received - last_pkts;
 164
 165                         printf("# %s RX=%"PRIu64", time %"PRIu64 "ms, "
 166                                         "avg %.3f mpps [current %.3f mpps]\n",
 167                                         __func__,
 168                                         received,
 169                                         total_ms,
 170                                         received / (total_ms * 1000.0),
 171                                         delta_pkts / (delta_ms * 1000.0));
 172                         last_pkts = received;
 173                         last_time = now;
 174                 }
 175
 176                 cdata.num_packets--;
 177                 if (cdata.num_packets <= 0)
 178                         fdata->done = 1;
 179         /* Be stuck in this loop if single. */
 180         } while (!fdata->done && fdata->tx_single);
 181
 182         return 0;
 183 }
 184
 185 static __rte_always_inline int
 186 consumer_burst(void)
 187 {
 188         const uint64_t freq_khz = rte_get_timer_hz() / 1000;
 189         struct rte_event packets[BATCH_SIZE];
 190
 191         static uint64_t received;
 192         static uint64_t last_pkts;
 193         static uint64_t last_time;
 194         static uint64_t start_time;
 195         unsigned int i, j;
 196         uint8_t dev_id = cons_data.dev_id;
 197         uint8_t port_id = cons_data.port_id;
 198         uint16_t nb_ports = rte_eth_dev_count();
 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                         for (j = 0; j < nb_ports; 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                         rte_eth_tx_buffer(outport, 0, fdata->tx_buf[outport],
 216                                         packets[i].mbuf);
 217
 218                         packets[i].op = RTE_EVENT_OP_RELEASE;
 219                 }
 220
 221                 if (cons_data.release) {
 222                         uint16_t nb_tx;
 223
 224                         nb_tx = rte_event_enqueue_burst(dev_id, port_id,
 225                                                                 packets, n);
 226                         while (nb_tx < n)
 227                                 nb_tx += rte_event_enqueue_burst(dev_id,
 228                                                 port_id, packets + nb_tx,
 229                                                 n - nb_tx);
 230                 }
 231
 232                 /* Print out mpps every 1<22 packets */
 233                 if (!cdata.quiet && received >= last_pkts + (1<<22)) {
 234                         const uint64_t now = rte_get_timer_cycles();
 235                         const uint64_t total_ms = (now - start_time) / freq_khz;
 236                         const uint64_t delta_ms = (now - last_time) / freq_khz;
 237                         uint64_t delta_pkts = received - last_pkts;
 238
 239                         printf("# consumer RX=%"PRIu64", time %"PRIu64 "ms, "
 240                                         "avg %.3f mpps [current %.3f mpps]\n",
 241                                         received,
 242                                         total_ms,
 243                                         received / (total_ms * 1000.0),
 244                                         delta_pkts / (delta_ms * 1000.0));
 245                         last_pkts = received;
 246                         last_time = now;
 247                 }
 248
 249                 cdata.num_packets -= n;
 250                 if (cdata.num_packets <= 0)
 251                         fdata->done = 1;
 252         /* Be stuck in this loop if single. */
 253         } while (!fdata->done && fdata->tx_single);
 254
 255         return 0;
 256 }
 257
 258 static int
 259 setup_eventdev_generic(struct cons_data *cons_data,
 260                 struct worker_data *worker_data)
 261 {
 262         const uint8_t dev_id = 0;
 263         /* +1 stages is for a SINGLE_LINK TX stage */
 264         const uint8_t nb_queues = cdata.num_stages + 1;
 265         /* + 1 is one port for consumer */
 266         const uint8_t nb_ports = cdata.num_workers + 1;
 267         struct rte_event_dev_config config = {
 268                         .nb_event_queues = nb_queues,
 269                         .nb_event_ports = nb_ports,
 270                         .nb_events_limit  = 4096,
 271                         .nb_event_queue_flows = 1024,
 272                         .nb_event_port_dequeue_depth = 128,
 273                         .nb_event_port_enqueue_depth = 128,
 274         };
 275         struct rte_event_port_conf wkr_p_conf = {
 276                         .dequeue_depth = cdata.worker_cq_depth,
 277                         .enqueue_depth = 64,
 278                         .new_event_threshold = 4096,
 279         };
 280         struct rte_event_queue_conf wkr_q_conf = {
 281                         .schedule_type = cdata.queue_type,
 282                         .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 283                         .nb_atomic_flows = 1024,
 284                 .nb_atomic_order_sequences = 1024,
 285         };
 286         struct rte_event_port_conf tx_p_conf = {
 287                         .dequeue_depth = 128,
 288                         .enqueue_depth = 128,
 289                         .new_event_threshold = 4096,
 290         };
 291         struct rte_event_queue_conf tx_q_conf = {
 292                         .priority = RTE_EVENT_DEV_PRIORITY_HIGHEST,
 293                         .event_queue_cfg = RTE_EVENT_QUEUE_CFG_SINGLE_LINK,
 294         };
 295
 296         struct port_link worker_queues[MAX_NUM_STAGES];
 297         uint8_t disable_implicit_release;
 298         struct port_link tx_queue;
 299         unsigned int i;
 300
 301         int ret, ndev = rte_event_dev_count();
 302         if (ndev < 1) {
 303                 printf("%d: No Eventdev Devices Found\n", __LINE__);
 304                 return -1;
 305         }
 306
 307         struct rte_event_dev_info dev_info;
 308         ret = rte_event_dev_info_get(dev_id, &dev_info);
 309         printf("\tEventdev %d: %s\n", dev_id, dev_info.driver_name);
 310
 311         disable_implicit_release = (dev_info.event_dev_cap &
 312                         RTE_EVENT_DEV_CAP_IMPLICIT_RELEASE_DISABLE);
 313
 314         wkr_p_conf.disable_implicit_release = disable_implicit_release;
 315         tx_p_conf.disable_implicit_release = disable_implicit_release;
 316
 317         if (dev_info.max_event_port_dequeue_depth <
 318                         config.nb_event_port_dequeue_depth)
 319                 config.nb_event_port_dequeue_depth =
 320                                 dev_info.max_event_port_dequeue_depth;
 321         if (dev_info.max_event_port_enqueue_depth <
 322                         config.nb_event_port_enqueue_depth)
 323                 config.nb_event_port_enqueue_depth =
 324                                 dev_info.max_event_port_enqueue_depth;
 325
 326         ret = rte_event_dev_configure(dev_id, &config);
 327         if (ret < 0) {
 328                 printf("%d: Error configuring device\n", __LINE__);
 329                 return -1;
 330         }
 331
 332         /* Q creation - one load balanced per pipeline stage*/
 333         printf("  Stages:\n");
 334         for (i = 0; i < cdata.num_stages; i++) {
 335                 if (rte_event_queue_setup(dev_id, i, &wkr_q_conf) < 0) {
 336                         printf("%d: error creating qid %d\n", __LINE__, i);
 337                         return -1;
 338                 }
 339                 cdata.qid[i] = i;
 340                 cdata.next_qid[i] = i+1;
 341                 worker_queues[i].queue_id = i;
 342                 if (cdata.enable_queue_priorities) {
 343                         /* calculate priority stepping for each stage, leaving
 344                          * headroom of 1 for the SINGLE_LINK TX below
 345                          */
 346                         const uint32_t prio_delta =
 347                                 (RTE_EVENT_DEV_PRIORITY_LOWEST-1) /  nb_queues;
 348
 349                         /* higher priority for queues closer to tx */
 350                         wkr_q_conf.priority =
 351                                 RTE_EVENT_DEV_PRIORITY_LOWEST - prio_delta * i;
 352                 }
 353
 354                 const char *type_str = "Atomic";
 355                 switch (wkr_q_conf.schedule_type) {
 356                 case RTE_SCHED_TYPE_ORDERED:
 357                         type_str = "Ordered";
 358                         break;
 359                 case RTE_SCHED_TYPE_PARALLEL:
 360                         type_str = "Parallel";
 361                         break;
 362                 }
 363                 printf("\tStage %d, Type %s\tPriority = %d\n", i, type_str,
 364                                 wkr_q_conf.priority);
 365         }
 366         printf("\n");
 367
 368         /* final queue for sending to TX core */
 369         if (rte_event_queue_setup(dev_id, i, &tx_q_conf) < 0) {
 370                 printf("%d: error creating qid %d\n", __LINE__, i);
 371                 return -1;
 372         }
 373         tx_queue.queue_id = i;
 374         tx_queue.priority = RTE_EVENT_DEV_PRIORITY_HIGHEST;
 375
 376         if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
 377                 wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
 378         if (wkr_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
 379                 wkr_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
 380
 381         /* set up one port per worker, linking to all stage queues */
 382         for (i = 0; i < cdata.num_workers; i++) {
 383                 struct worker_data *w = &worker_data[i];
 384                 w->dev_id = dev_id;
 385                 if (rte_event_port_setup(dev_id, i, &wkr_p_conf) < 0) {
 386                         printf("Error setting up port %d\n", i);
 387                         return -1;
 388                 }
 389
 390                 uint32_t s;
 391                 for (s = 0; s < cdata.num_stages; s++) {
 392                         if (rte_event_port_link(dev_id, i,
 393                                                 &worker_queues[s].queue_id,
 394                                                 &worker_queues[s].priority,
 395                                                 1) != 1) {
 396                                 printf("%d: error creating link for port %d\n",
 397                                                 __LINE__, i);
 398                                 return -1;
 399                         }
 400                 }
 401                 w->port_id = i;
 402         }
 403
 404         if (tx_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
 405                 tx_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
 406         if (tx_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
 407                 tx_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
 408
 409         /* port for consumer, linked to TX queue */
 410         if (rte_event_port_setup(dev_id, i, &tx_p_conf) < 0) {
 411                 printf("Error setting up port %d\n", i);
 412                 return -1;
 413         }
 414         if (rte_event_port_link(dev_id, i, &tx_queue.queue_id,
 415                                 &tx_queue.priority, 1) != 1) {
 416                 printf("%d: error creating link for port %d\n",
 417                                 __LINE__, i);
 418                 return -1;
 419         }
 420         *cons_data = (struct cons_data){.dev_id = dev_id,
 421                                         .port_id = i,
 422                                         .release = disable_implicit_release };
 423
 424         ret = rte_event_dev_service_id_get(dev_id,
 425                                 &fdata->evdev_service_id);
 426         if (ret != -ESRCH && ret != 0) {
 427                 printf("Error getting the service ID for sw eventdev\n");
 428                 return -1;
 429         }
 430         rte_service_runstate_set(fdata->evdev_service_id, 1);
 431         rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0);
 432         if (rte_event_dev_start(dev_id) < 0) {
 433                 printf("Error starting eventdev\n");
 434                 return -1;
 435         }
 436
 437         return dev_id;
 438 }
 439
 440 static void
 441 init_rx_adapter(uint16_t nb_ports)
 442 {
 443         int i;
 444         int ret;
 445         uint8_t evdev_id = 0;
 446         struct rte_event_dev_info dev_info;
 447
 448         ret = rte_event_dev_info_get(evdev_id, &dev_info);
 449
 450         struct rte_event_port_conf rx_p_conf = {
 451                 .dequeue_depth = 8,
 452                 .enqueue_depth = 8,
 453                 .new_event_threshold = 1200,
 454         };
 455
 456         if (rx_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
 457                 rx_p_conf.dequeue_depth = dev_info.max_event_port_dequeue_depth;
 458         if (rx_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
 459                 rx_p_conf.enqueue_depth = dev_info.max_event_port_enqueue_depth;
 460
 461         /* Create one adapter for all the ethernet ports. */
 462         ret = rte_event_eth_rx_adapter_create(cdata.rx_adapter_id, evdev_id,
 463                         &rx_p_conf);
 464         if (ret)
 465                 rte_exit(EXIT_FAILURE, "failed to create rx adapter[%d]",
 466                                 cdata.rx_adapter_id);
 467
 468         struct rte_event_eth_rx_adapter_queue_conf queue_conf = {
 469                 .ev.sched_type = cdata.queue_type,
 470                 .ev.queue_id = cdata.qid[0],
 471         };
 472
 473         for (i = 0; i < nb_ports; i++) {
 474                 uint32_t cap;
 475
 476                 ret = rte_event_eth_rx_adapter_caps_get(evdev_id, i, &cap);
 477                 if (ret)
 478                         rte_exit(EXIT_FAILURE,
 479                                         "failed to get event rx adapter "
 480                                         "capabilities");
 481
 482                 ret = rte_event_eth_rx_adapter_queue_add(cdata.rx_adapter_id, i,
 483                                 -1, &queue_conf);
 484                 if (ret)
 485                         rte_exit(EXIT_FAILURE,
 486                                         "Failed to add queues to Rx adapter");
 487         }
 488
 489         ret = rte_event_eth_rx_adapter_service_id_get(cdata.rx_adapter_id,
 490                                 &fdata->rxadptr_service_id);
 491         if (ret != -ESRCH && ret != 0) {
 492                 rte_exit(EXIT_FAILURE,
 493                         "Error getting the service ID for sw eventdev\n");
 494         }
 495         rte_service_runstate_set(fdata->rxadptr_service_id, 1);
 496         rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id, 0);
 497
 498         ret = rte_event_eth_rx_adapter_start(cdata.rx_adapter_id);
 499         if (ret)
 500                 rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
 501                                 cdata.rx_adapter_id);
 502 }
 503
 504 static void
 505 generic_opt_check(void)
 506 {
 507         int i;
 508         int ret;
 509         uint32_t cap = 0;
 510         uint8_t rx_needed = 0;
 511         struct rte_event_dev_info eventdev_info;
 512
 513         memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
 514         rte_event_dev_info_get(0, &eventdev_info);
 515
 516         for (i = 0; i < rte_eth_dev_count(); i++) {
 517                 ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap);
 518                 if (ret)
 519                         rte_exit(EXIT_FAILURE,
 520                                 "failed to get event rx adapter capabilities");
 521                 rx_needed |=
 522                         !(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT);
 523         }
 524
 525         if (cdata.worker_lcore_mask == 0 ||
 526                         (rx_needed && cdata.rx_lcore_mask == 0) ||
 527                         cdata.tx_lcore_mask == 0 || (cdata.sched_lcore_mask == 0
 528                                 && !(eventdev_info.event_dev_cap &
 529                                         RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED))) {
 530                 printf("Core part of pipeline was not assigned any cores. "
 531                         "This will stall the pipeline, please check core masks "
 532                         "(use -h for details on setting core masks):\n"
 533                         "\trx: %"PRIu64"\n\ttx: %"PRIu64"\n\tsched: %"PRIu64
 534                         "\n\tworkers: %"PRIu64"\n",
 535                         cdata.rx_lcore_mask, cdata.tx_lcore_mask,
 536                         cdata.sched_lcore_mask,
 537                         cdata.worker_lcore_mask);
 538                 rte_exit(-1, "Fix core masks\n");
 539         }
 540
 541         if (eventdev_info.event_dev_cap & RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED)
 542                 memset(fdata->sched_core, 0,
 543                                 sizeof(unsigned int) * MAX_NUM_CORE);
 544 }
 545
 546 void
 547 set_worker_generic_setup_data(struct setup_data *caps, bool burst)
 548 {
 549         if (burst) {
 550                 caps->consumer = consumer_burst;
 551                 caps->worker = worker_generic_burst;
 552         } else {
 553                 caps->consumer = consumer;
 554                 caps->worker = worker_generic;
 555         }
 556
 557         caps->adptr_setup = init_rx_adapter;
 558         caps->scheduler = schedule_devices;
 559         caps->evdev_setup = setup_eventdev_generic;
 560         caps->check_opt = generic_opt_check;
 561 }