examples/eventdev_pipeline_sw_pmd/pipeline_worker_tx.c

   1 /*
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  * Copyright(c) 2010-2014 Intel Corporation
   4  * Copyright 2017 Cavium, Inc.
   5  */
   6
   7 #include "pipeline_common.h"
   8
   9 static __rte_always_inline void
  10 worker_fwd_event(struct rte_event *ev, uint8_t sched)
  11 {
  12         ev->event_type = RTE_EVENT_TYPE_CPU;
  13         ev->op = RTE_EVENT_OP_FORWARD;
  14         ev->sched_type = sched;
  15 }
  16
  17 static __rte_always_inline void
  18 worker_event_enqueue(const uint8_t dev, const uint8_t port,
  19                 struct rte_event *ev)
  20 {
  21         while (rte_event_enqueue_burst(dev, port, ev, 1) != 1)
  22                 rte_pause();
  23 }
  24
  25 static __rte_always_inline void
  26 worker_event_enqueue_burst(const uint8_t dev, const uint8_t port,
  27                 struct rte_event *ev, const uint16_t nb_rx)
  28 {
  29         uint16_t enq;
  30
  31         enq = rte_event_enqueue_burst(dev, port, ev, nb_rx);
  32         while (enq < nb_rx) {
  33                 enq += rte_event_enqueue_burst(dev, port,
  34                                                 ev + enq, nb_rx - enq);
  35         }
  36 }
  37
  38 static __rte_always_inline void
  39 worker_tx_pkt(struct rte_mbuf *mbuf)
  40 {
  41         exchange_mac(mbuf);
  42         while (rte_eth_tx_burst(mbuf->port, 0, &mbuf, 1) != 1)
  43                 rte_pause();
  44 }
  45
  46 /* Multi stage Pipeline Workers */
  47
  48 static int
  49 worker_do_tx(void *arg)
  50 {
  51         struct rte_event ev;
  52
  53         struct worker_data *data = (struct worker_data *)arg;
  54         const uint8_t dev = data->dev_id;
  55         const uint8_t port = data->port_id;
  56         const uint8_t lst_qid = cdata.num_stages - 1;
  57         size_t fwd = 0, received = 0, tx = 0;
  58
  59
  60         while (!fdata->done) {
  61
  62                 if (!rte_event_dequeue_burst(dev, port, &ev, 1, 0)) {
  63                         rte_pause();
  64                         continue;
  65                 }
  66
  67                 received++;
  68                 const uint8_t cq_id = ev.queue_id % cdata.num_stages;
  69
  70                 if (cq_id >= lst_qid) {
  71                         if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
  72                                 worker_tx_pkt(ev.mbuf);
  73                                 tx++;
  74                                 continue;
  75                         }
  76
  77                         worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
  78                         ev.queue_id = (cq_id == lst_qid) ?
  79                                 cdata.next_qid[ev.queue_id] : ev.queue_id;
  80                 } else {
  81                         ev.queue_id = cdata.next_qid[ev.queue_id];
  82                         worker_fwd_event(&ev, cdata.queue_type);
  83                 }
  84                 work();
  85
  86                 worker_event_enqueue(dev, port, &ev);
  87                 fwd++;
  88         }
  89
  90         if (!cdata.quiet)
  91                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
  92                                 rte_lcore_id(), received, fwd, tx);
  93
  94         return 0;
  95 }
  96
  97 static int
  98 worker_do_tx_burst(void *arg)
  99 {
 100         struct rte_event ev[BATCH_SIZE];
 101
 102         struct worker_data *data = (struct worker_data *)arg;
 103         uint8_t dev = data->dev_id;
 104         uint8_t port = data->port_id;
 105         uint8_t lst_qid = cdata.num_stages - 1;
 106         size_t fwd = 0, received = 0, tx = 0;
 107
 108         while (!fdata->done) {
 109                 uint16_t i;
 110                 const uint16_t nb_rx = rte_event_dequeue_burst(dev, port,
 111                                 ev, BATCH_SIZE, 0);
 112
 113                 if (nb_rx == 0) {
 114                         rte_pause();
 115                         continue;
 116                 }
 117                 received += nb_rx;
 118
 119                 for (i = 0; i < nb_rx; i++) {
 120                         const uint8_t cq_id = ev[i].queue_id % cdata.num_stages;
 121
 122                         if (cq_id >= lst_qid) {
 123                                 if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
 124                                         worker_tx_pkt(ev[i].mbuf);
 125                                         tx++;
 126                                         ev[i].op = RTE_EVENT_OP_RELEASE;
 127                                         continue;
 128                                 }
 129                                 ev[i].queue_id = (cq_id == lst_qid) ?
 130                                         cdata.next_qid[ev[i].queue_id] :
 131                                         ev[i].queue_id;
 132
 133                                 worker_fwd_event(&ev[i], RTE_SCHED_TYPE_ATOMIC);
 134                         } else {
 135                                 ev[i].queue_id = cdata.next_qid[ev[i].queue_id];
 136                                 worker_fwd_event(&ev[i], cdata.queue_type);
 137                         }
 138                         work();
 139                 }
 140                 worker_event_enqueue_burst(dev, port, ev, nb_rx);
 141
 142                 fwd += nb_rx;
 143         }
 144
 145         if (!cdata.quiet)
 146                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
 147                                 rte_lcore_id(), received, fwd, tx);
 148
 149         return 0;
 150 }
 151
 152 static int
 153 setup_eventdev_worker_tx(struct cons_data *cons_data,
 154                 struct worker_data *worker_data)
 155 {
 156         RTE_SET_USED(cons_data);
 157         uint8_t i;
 158         const uint8_t dev_id = 0;
 159         const uint8_t nb_ports = cdata.num_workers;
 160         uint8_t nb_slots = 0;
 161         uint8_t nb_queues = rte_eth_dev_count() * cdata.num_stages;
 162         nb_queues += rte_eth_dev_count();
 163
 164         struct rte_event_dev_config config = {
 165                         .nb_event_queues = nb_queues,
 166                         .nb_event_ports = nb_ports,
 167                         .nb_events_limit  = 4096,
 168                         .nb_event_queue_flows = 1024,
 169                         .nb_event_port_dequeue_depth = 128,
 170                         .nb_event_port_enqueue_depth = 128,
 171         };
 172         struct rte_event_port_conf wkr_p_conf = {
 173                         .dequeue_depth = cdata.worker_cq_depth,
 174                         .enqueue_depth = 64,
 175                         .new_event_threshold = 4096,
 176         };
 177         struct rte_event_queue_conf wkr_q_conf = {
 178                         .schedule_type = cdata.queue_type,
 179                         .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 180                         .nb_atomic_flows = 1024,
 181                         .nb_atomic_order_sequences = 1024,
 182         };
 183
 184         int ret, ndev = rte_event_dev_count();
 185
 186         if (ndev < 1) {
 187                 printf("%d: No Eventdev Devices Found\n", __LINE__);
 188                 return -1;
 189         }
 190
 191
 192         struct rte_event_dev_info dev_info;
 193         ret = rte_event_dev_info_get(dev_id, &dev_info);
 194         printf("\tEventdev %d: %s\n", dev_id, dev_info.driver_name);
 195
 196         if (dev_info.max_event_port_dequeue_depth <
 197                         config.nb_event_port_dequeue_depth)
 198                 config.nb_event_port_dequeue_depth =
 199                                 dev_info.max_event_port_dequeue_depth;
 200         if (dev_info.max_event_port_enqueue_depth <
 201                         config.nb_event_port_enqueue_depth)
 202                 config.nb_event_port_enqueue_depth =
 203                                 dev_info.max_event_port_enqueue_depth;
 204
 205         ret = rte_event_dev_configure(dev_id, &config);
 206         if (ret < 0) {
 207                 printf("%d: Error configuring device\n", __LINE__);
 208                 return -1;
 209         }
 210
 211         printf("  Stages:\n");
 212         for (i = 0; i < nb_queues; i++) {
 213
 214                 uint8_t slot;
 215
 216                 nb_slots = cdata.num_stages + 1;
 217                 slot = i % nb_slots;
 218                 wkr_q_conf.schedule_type = slot == cdata.num_stages ?
 219                         RTE_SCHED_TYPE_ATOMIC : cdata.queue_type;
 220
 221                 if (rte_event_queue_setup(dev_id, i, &wkr_q_conf) < 0) {
 222                         printf("%d: error creating qid %d\n", __LINE__, i);
 223                         return -1;
 224                 }
 225                 cdata.qid[i] = i;
 226                 cdata.next_qid[i] = i+1;
 227                 if (cdata.enable_queue_priorities) {
 228                         const uint32_t prio_delta =
 229                                 (RTE_EVENT_DEV_PRIORITY_LOWEST) /
 230                                 nb_slots;
 231
 232                         /* higher priority for queues closer to tx */
 233                         wkr_q_conf.priority =
 234                                 RTE_EVENT_DEV_PRIORITY_LOWEST - prio_delta *
 235                                 (i % nb_slots);
 236                 }
 237
 238                 const char *type_str = "Atomic";
 239                 switch (wkr_q_conf.schedule_type) {
 240                 case RTE_SCHED_TYPE_ORDERED:
 241                         type_str = "Ordered";
 242                         break;
 243                 case RTE_SCHED_TYPE_PARALLEL:
 244                         type_str = "Parallel";
 245                         break;
 246                 }
 247                 printf("\tStage %d, Type %s\tPriority = %d\n", i, type_str,
 248                                 wkr_q_conf.priority);
 249         }
 250
 251         printf("\n");
 252         if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
 253                 wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
 254         if (wkr_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
 255                 wkr_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
 256
 257         /* set up one port per worker, linking to all stage queues */
 258         for (i = 0; i < cdata.num_workers; i++) {
 259                 struct worker_data *w = &worker_data[i];
 260                 w->dev_id = dev_id;
 261                 if (rte_event_port_setup(dev_id, i, &wkr_p_conf) < 0) {
 262                         printf("Error setting up port %d\n", i);
 263                         return -1;
 264                 }
 265
 266                 if (rte_event_port_link(dev_id, i, NULL, NULL, 0)
 267                                 != nb_queues) {
 268                         printf("%d: error creating link for port %d\n",
 269                                         __LINE__, i);
 270                         return -1;
 271                 }
 272                 w->port_id = i;
 273         }
 274         /*
 275          * Reduce the load on ingress event queue by splitting the traffic
 276          * across multiple event queues.
 277          * for example, nb_stages =  2 and nb_ethdev = 2 then
 278          *
 279          *      nb_queues = (2 * 2) + 2 = 6 (non atq)
 280          *      rx_stride = 3
 281          *
 282          * So, traffic is split across queue 0 and queue 3 since queue id for
 283          * rx adapter is chosen <ethport_id> * <rx_stride> i.e in the above
 284          * case eth port 0, 1 will inject packets into event queue 0, 3
 285          * respectively.
 286          *
 287          * This forms two set of queue pipelines 0->1->2->tx and 3->4->5->tx.
 288          */
 289         cdata.rx_stride = nb_slots;
 290         ret = rte_event_dev_service_id_get(dev_id,
 291                                 &fdata->evdev_service_id);
 292         if (ret != -ESRCH && ret != 0) {
 293                 printf("Error getting the service ID\n");
 294                 return -1;
 295         }
 296         rte_service_runstate_set(fdata->evdev_service_id, 1);
 297         rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0);
 298         if (rte_event_dev_start(dev_id) < 0) {
 299                 printf("Error starting eventdev\n");
 300                 return -1;
 301         }
 302
 303         return dev_id;
 304 }
 305
 306
 307 struct rx_adptr_services {
 308         uint16_t nb_rx_adptrs;
 309         uint32_t *rx_adpt_arr;
 310 };
 311
 312 static int32_t
 313 service_rx_adapter(void *arg)
 314 {
 315         int i;
 316         struct rx_adptr_services *adptr_services = arg;
 317
 318         for (i = 0; i < adptr_services->nb_rx_adptrs; i++)
 319                 rte_service_run_iter_on_app_lcore(
 320                                 adptr_services->rx_adpt_arr[i], 1);
 321         return 0;
 322 }
 323
 324 static void
 325 init_rx_adapter(uint16_t nb_ports)
 326 {
 327         int i;
 328         int ret;
 329         uint8_t evdev_id = 0;
 330         struct rx_adptr_services *adptr_services = NULL;
 331         struct rte_event_dev_info dev_info;
 332
 333         ret = rte_event_dev_info_get(evdev_id, &dev_info);
 334         adptr_services = rte_zmalloc(NULL, sizeof(struct rx_adptr_services), 0);
 335
 336         struct rte_event_port_conf rx_p_conf = {
 337                 .dequeue_depth = 8,
 338                 .enqueue_depth = 8,
 339                 .new_event_threshold = 1200,
 340         };
 341
 342         if (rx_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
 343                 rx_p_conf.dequeue_depth = dev_info.max_event_port_dequeue_depth;
 344         if (rx_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
 345                 rx_p_conf.enqueue_depth = dev_info.max_event_port_enqueue_depth;
 346
 347
 348         struct rte_event_eth_rx_adapter_queue_conf queue_conf = {
 349                 .ev.sched_type = cdata.queue_type,
 350         };
 351
 352         for (i = 0; i < nb_ports; i++) {
 353                 uint32_t cap;
 354                 uint32_t service_id;
 355
 356                 ret = rte_event_eth_rx_adapter_create(i, evdev_id, &rx_p_conf);
 357                 if (ret)
 358                         rte_exit(EXIT_FAILURE,
 359                                         "failed to create rx adapter[%d]",
 360                                         cdata.rx_adapter_id);
 361
 362                 ret = rte_event_eth_rx_adapter_caps_get(evdev_id, i, &cap);
 363                 if (ret)
 364                         rte_exit(EXIT_FAILURE,
 365                                         "failed to get event rx adapter "
 366                                         "capabilities");
 367
 368                 queue_conf.ev.queue_id = cdata.rx_stride ?
 369                         (i * cdata.rx_stride)
 370                         : (uint8_t)cdata.qid[0];
 371
 372                 ret = rte_event_eth_rx_adapter_queue_add(i, i, -1, &queue_conf);
 373                 if (ret)
 374                         rte_exit(EXIT_FAILURE,
 375                                         "Failed to add queues to Rx adapter");
 376
 377
 378                 /* Producer needs to be scheduled. */
 379                 if (!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT)) {
 380                         ret = rte_event_eth_rx_adapter_service_id_get(i,
 381                                         &service_id);
 382                         if (ret != -ESRCH && ret != 0) {
 383                                 rte_exit(EXIT_FAILURE,
 384                                 "Error getting the service ID for rx adptr\n");
 385                         }
 386
 387                         rte_service_runstate_set(service_id, 1);
 388                         rte_service_set_runstate_mapped_check(service_id, 0);
 389
 390                         adptr_services->nb_rx_adptrs++;
 391                         adptr_services->rx_adpt_arr = rte_realloc(
 392                                         adptr_services->rx_adpt_arr,
 393                                         adptr_services->nb_rx_adptrs *
 394                                         sizeof(uint32_t), 0);
 395                         adptr_services->rx_adpt_arr[
 396                                 adptr_services->nb_rx_adptrs - 1] =
 397                                 service_id;
 398                 }
 399
 400                 ret = rte_event_eth_rx_adapter_start(i);
 401                 if (ret)
 402                         rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
 403                                         cdata.rx_adapter_id);
 404         }
 405
 406         if (adptr_services->nb_rx_adptrs) {
 407                 struct rte_service_spec service;
 408
 409                 memset(&service, 0, sizeof(struct rte_service_spec));
 410                 snprintf(service.name, sizeof(service.name), "rx_service");
 411                 service.callback = service_rx_adapter;
 412                 service.callback_userdata = (void *)adptr_services;
 413
 414                 int32_t ret = rte_service_component_register(&service,
 415                                 &fdata->rxadptr_service_id);
 416                 if (ret)
 417                         rte_exit(EXIT_FAILURE,
 418                                 "Rx adapter[%d] service register failed",
 419                                 cdata.rx_adapter_id);
 420
 421                 rte_service_runstate_set(fdata->rxadptr_service_id, 1);
 422                 rte_service_component_runstate_set(fdata->rxadptr_service_id,
 423                                 1);
 424                 rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id,
 425                                 0);
 426         } else {
 427                 memset(fdata->rx_core, 0, sizeof(unsigned int) * MAX_NUM_CORE);
 428                 rte_free(adptr_services);
 429         }
 430
 431         if (!adptr_services->nb_rx_adptrs && fdata->cap.consumer == NULL &&
 432                         (dev_info.event_dev_cap &
 433                          RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED))
 434                 fdata->cap.scheduler = NULL;
 435
 436         if (dev_info.event_dev_cap & RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED)
 437                 memset(fdata->sched_core, 0,
 438                                 sizeof(unsigned int) * MAX_NUM_CORE);
 439 }
 440
 441 static void
 442 worker_tx_opt_check(void)
 443 {
 444         int i;
 445         int ret;
 446         uint32_t cap = 0;
 447         uint8_t rx_needed = 0;
 448         struct rte_event_dev_info eventdev_info;
 449
 450         memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
 451         rte_event_dev_info_get(0, &eventdev_info);
 452
 453         for (i = 0; i < rte_eth_dev_count(); i++) {
 454                 ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap);
 455                 if (ret)
 456                         rte_exit(EXIT_FAILURE,
 457                                         "failed to get event rx adapter "
 458                                         "capabilities");
 459                 rx_needed |=
 460                         !(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT);
 461         }
 462
 463         if (cdata.worker_lcore_mask == 0 ||
 464                         (rx_needed && cdata.rx_lcore_mask == 0) ||
 465                         (cdata.sched_lcore_mask == 0 &&
 466                          !(eventdev_info.event_dev_cap &
 467                                  RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED))) {
 468                 printf("Core part of pipeline was not assigned any cores. "
 469                         "This will stall the pipeline, please check core masks "
 470                         "(use -h for details on setting core masks):\n"
 471                         "\trx: %"PRIu64"\n\ttx: %"PRIu64"\n\tsched: %"PRIu64
 472                         "\n\tworkers: %"PRIu64"\n",
 473                         cdata.rx_lcore_mask, cdata.tx_lcore_mask,
 474                         cdata.sched_lcore_mask,
 475                         cdata.worker_lcore_mask);
 476                 rte_exit(-1, "Fix core masks\n");
 477         }
 478 }
 479
 480 void
 481 set_worker_tx_setup_data(struct setup_data *caps, bool burst)
 482 {
 483         if (burst)
 484                 caps->worker = worker_do_tx_burst;
 485         else
 486                 caps->worker = worker_do_tx;
 487
 488         memset(fdata->tx_core, 0, sizeof(unsigned int) * MAX_NUM_CORE);
 489
 490         caps->check_opt = worker_tx_opt_check;
 491         caps->consumer = NULL;
 492         caps->scheduler = schedule_devices;
 493         caps->evdev_setup = setup_eventdev_worker_tx;
 494         caps->adptr_setup = init_rx_adapter;
 495 }