examples/eventdev_pipeline/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) && !fdata->done)
  22                 rte_pause();
  23 }
  24
  25 static __rte_always_inline uint16_t
  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 && !fdata->done)
  33                 enq += rte_event_enqueue_burst(dev, port,
  34                                                 ev + enq, nb_rx - enq);
  35
  36         return enq;
  37 }
  38
  39 static __rte_always_inline void
  40 worker_tx_pkt(const uint8_t dev, const uint8_t port, struct rte_event *ev)
  41 {
  42         exchange_mac(ev->mbuf);
  43         rte_event_eth_tx_adapter_txq_set(ev->mbuf, 0);
  44         while (!rte_event_eth_tx_adapter_enqueue(dev, port, ev, 1, 0) &&
  45                !fdata->done)
  46                 rte_pause();
  47 }
  48
  49 /* Single stage pipeline workers */
  50
  51 static int
  52 worker_do_tx_single(void *arg)
  53 {
  54         struct worker_data *data = (struct worker_data *)arg;
  55         const uint8_t dev = data->dev_id;
  56         const uint8_t port = data->port_id;
  57         size_t fwd = 0, received = 0, tx = 0;
  58         struct rte_event ev;
  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
  69                 if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
  70                         worker_tx_pkt(dev, port, &ev);
  71                         tx++;
  72                 } else {
  73                         work();
  74                         ev.queue_id++;
  75                         worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
  76                         worker_event_enqueue(dev, port, &ev);
  77                         fwd++;
  78                 }
  79         }
  80
  81         if (ev.u64) {
  82                 ev.op = RTE_EVENT_OP_RELEASE;
  83                 rte_event_enqueue_burst(dev, port, &ev, 1);
  84         }
  85
  86         if (!cdata.quiet)
  87                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
  88                                 rte_lcore_id(), received, fwd, tx);
  89         return 0;
  90 }
  91
  92 static int
  93 worker_do_tx_single_atq(void *arg)
  94 {
  95         struct worker_data *data = (struct worker_data *)arg;
  96         const uint8_t dev = data->dev_id;
  97         const uint8_t port = data->port_id;
  98         size_t fwd = 0, received = 0, tx = 0;
  99         struct rte_event ev;
 100
 101         while (!fdata->done) {
 102
 103                 if (!rte_event_dequeue_burst(dev, port, &ev, 1, 0)) {
 104                         rte_pause();
 105                         continue;
 106                 }
 107
 108                 received++;
 109
 110                 if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
 111                         worker_tx_pkt(dev, port, &ev);
 112                         tx++;
 113                 } else {
 114                         work();
 115                         worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
 116                         worker_event_enqueue(dev, port, &ev);
 117                         fwd++;
 118                 }
 119         }
 120
 121         if (ev.u64) {
 122                 ev.op = RTE_EVENT_OP_RELEASE;
 123                 rte_event_enqueue_burst(dev, port, &ev, 1);
 124         }
 125
 126         if (!cdata.quiet)
 127                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
 128                                 rte_lcore_id(), received, fwd, tx);
 129         return 0;
 130 }
 131
 132 static int
 133 worker_do_tx_single_burst(void *arg)
 134 {
 135         struct rte_event ev[BATCH_SIZE + 1];
 136
 137         struct worker_data *data = (struct worker_data *)arg;
 138         const uint8_t dev = data->dev_id;
 139         const uint8_t port = data->port_id;
 140         size_t fwd = 0, received = 0, tx = 0;
 141         uint16_t nb_tx = 0, nb_rx = 0, i;
 142
 143         while (!fdata->done) {
 144                 nb_rx = rte_event_dequeue_burst(dev, port, ev, BATCH_SIZE, 0);
 145
 146                 if (!nb_rx) {
 147                         rte_pause();
 148                         continue;
 149                 }
 150                 received += nb_rx;
 151
 152                 for (i = 0; i < nb_rx; i++) {
 153                         rte_prefetch0(ev[i + 1].mbuf);
 154                         if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
 155
 156                                 worker_tx_pkt(dev, port, &ev[i]);
 157                                 ev[i].op = RTE_EVENT_OP_RELEASE;
 158                                 tx++;
 159
 160                         } else {
 161                                 ev[i].queue_id++;
 162                                 worker_fwd_event(&ev[i], RTE_SCHED_TYPE_ATOMIC);
 163                         }
 164                         work();
 165                 }
 166
 167                 nb_tx = worker_event_enqueue_burst(dev, port, ev, nb_rx);
 168                 fwd += nb_tx;
 169         }
 170
 171         worker_cleanup(dev, port, ev, nb_tx, nb_rx);
 172
 173         if (!cdata.quiet)
 174                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
 175                                 rte_lcore_id(), received, fwd, tx);
 176         return 0;
 177 }
 178
 179 static int
 180 worker_do_tx_single_burst_atq(void *arg)
 181 {
 182         struct rte_event ev[BATCH_SIZE + 1];
 183
 184         struct worker_data *data = (struct worker_data *)arg;
 185         const uint8_t dev = data->dev_id;
 186         const uint8_t port = data->port_id;
 187         size_t fwd = 0, received = 0, tx = 0;
 188         uint16_t i, nb_rx = 0, nb_tx = 0;
 189
 190         while (!fdata->done) {
 191                 nb_rx = rte_event_dequeue_burst(dev, port, ev, BATCH_SIZE, 0);
 192
 193                 if (!nb_rx) {
 194                         rte_pause();
 195                         continue;
 196                 }
 197
 198                 received += nb_rx;
 199
 200                 for (i = 0; i < nb_rx; i++) {
 201                         rte_prefetch0(ev[i + 1].mbuf);
 202                         if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
 203
 204                                 worker_tx_pkt(dev, port, &ev[i]);
 205                                 ev[i].op = RTE_EVENT_OP_RELEASE;
 206                                 tx++;
 207                         } else
 208                                 worker_fwd_event(&ev[i], RTE_SCHED_TYPE_ATOMIC);
 209                         work();
 210                 }
 211
 212                 nb_tx = worker_event_enqueue_burst(dev, port, ev, nb_rx);
 213                 fwd += nb_tx;
 214         }
 215
 216         worker_cleanup(dev, port, ev, nb_tx, nb_rx);
 217
 218         if (!cdata.quiet)
 219                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
 220                                 rte_lcore_id(), received, fwd, tx);
 221         return 0;
 222 }
 223
 224 /* Multi stage Pipeline Workers */
 225
 226 static int
 227 worker_do_tx(void *arg)
 228 {
 229         struct rte_event ev;
 230
 231         struct worker_data *data = (struct worker_data *)arg;
 232         const uint8_t dev = data->dev_id;
 233         const uint8_t port = data->port_id;
 234         const uint8_t lst_qid = cdata.num_stages - 1;
 235         size_t fwd = 0, received = 0, tx = 0;
 236
 237
 238         while (!fdata->done) {
 239
 240                 if (!rte_event_dequeue_burst(dev, port, &ev, 1, 0)) {
 241                         rte_pause();
 242                         continue;
 243                 }
 244
 245                 received++;
 246                 const uint8_t cq_id = ev.queue_id % cdata.num_stages;
 247
 248                 if (cq_id >= lst_qid) {
 249                         if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
 250                                 worker_tx_pkt(dev, port, &ev);
 251                                 tx++;
 252                                 continue;
 253                         }
 254
 255                         worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
 256                         ev.queue_id = (cq_id == lst_qid) ?
 257                                 cdata.next_qid[ev.queue_id] : ev.queue_id;
 258                 } else {
 259                         ev.queue_id = cdata.next_qid[ev.queue_id];
 260                         worker_fwd_event(&ev, cdata.queue_type);
 261                 }
 262                 work();
 263
 264                 worker_event_enqueue(dev, port, &ev);
 265                 fwd++;
 266         }
 267
 268         if (ev.u64) {
 269                 ev.op = RTE_EVENT_OP_RELEASE;
 270                 rte_event_enqueue_burst(dev, port, &ev, 1);
 271         }
 272
 273         if (!cdata.quiet)
 274                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
 275                                 rte_lcore_id(), received, fwd, tx);
 276
 277         return 0;
 278 }
 279
 280 static int
 281 worker_do_tx_atq(void *arg)
 282 {
 283         struct rte_event ev;
 284
 285         struct worker_data *data = (struct worker_data *)arg;
 286         const uint8_t dev = data->dev_id;
 287         const uint8_t port = data->port_id;
 288         const uint8_t lst_qid = cdata.num_stages - 1;
 289         size_t fwd = 0, received = 0, tx = 0;
 290
 291         while (!fdata->done) {
 292
 293                 if (!rte_event_dequeue_burst(dev, port, &ev, 1, 0)) {
 294                         rte_pause();
 295                         continue;
 296                 }
 297
 298                 received++;
 299                 const uint8_t cq_id = ev.sub_event_type % cdata.num_stages;
 300
 301                 if (cq_id == lst_qid) {
 302                         if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
 303                                 worker_tx_pkt(dev, port, &ev);
 304                                 tx++;
 305                                 continue;
 306                         }
 307
 308                         worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
 309                 } else {
 310                         ev.sub_event_type++;
 311                         worker_fwd_event(&ev, cdata.queue_type);
 312                 }
 313                 work();
 314
 315                 worker_event_enqueue(dev, port, &ev);
 316                 fwd++;
 317         }
 318
 319         if (ev.u64) {
 320                 ev.op = RTE_EVENT_OP_RELEASE;
 321                 rte_event_enqueue_burst(dev, port, &ev, 1);
 322         }
 323
 324         if (!cdata.quiet)
 325                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
 326                                 rte_lcore_id(), received, fwd, tx);
 327
 328         return 0;
 329 }
 330
 331 static int
 332 worker_do_tx_burst(void *arg)
 333 {
 334         struct rte_event ev[BATCH_SIZE];
 335
 336         struct worker_data *data = (struct worker_data *)arg;
 337         uint8_t dev = data->dev_id;
 338         uint8_t port = data->port_id;
 339         uint8_t lst_qid = cdata.num_stages - 1;
 340         size_t fwd = 0, received = 0, tx = 0;
 341         uint16_t i, nb_rx = 0, nb_tx = 0;
 342
 343         while (!fdata->done) {
 344                 nb_rx = rte_event_dequeue_burst(dev, port, ev, BATCH_SIZE, 0);
 345
 346                 if (nb_rx == 0) {
 347                         rte_pause();
 348                         continue;
 349                 }
 350                 received += nb_rx;
 351
 352                 for (i = 0; i < nb_rx; i++) {
 353                         const uint8_t cq_id = ev[i].queue_id % cdata.num_stages;
 354
 355                         if (cq_id >= lst_qid) {
 356                                 if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
 357                                         worker_tx_pkt(dev, port, &ev[i]);
 358                                         tx++;
 359                                         ev[i].op = RTE_EVENT_OP_RELEASE;
 360                                         continue;
 361                                 }
 362                                 ev[i].queue_id = (cq_id == lst_qid) ?
 363                                         cdata.next_qid[ev[i].queue_id] :
 364                                         ev[i].queue_id;
 365
 366                                 worker_fwd_event(&ev[i], RTE_SCHED_TYPE_ATOMIC);
 367                         } else {
 368                                 ev[i].queue_id = cdata.next_qid[ev[i].queue_id];
 369                                 worker_fwd_event(&ev[i], cdata.queue_type);
 370                         }
 371                         work();
 372                 }
 373
 374                 nb_tx = worker_event_enqueue_burst(dev, port, ev, nb_rx);
 375                 fwd += nb_tx;
 376         }
 377
 378         worker_cleanup(dev, port, ev, nb_tx, nb_rx);
 379
 380         if (!cdata.quiet)
 381                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
 382                                 rte_lcore_id(), received, fwd, tx);
 383
 384         return 0;
 385 }
 386
 387 static int
 388 worker_do_tx_burst_atq(void *arg)
 389 {
 390         struct rte_event ev[BATCH_SIZE];
 391
 392         struct worker_data *data = (struct worker_data *)arg;
 393         uint8_t dev = data->dev_id;
 394         uint8_t port = data->port_id;
 395         uint8_t lst_qid = cdata.num_stages - 1;
 396         size_t fwd = 0, received = 0, tx = 0;
 397         uint16_t i, nb_rx = 0, nb_tx = 0;
 398
 399         while (!fdata->done) {
 400                 nb_rx = rte_event_dequeue_burst(dev, port, ev, BATCH_SIZE, 0);
 401
 402                 if (nb_rx == 0) {
 403                         rte_pause();
 404                         continue;
 405                 }
 406                 received += nb_rx;
 407
 408                 for (i = 0; i < nb_rx; i++) {
 409                         const uint8_t cq_id = ev[i].sub_event_type %
 410                                 cdata.num_stages;
 411
 412                         if (cq_id == lst_qid) {
 413                                 if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
 414                                         worker_tx_pkt(dev, port, &ev[i]);
 415                                         tx++;
 416                                         ev[i].op = RTE_EVENT_OP_RELEASE;
 417                                         continue;
 418                                 }
 419
 420                                 worker_fwd_event(&ev[i], RTE_SCHED_TYPE_ATOMIC);
 421                         } else {
 422                                 ev[i].sub_event_type++;
 423                                 worker_fwd_event(&ev[i], cdata.queue_type);
 424                         }
 425                         work();
 426                 }
 427
 428                 nb_tx = worker_event_enqueue_burst(dev, port, ev, nb_rx);
 429                 fwd += nb_tx;
 430         }
 431
 432         worker_cleanup(dev, port, ev, nb_tx, nb_rx);
 433
 434         if (!cdata.quiet)
 435                 printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
 436                                 rte_lcore_id(), received, fwd, tx);
 437
 438         return 0;
 439 }
 440
 441 static int
 442 setup_eventdev_worker_tx_enq(struct worker_data *worker_data)
 443 {
 444         uint8_t i;
 445         const uint8_t atq = cdata.all_type_queues ? 1 : 0;
 446         const uint8_t dev_id = 0;
 447         const uint8_t nb_ports = cdata.num_workers;
 448         uint8_t nb_slots = 0;
 449         uint8_t nb_queues = rte_eth_dev_count_avail();
 450
 451         /*
 452          * In case where all type queues are not enabled, use queues equal to
 453          * number of stages * eth_dev_count and one extra queue per pipeline
 454          * for Tx.
 455          */
 456         if (!atq) {
 457                 nb_queues *= cdata.num_stages;
 458                 nb_queues += rte_eth_dev_count_avail();
 459         }
 460
 461         struct rte_event_dev_config config = {
 462                         .nb_event_queues = nb_queues,
 463                         .nb_event_ports = nb_ports,
 464                         .nb_single_link_event_port_queues = 0,
 465                         .nb_events_limit  = 4096,
 466                         .nb_event_queue_flows = 1024,
 467                         .nb_event_port_dequeue_depth = 128,
 468                         .nb_event_port_enqueue_depth = 128,
 469         };
 470         struct rte_event_port_conf wkr_p_conf = {
 471                         .dequeue_depth = cdata.worker_cq_depth,
 472                         .enqueue_depth = 64,
 473                         .new_event_threshold = 4096,
 474                         .event_port_cfg = RTE_EVENT_PORT_CFG_HINT_WORKER,
 475         };
 476         struct rte_event_queue_conf wkr_q_conf = {
 477                         .schedule_type = cdata.queue_type,
 478                         .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 479                         .nb_atomic_flows = 1024,
 480                         .nb_atomic_order_sequences = 1024,
 481         };
 482
 483         int ret, ndev = rte_event_dev_count();
 484
 485         if (ndev < 1) {
 486                 printf("%d: No Eventdev Devices Found\n", __LINE__);
 487                 return -1;
 488         }
 489
 490
 491         struct rte_event_dev_info dev_info;
 492         ret = rte_event_dev_info_get(dev_id, &dev_info);
 493         printf("\tEventdev %d: %s\n", dev_id, dev_info.driver_name);
 494
 495         if (dev_info.max_num_events < config.nb_events_limit)
 496                 config.nb_events_limit = dev_info.max_num_events;
 497         if (dev_info.max_event_port_dequeue_depth <
 498                         config.nb_event_port_dequeue_depth)
 499                 config.nb_event_port_dequeue_depth =
 500                                 dev_info.max_event_port_dequeue_depth;
 501         if (dev_info.max_event_port_enqueue_depth <
 502                         config.nb_event_port_enqueue_depth)
 503                 config.nb_event_port_enqueue_depth =
 504                                 dev_info.max_event_port_enqueue_depth;
 505
 506         ret = rte_event_dev_configure(dev_id, &config);
 507         if (ret < 0) {
 508                 printf("%d: Error configuring device\n", __LINE__);
 509                 return -1;
 510         }
 511
 512         printf("  Stages:\n");
 513         for (i = 0; i < nb_queues; i++) {
 514
 515                 if (atq) {
 516
 517                         nb_slots = cdata.num_stages;
 518                         wkr_q_conf.event_queue_cfg =
 519                                 RTE_EVENT_QUEUE_CFG_ALL_TYPES;
 520                 } else {
 521                         uint8_t slot;
 522
 523                         nb_slots = cdata.num_stages + 1;
 524                         slot = i % nb_slots;
 525                         wkr_q_conf.schedule_type = slot == cdata.num_stages ?
 526                                 RTE_SCHED_TYPE_ATOMIC : cdata.queue_type;
 527                 }
 528
 529                 if (rte_event_queue_setup(dev_id, i, &wkr_q_conf) < 0) {
 530                         printf("%d: error creating qid %d\n", __LINE__, i);
 531                         return -1;
 532                 }
 533                 cdata.qid[i] = i;
 534                 cdata.next_qid[i] = i+1;
 535                 if (cdata.enable_queue_priorities) {
 536                         const uint32_t prio_delta =
 537                                 (RTE_EVENT_DEV_PRIORITY_LOWEST) /
 538                                 nb_slots;
 539
 540                         /* higher priority for queues closer to tx */
 541                         wkr_q_conf.priority =
 542                                 RTE_EVENT_DEV_PRIORITY_LOWEST - prio_delta *
 543                                 (i % nb_slots);
 544                 }
 545
 546                 const char *type_str = "Atomic";
 547                 switch (wkr_q_conf.schedule_type) {
 548                 case RTE_SCHED_TYPE_ORDERED:
 549                         type_str = "Ordered";
 550                         break;
 551                 case RTE_SCHED_TYPE_PARALLEL:
 552                         type_str = "Parallel";
 553                         break;
 554                 }
 555                 printf("\tStage %d, Type %s\tPriority = %d\n", i, type_str,
 556                                 wkr_q_conf.priority);
 557         }
 558
 559         printf("\n");
 560         if (wkr_p_conf.new_event_threshold > config.nb_events_limit)
 561                 wkr_p_conf.new_event_threshold = config.nb_events_limit;
 562         if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
 563                 wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
 564         if (wkr_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
 565                 wkr_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
 566
 567         /* set up one port per worker, linking to all stage queues */
 568         for (i = 0; i < cdata.num_workers; i++) {
 569                 struct worker_data *w = &worker_data[i];
 570                 w->dev_id = dev_id;
 571                 if (rte_event_port_setup(dev_id, i, &wkr_p_conf) < 0) {
 572                         printf("Error setting up port %d\n", i);
 573                         return -1;
 574                 }
 575
 576                 if (rte_event_port_link(dev_id, i, NULL, NULL, 0)
 577                                 != nb_queues) {
 578                         printf("%d: error creating link for port %d\n",
 579                                         __LINE__, i);
 580                         return -1;
 581                 }
 582                 w->port_id = i;
 583         }
 584         /*
 585          * Reduce the load on ingress event queue by splitting the traffic
 586          * across multiple event queues.
 587          * for example, nb_stages =  2 and nb_ethdev = 2 then
 588          *
 589          *      nb_queues = (2 * 2) + 2 = 6 (non atq)
 590          *      rx_stride = 3
 591          *
 592          * So, traffic is split across queue 0 and queue 3 since queue id for
 593          * rx adapter is chosen <ethport_id> * <rx_stride> i.e in the above
 594          * case eth port 0, 1 will inject packets into event queue 0, 3
 595          * respectively.
 596          *
 597          * This forms two set of queue pipelines 0->1->2->tx and 3->4->5->tx.
 598          */
 599         cdata.rx_stride = atq ? 1 : nb_slots;
 600         ret = rte_event_dev_service_id_get(dev_id,
 601                                 &fdata->evdev_service_id);
 602         if (ret != -ESRCH && ret != 0) {
 603                 printf("Error getting the service ID\n");
 604                 return -1;
 605         }
 606         rte_service_runstate_set(fdata->evdev_service_id, 1);
 607         rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0);
 608
 609         if (rte_event_dev_start(dev_id) < 0)
 610                 rte_exit(EXIT_FAILURE, "Error starting eventdev");
 611
 612         return dev_id;
 613 }
 614
 615
 616 struct rx_adptr_services {
 617         uint16_t nb_rx_adptrs;
 618         uint32_t *rx_adpt_arr;
 619 };
 620
 621 static int32_t
 622 service_rx_adapter(void *arg)
 623 {
 624         int i;
 625         struct rx_adptr_services *adptr_services = arg;
 626
 627         for (i = 0; i < adptr_services->nb_rx_adptrs; i++)
 628                 rte_service_run_iter_on_app_lcore(
 629                                 adptr_services->rx_adpt_arr[i], 1);
 630         return 0;
 631 }
 632
 633 /*
 634  * Initializes a given port using global settings and with the RX buffers
 635  * coming from the mbuf_pool passed as a parameter.
 636  */
 637 static inline int
 638 port_init(uint8_t port, struct rte_mempool *mbuf_pool)
 639 {
 640         struct rte_eth_rxconf rx_conf;
 641         static const struct rte_eth_conf port_conf_default = {
 642                 .rxmode = {
 643                         .mq_mode = RTE_ETH_MQ_RX_RSS,
 644                 },
 645                 .rx_adv_conf = {
 646                         .rss_conf = {
 647                                 .rss_hf = RTE_ETH_RSS_IP |
 648                                           RTE_ETH_RSS_TCP |
 649                                           RTE_ETH_RSS_UDP,
 650                         }
 651                 }
 652         };
 653         const uint16_t rx_rings = 1, tx_rings = 1;
 654         const uint16_t rx_ring_size = 512, tx_ring_size = 512;
 655         struct rte_eth_conf port_conf = port_conf_default;
 656         int retval;
 657         uint16_t q;
 658         struct rte_eth_dev_info dev_info;
 659         struct rte_eth_txconf txconf;
 660
 661         if (!rte_eth_dev_is_valid_port(port))
 662                 return -1;
 663
 664         retval = rte_eth_dev_info_get(port, &dev_info);
 665         if (retval != 0) {
 666                 printf("Error during getting device (port %u) info: %s\n",
 667                                 port, strerror(-retval));
 668                 return retval;
 669         }
 670
 671         if (dev_info.tx_offload_capa & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE)
 672                 port_conf.txmode.offloads |=
 673                         RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE;
 674         rx_conf = dev_info.default_rxconf;
 675         rx_conf.offloads = port_conf.rxmode.offloads;
 676
 677         port_conf.rx_adv_conf.rss_conf.rss_hf &=
 678                 dev_info.flow_type_rss_offloads;
 679         if (port_conf.rx_adv_conf.rss_conf.rss_hf !=
 680                         port_conf_default.rx_adv_conf.rss_conf.rss_hf) {
 681                 printf("Port %u modified RSS hash function based on hardware support,"
 682                         "requested:%#"PRIx64" configured:%#"PRIx64"\n",
 683                         port,
 684                         port_conf_default.rx_adv_conf.rss_conf.rss_hf,
 685                         port_conf.rx_adv_conf.rss_conf.rss_hf);
 686         }
 687
 688         /* Configure the Ethernet device. */
 689         retval = rte_eth_dev_configure(port, rx_rings, tx_rings, &port_conf);
 690         if (retval != 0)
 691                 return retval;
 692
 693         /* Allocate and set up 1 RX queue per Ethernet port. */
 694         for (q = 0; q < rx_rings; q++) {
 695                 retval = rte_eth_rx_queue_setup(port, q, rx_ring_size,
 696                                 rte_eth_dev_socket_id(port), &rx_conf,
 697                                 mbuf_pool);
 698                 if (retval < 0)
 699                         return retval;
 700         }
 701
 702         txconf = dev_info.default_txconf;
 703         txconf.offloads = port_conf_default.txmode.offloads;
 704         /* Allocate and set up 1 TX queue per Ethernet port. */
 705         for (q = 0; q < tx_rings; q++) {
 706                 retval = rte_eth_tx_queue_setup(port, q, tx_ring_size,
 707                                 rte_eth_dev_socket_id(port), &txconf);
 708                 if (retval < 0)
 709                         return retval;
 710         }
 711
 712         /* Display the port MAC address. */
 713         struct rte_ether_addr addr;
 714         retval = rte_eth_macaddr_get(port, &addr);
 715         if (retval != 0) {
 716                 printf("Failed to get MAC address (port %u): %s\n",
 717                                 port, rte_strerror(-retval));
 718                 return retval;
 719         }
 720
 721         printf("Port %u MAC: %02" PRIx8 " %02" PRIx8 " %02" PRIx8
 722                         " %02" PRIx8 " %02" PRIx8 " %02" PRIx8 "\n",
 723                         (unsigned int)port, RTE_ETHER_ADDR_BYTES(&addr));
 724
 725         /* Enable RX in promiscuous mode for the Ethernet device. */
 726         retval = rte_eth_promiscuous_enable(port);
 727         if (retval != 0)
 728                 return retval;
 729
 730         return 0;
 731 }
 732
 733 static int
 734 init_ports(uint16_t num_ports)
 735 {
 736         uint16_t portid;
 737
 738         if (!cdata.num_mbuf)
 739                 cdata.num_mbuf = 16384 * num_ports;
 740
 741         struct rte_mempool *mp = rte_pktmbuf_pool_create("packet_pool",
 742                         /* mbufs */ cdata.num_mbuf,
 743                         /* cache_size */ 512,
 744                         /* priv_size*/ 0,
 745                         /* data_room_size */ RTE_MBUF_DEFAULT_BUF_SIZE,
 746                         rte_socket_id());
 747
 748         RTE_ETH_FOREACH_DEV(portid)
 749                 if (port_init(portid, mp) != 0)
 750                         rte_exit(EXIT_FAILURE, "Cannot init port %"PRIu16 "\n",
 751                                         portid);
 752
 753         return 0;
 754 }
 755
 756 static void
 757 init_adapters(uint16_t nb_ports)
 758 {
 759         int i;
 760         int ret;
 761         uint8_t evdev_id = 0;
 762         struct rx_adptr_services *adptr_services = NULL;
 763         struct rte_event_dev_info dev_info;
 764
 765         ret = rte_event_dev_info_get(evdev_id, &dev_info);
 766         adptr_services = rte_zmalloc(NULL, sizeof(struct rx_adptr_services), 0);
 767
 768         struct rte_event_port_conf adptr_p_conf = {
 769                 .dequeue_depth = cdata.worker_cq_depth,
 770                 .enqueue_depth = 64,
 771                 .new_event_threshold = 4096,
 772                 .event_port_cfg = RTE_EVENT_PORT_CFG_HINT_PRODUCER,
 773         };
 774
 775         init_ports(nb_ports);
 776         if (adptr_p_conf.new_event_threshold > dev_info.max_num_events)
 777                 adptr_p_conf.new_event_threshold = dev_info.max_num_events;
 778         if (adptr_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
 779                 adptr_p_conf.dequeue_depth =
 780                         dev_info.max_event_port_dequeue_depth;
 781         if (adptr_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
 782                 adptr_p_conf.enqueue_depth =
 783                         dev_info.max_event_port_enqueue_depth;
 784
 785         struct rte_event_eth_rx_adapter_queue_conf queue_conf;
 786         memset(&queue_conf, 0, sizeof(queue_conf));
 787         queue_conf.ev.sched_type = cdata.queue_type;
 788
 789         for (i = 0; i < nb_ports; i++) {
 790                 uint32_t cap;
 791                 uint32_t service_id;
 792
 793                 ret = rte_event_eth_rx_adapter_create(i, evdev_id,
 794                                 &adptr_p_conf);
 795                 if (ret)
 796                         rte_exit(EXIT_FAILURE,
 797                                         "failed to create rx adapter[%d]", i);
 798
 799                 ret = rte_event_eth_rx_adapter_caps_get(evdev_id, i, &cap);
 800                 if (ret)
 801                         rte_exit(EXIT_FAILURE,
 802                                         "failed to get event rx adapter "
 803                                         "capabilities");
 804
 805                 queue_conf.ev.queue_id = cdata.rx_stride ?
 806                         (i * cdata.rx_stride)
 807                         : (uint8_t)cdata.qid[0];
 808
 809                 ret = rte_event_eth_rx_adapter_queue_add(i, i, -1, &queue_conf);
 810                 if (ret)
 811                         rte_exit(EXIT_FAILURE,
 812                                         "Failed to add queues to Rx adapter");
 813
 814                 /* Producer needs to be scheduled. */
 815                 if (!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT)) {
 816                         ret = rte_event_eth_rx_adapter_service_id_get(i,
 817                                         &service_id);
 818                         if (ret != -ESRCH && ret != 0) {
 819                                 rte_exit(EXIT_FAILURE,
 820                                 "Error getting the service ID for rx adptr\n");
 821                         }
 822
 823                         rte_service_runstate_set(service_id, 1);
 824                         rte_service_set_runstate_mapped_check(service_id, 0);
 825
 826                         adptr_services->nb_rx_adptrs++;
 827                         adptr_services->rx_adpt_arr = rte_realloc(
 828                                         adptr_services->rx_adpt_arr,
 829                                         adptr_services->nb_rx_adptrs *
 830                                         sizeof(uint32_t), 0);
 831                         adptr_services->rx_adpt_arr[
 832                                 adptr_services->nb_rx_adptrs - 1] =
 833                                 service_id;
 834                 }
 835
 836                 ret = rte_event_eth_rx_adapter_start(i);
 837                 if (ret)
 838                         rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
 839                                         i);
 840         }
 841
 842         /* We already know that Tx adapter has INTERNAL port cap*/
 843         ret = rte_event_eth_tx_adapter_create(cdata.tx_adapter_id, evdev_id,
 844                         &adptr_p_conf);
 845         if (ret)
 846                 rte_exit(EXIT_FAILURE, "failed to create tx adapter[%d]",
 847                                 cdata.tx_adapter_id);
 848
 849         for (i = 0; i < nb_ports; i++) {
 850                 ret = rte_event_eth_tx_adapter_queue_add(cdata.tx_adapter_id, i,
 851                                 -1);
 852                 if (ret)
 853                         rte_exit(EXIT_FAILURE,
 854                                         "Failed to add queues to Tx adapter");
 855         }
 856
 857         ret = rte_event_eth_tx_adapter_start(cdata.tx_adapter_id);
 858         if (ret)
 859                 rte_exit(EXIT_FAILURE, "Tx adapter[%d] start failed",
 860                                 cdata.tx_adapter_id);
 861
 862         if (adptr_services->nb_rx_adptrs) {
 863                 struct rte_service_spec service;
 864
 865                 memset(&service, 0, sizeof(struct rte_service_spec));
 866                 snprintf(service.name, sizeof(service.name), "rx_service");
 867                 service.callback = service_rx_adapter;
 868                 service.callback_userdata = (void *)adptr_services;
 869
 870                 int32_t ret = rte_service_component_register(&service,
 871                                 &fdata->rxadptr_service_id);
 872                 if (ret)
 873                         rte_exit(EXIT_FAILURE,
 874                                 "Rx adapter service register failed");
 875
 876                 rte_service_runstate_set(fdata->rxadptr_service_id, 1);
 877                 rte_service_component_runstate_set(fdata->rxadptr_service_id,
 878                                 1);
 879                 rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id,
 880                                 0);
 881         } else {
 882                 memset(fdata->rx_core, 0, sizeof(unsigned int) * MAX_NUM_CORE);
 883                 rte_free(adptr_services);
 884         }
 885
 886         if (!adptr_services->nb_rx_adptrs && (dev_info.event_dev_cap &
 887                          RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED))
 888                 fdata->cap.scheduler = NULL;
 889 }
 890
 891 static void
 892 worker_tx_enq_opt_check(void)
 893 {
 894         int i;
 895         int ret;
 896         uint32_t cap = 0;
 897         uint8_t rx_needed = 0;
 898         uint8_t sched_needed = 0;
 899         struct rte_event_dev_info eventdev_info;
 900
 901         memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
 902         rte_event_dev_info_get(0, &eventdev_info);
 903
 904         if (cdata.all_type_queues && !(eventdev_info.event_dev_cap &
 905                                 RTE_EVENT_DEV_CAP_QUEUE_ALL_TYPES))
 906                 rte_exit(EXIT_FAILURE,
 907                                 "Event dev doesn't support all type queues\n");
 908         sched_needed = !(eventdev_info.event_dev_cap &
 909                 RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED);
 910
 911         RTE_ETH_FOREACH_DEV(i) {
 912                 ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap);
 913                 if (ret)
 914                         rte_exit(EXIT_FAILURE,
 915                                 "failed to get event rx adapter capabilities");
 916                 rx_needed |=
 917                         !(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT);
 918         }
 919
 920         if (cdata.worker_lcore_mask == 0 ||
 921                         (rx_needed && cdata.rx_lcore_mask == 0) ||
 922                         (sched_needed && cdata.sched_lcore_mask == 0)) {
 923                 printf("Core part of pipeline was not assigned any cores. "
 924                         "This will stall the pipeline, please check core masks "
 925                         "(use -h for details on setting core masks):\n"
 926                         "\trx: %"PRIu64"\n\tsched: %"PRIu64
 927                         "\n\tworkers: %"PRIu64"\n", cdata.rx_lcore_mask,
 928                         cdata.sched_lcore_mask, cdata.worker_lcore_mask);
 929                 rte_exit(-1, "Fix core masks\n");
 930         }
 931
 932         if (!sched_needed)
 933                 memset(fdata->sched_core, 0,
 934                                 sizeof(unsigned int) * MAX_NUM_CORE);
 935         if (!rx_needed)
 936                 memset(fdata->rx_core, 0,
 937                                 sizeof(unsigned int) * MAX_NUM_CORE);
 938
 939         memset(fdata->tx_core, 0, sizeof(unsigned int) * MAX_NUM_CORE);
 940 }
 941
 942 static worker_loop
 943 get_worker_loop_single_burst(uint8_t atq)
 944 {
 945         if (atq)
 946                 return worker_do_tx_single_burst_atq;
 947
 948         return worker_do_tx_single_burst;
 949 }
 950
 951 static worker_loop
 952 get_worker_loop_single_non_burst(uint8_t atq)
 953 {
 954         if (atq)
 955                 return worker_do_tx_single_atq;
 956
 957         return worker_do_tx_single;
 958 }
 959
 960 static worker_loop
 961 get_worker_loop_burst(uint8_t atq)
 962 {
 963         if (atq)
 964                 return worker_do_tx_burst_atq;
 965
 966         return worker_do_tx_burst;
 967 }
 968
 969 static worker_loop
 970 get_worker_loop_non_burst(uint8_t atq)
 971 {
 972         if (atq)
 973                 return worker_do_tx_atq;
 974
 975         return worker_do_tx;
 976 }
 977
 978 static worker_loop
 979 get_worker_single_stage(bool burst)
 980 {
 981         uint8_t atq = cdata.all_type_queues ? 1 : 0;
 982
 983         if (burst)
 984                 return get_worker_loop_single_burst(atq);
 985
 986         return get_worker_loop_single_non_burst(atq);
 987 }
 988
 989 static worker_loop
 990 get_worker_multi_stage(bool burst)
 991 {
 992         uint8_t atq = cdata.all_type_queues ? 1 : 0;
 993
 994         if (burst)
 995                 return get_worker_loop_burst(atq);
 996
 997         return get_worker_loop_non_burst(atq);
 998 }
 999
1000 void
1001 set_worker_tx_enq_setup_data(struct setup_data *caps, bool burst)
1002 {
1003         if (cdata.num_stages == 1)
1004                 caps->worker = get_worker_single_stage(burst);
1005         else
1006                 caps->worker = get_worker_multi_stage(burst);
1007
1008         caps->check_opt = worker_tx_enq_opt_check;
1009         caps->scheduler = schedule_devices;
1010         caps->evdev_setup = setup_eventdev_worker_tx_enq;
1011         caps->adptr_setup = init_adapters;
1012 }