drivers/event/sw/sw_evdev.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016-2017 Intel Corporation
   3  */
   4
   5 #include <inttypes.h>
   6 #include <string.h>
   7
   8 #include <rte_bus_vdev.h>
   9 #include <rte_kvargs.h>
  10 #include <rte_ring.h>
  11 #include <rte_errno.h>
  12 #include <rte_event_ring.h>
  13 #include <rte_service_component.h>
  14
  15 #include "sw_evdev.h"
  16 #include "iq_chunk.h"
  17 #include "event_ring.h"
  18
  19 #define EVENTDEV_NAME_SW_PMD event_sw
  20 #define NUMA_NODE_ARG "numa_node"
  21 #define SCHED_QUANTA_ARG "sched_quanta"
  22 #define CREDIT_QUANTA_ARG "credit_quanta"
  23 #define MIN_BURST_SIZE_ARG "min_burst"
  24 #define DEQ_BURST_SIZE_ARG "deq_burst"
  25 #define REFIL_ONCE_ARG "refill_once"
  26
  27 static void
  28 sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info);
  29
  30 static int
  31 sw_port_link(struct rte_eventdev *dev, void *port, const uint8_t queues[],
  32                 const uint8_t priorities[], uint16_t num)
  33 {
  34         struct sw_port *p = port;
  35         struct sw_evdev *sw = sw_pmd_priv(dev);
  36         int i;
  37
  38         RTE_SET_USED(priorities);
  39         for (i = 0; i < num; i++) {
  40                 struct sw_qid *q = &sw->qids[queues[i]];
  41                 unsigned int j;
  42
  43                 /* check for qid map overflow */
  44                 if (q->cq_num_mapped_cqs >= RTE_DIM(q->cq_map)) {
  45                         rte_errno = EDQUOT;
  46                         break;
  47                 }
  48
  49                 if (p->is_directed && p->num_qids_mapped > 0) {
  50                         rte_errno = EDQUOT;
  51                         break;
  52                 }
  53
  54                 for (j = 0; j < q->cq_num_mapped_cqs; j++) {
  55                         if (q->cq_map[j] == p->id)
  56                                 break;
  57                 }
  58
  59                 /* check if port is already linked */
  60                 if (j < q->cq_num_mapped_cqs)
  61                         continue;
  62
  63                 if (q->type == SW_SCHED_TYPE_DIRECT) {
  64                         /* check directed qids only map to one port */
  65                         if (p->num_qids_mapped > 0) {
  66                                 rte_errno = EDQUOT;
  67                                 break;
  68                         }
  69                         /* check port only takes a directed flow */
  70                         if (num > 1) {
  71                                 rte_errno = EDQUOT;
  72                                 break;
  73                         }
  74
  75                         p->is_directed = 1;
  76                         p->num_qids_mapped = 1;
  77                 } else if (q->type == RTE_SCHED_TYPE_ORDERED) {
  78                         p->num_ordered_qids++;
  79                         p->num_qids_mapped++;
  80                 } else if (q->type == RTE_SCHED_TYPE_ATOMIC ||
  81                                 q->type == RTE_SCHED_TYPE_PARALLEL) {
  82                         p->num_qids_mapped++;
  83                 }
  84
  85                 q->cq_map[q->cq_num_mapped_cqs] = p->id;
  86                 rte_smp_wmb();
  87                 q->cq_num_mapped_cqs++;
  88         }
  89         return i;
  90 }
  91
  92 static int
  93 sw_port_unlink(struct rte_eventdev *dev, void *port, uint8_t queues[],
  94                 uint16_t nb_unlinks)
  95 {
  96         struct sw_port *p = port;
  97         struct sw_evdev *sw = sw_pmd_priv(dev);
  98         unsigned int i, j;
  99
 100         int unlinked = 0;
 101         for (i = 0; i < nb_unlinks; i++) {
 102                 struct sw_qid *q = &sw->qids[queues[i]];
 103                 for (j = 0; j < q->cq_num_mapped_cqs; j++) {
 104                         if (q->cq_map[j] == p->id) {
 105                                 q->cq_map[j] =
 106                                         q->cq_map[q->cq_num_mapped_cqs - 1];
 107                                 rte_smp_wmb();
 108                                 q->cq_num_mapped_cqs--;
 109                                 unlinked++;
 110
 111                                 p->num_qids_mapped--;
 112
 113                                 if (q->type == RTE_SCHED_TYPE_ORDERED)
 114                                         p->num_ordered_qids--;
 115
 116                                 continue;
 117                         }
 118                 }
 119         }
 120
 121         p->unlinks_in_progress += unlinked;
 122         rte_smp_mb();
 123
 124         return unlinked;
 125 }
 126
 127 static int
 128 sw_port_unlinks_in_progress(struct rte_eventdev *dev, void *port)
 129 {
 130         RTE_SET_USED(dev);
 131         struct sw_port *p = port;
 132         return p->unlinks_in_progress;
 133 }
 134
 135 static int
 136 sw_port_setup(struct rte_eventdev *dev, uint8_t port_id,
 137                 const struct rte_event_port_conf *conf)
 138 {
 139         struct sw_evdev *sw = sw_pmd_priv(dev);
 140         struct sw_port *p = &sw->ports[port_id];
 141         char buf[RTE_RING_NAMESIZE];
 142         unsigned int i;
 143
 144         struct rte_event_dev_info info;
 145         sw_info_get(dev, &info);
 146
 147         /* detect re-configuring and return credits to instance if needed */
 148         if (p->initialized) {
 149                 /* taking credits from pool is done one quanta at a time, and
 150                  * credits may be spend (counted in p->inflights) or still
 151                  * available in the port (p->inflight_credits). We must return
 152                  * the sum to no leak credits
 153                  */
 154                 int possible_inflights = p->inflight_credits + p->inflights;
 155                 rte_atomic32_sub(&sw->inflights, possible_inflights);
 156         }
 157
 158         *p = (struct sw_port){0}; /* zero entire structure */
 159         p->id = port_id;
 160         p->sw = sw;
 161
 162         /* check to see if rings exists - port_setup() can be called multiple
 163          * times legally (assuming device is stopped). If ring exists, free it
 164          * to so it gets re-created with the correct size
 165          */
 166         snprintf(buf, sizeof(buf), "sw%d_p%u_%s", dev->data->dev_id,
 167                         port_id, "rx_worker_ring");
 168         struct rte_event_ring *existing_ring = rte_event_ring_lookup(buf);
 169         if (existing_ring)
 170                 rte_event_ring_free(existing_ring);
 171
 172         p->rx_worker_ring = rte_event_ring_create(buf, MAX_SW_PROD_Q_DEPTH,
 173                         dev->data->socket_id,
 174                         RING_F_SP_ENQ | RING_F_SC_DEQ | RING_F_EXACT_SZ);
 175         if (p->rx_worker_ring == NULL) {
 176                 SW_LOG_ERR("Error creating RX worker ring for port %d\n",
 177                                 port_id);
 178                 return -1;
 179         }
 180
 181         p->inflight_max = conf->new_event_threshold;
 182         p->implicit_release = !(conf->event_port_cfg &
 183                                 RTE_EVENT_PORT_CFG_DISABLE_IMPL_REL);
 184
 185         /* check if ring exists, same as rx_worker above */
 186         snprintf(buf, sizeof(buf), "sw%d_p%u, %s", dev->data->dev_id,
 187                         port_id, "cq_worker_ring");
 188         existing_ring = rte_event_ring_lookup(buf);
 189         if (existing_ring)
 190                 rte_event_ring_free(existing_ring);
 191
 192         p->cq_worker_ring = rte_event_ring_create(buf, conf->dequeue_depth,
 193                         dev->data->socket_id,
 194                         RING_F_SP_ENQ | RING_F_SC_DEQ | RING_F_EXACT_SZ);
 195         if (p->cq_worker_ring == NULL) {
 196                 rte_event_ring_free(p->rx_worker_ring);
 197                 SW_LOG_ERR("Error creating CQ worker ring for port %d\n",
 198                                 port_id);
 199                 return -1;
 200         }
 201         sw->cq_ring_space[port_id] = conf->dequeue_depth;
 202
 203         /* set hist list contents to empty */
 204         for (i = 0; i < SW_PORT_HIST_LIST; i++) {
 205                 p->hist_list[i].fid = -1;
 206                 p->hist_list[i].qid = -1;
 207         }
 208         dev->data->ports[port_id] = p;
 209
 210         rte_smp_wmb();
 211         p->initialized = 1;
 212         return 0;
 213 }
 214
 215 static void
 216 sw_port_release(void *port)
 217 {
 218         struct sw_port *p = (void *)port;
 219         if (p == NULL)
 220                 return;
 221
 222         rte_event_ring_free(p->rx_worker_ring);
 223         rte_event_ring_free(p->cq_worker_ring);
 224         memset(p, 0, sizeof(*p));
 225 }
 226
 227 static int32_t
 228 qid_init(struct sw_evdev *sw, unsigned int idx, int type,
 229                 const struct rte_event_queue_conf *queue_conf)
 230 {
 231         unsigned int i;
 232         int dev_id = sw->data->dev_id;
 233         int socket_id = sw->data->socket_id;
 234         char buf[IQ_ROB_NAMESIZE];
 235         struct sw_qid *qid = &sw->qids[idx];
 236
 237         /* Initialize the FID structures to no pinning (-1), and zero packets */
 238         const struct sw_fid_t fid = {.cq = -1, .pcount = 0};
 239         for (i = 0; i < RTE_DIM(qid->fids); i++)
 240                 qid->fids[i] = fid;
 241
 242         qid->id = idx;
 243         qid->type = type;
 244         qid->priority = queue_conf->priority;
 245
 246         if (qid->type == RTE_SCHED_TYPE_ORDERED) {
 247                 uint32_t window_size;
 248
 249                 /* rte_ring and window_size_mask require require window_size to
 250                  * be a power-of-2.
 251                  */
 252                 window_size = rte_align32pow2(
 253                                 queue_conf->nb_atomic_order_sequences);
 254
 255                 qid->window_size = window_size - 1;
 256
 257                 if (!window_size) {
 258                         SW_LOG_DBG(
 259                                 "invalid reorder_window_size for ordered queue\n"
 260                                 );
 261                         goto cleanup;
 262                 }
 263
 264                 snprintf(buf, sizeof(buf), "sw%d_iq_%d_rob", dev_id, i);
 265                 qid->reorder_buffer = rte_zmalloc_socket(buf,
 266                                 window_size * sizeof(qid->reorder_buffer[0]),
 267                                 0, socket_id);
 268                 if (!qid->reorder_buffer) {
 269                         SW_LOG_DBG("reorder_buffer malloc failed\n");
 270                         goto cleanup;
 271                 }
 272
 273                 memset(&qid->reorder_buffer[0],
 274                        0,
 275                        window_size * sizeof(qid->reorder_buffer[0]));
 276
 277                 qid->reorder_buffer_freelist = rob_ring_create(window_size,
 278                                 socket_id);
 279                 if (!qid->reorder_buffer_freelist) {
 280                         SW_LOG_DBG("freelist ring create failed");
 281                         goto cleanup;
 282                 }
 283
 284                 /* Populate the freelist with reorder buffer entries. Enqueue
 285                  * 'window_size - 1' entries because the rte_ring holds only
 286                  * that many.
 287                  */
 288                 for (i = 0; i < window_size - 1; i++) {
 289                         if (rob_ring_enqueue(qid->reorder_buffer_freelist,
 290                                                 &qid->reorder_buffer[i]) != 1)
 291                                 goto cleanup;
 292                 }
 293
 294                 qid->reorder_buffer_index = 0;
 295                 qid->cq_next_tx = 0;
 296         }
 297
 298         qid->initialized = 1;
 299
 300         return 0;
 301
 302 cleanup:
 303         if (qid->reorder_buffer) {
 304                 rte_free(qid->reorder_buffer);
 305                 qid->reorder_buffer = NULL;
 306         }
 307
 308         if (qid->reorder_buffer_freelist) {
 309                 rob_ring_free(qid->reorder_buffer_freelist);
 310                 qid->reorder_buffer_freelist = NULL;
 311         }
 312
 313         return -EINVAL;
 314 }
 315
 316 static void
 317 sw_queue_release(struct rte_eventdev *dev, uint8_t id)
 318 {
 319         struct sw_evdev *sw = sw_pmd_priv(dev);
 320         struct sw_qid *qid = &sw->qids[id];
 321
 322         if (qid->type == RTE_SCHED_TYPE_ORDERED) {
 323                 rte_free(qid->reorder_buffer);
 324                 rob_ring_free(qid->reorder_buffer_freelist);
 325         }
 326         memset(qid, 0, sizeof(*qid));
 327 }
 328
 329 static int
 330 sw_queue_setup(struct rte_eventdev *dev, uint8_t queue_id,
 331                 const struct rte_event_queue_conf *conf)
 332 {
 333         int type;
 334
 335         type = conf->schedule_type;
 336
 337         if (RTE_EVENT_QUEUE_CFG_SINGLE_LINK & conf->event_queue_cfg) {
 338                 type = SW_SCHED_TYPE_DIRECT;
 339         } else if (RTE_EVENT_QUEUE_CFG_ALL_TYPES
 340                         & conf->event_queue_cfg) {
 341                 SW_LOG_ERR("QUEUE_CFG_ALL_TYPES not supported\n");
 342                 return -ENOTSUP;
 343         }
 344
 345         struct sw_evdev *sw = sw_pmd_priv(dev);
 346
 347         if (sw->qids[queue_id].initialized)
 348                 sw_queue_release(dev, queue_id);
 349
 350         return qid_init(sw, queue_id, type, conf);
 351 }
 352
 353 static void
 354 sw_init_qid_iqs(struct sw_evdev *sw)
 355 {
 356         int i, j;
 357
 358         /* Initialize the IQ memory of all configured qids */
 359         for (i = 0; i < RTE_EVENT_MAX_QUEUES_PER_DEV; i++) {
 360                 struct sw_qid *qid = &sw->qids[i];
 361
 362                 if (!qid->initialized)
 363                         continue;
 364
 365                 for (j = 0; j < SW_IQS_MAX; j++)
 366                         iq_init(sw, &qid->iq[j]);
 367         }
 368 }
 369
 370 static int
 371 sw_qids_empty(struct sw_evdev *sw)
 372 {
 373         unsigned int i, j;
 374
 375         for (i = 0; i < sw->qid_count; i++) {
 376                 for (j = 0; j < SW_IQS_MAX; j++) {
 377                         if (iq_count(&sw->qids[i].iq[j]))
 378                                 return 0;
 379                 }
 380         }
 381
 382         return 1;
 383 }
 384
 385 static int
 386 sw_ports_empty(struct sw_evdev *sw)
 387 {
 388         unsigned int i;
 389
 390         for (i = 0; i < sw->port_count; i++) {
 391                 if ((rte_event_ring_count(sw->ports[i].rx_worker_ring)) ||
 392                      rte_event_ring_count(sw->ports[i].cq_worker_ring))
 393                         return 0;
 394         }
 395
 396         return 1;
 397 }
 398
 399 static void
 400 sw_drain_ports(struct rte_eventdev *dev)
 401 {
 402         struct sw_evdev *sw = sw_pmd_priv(dev);
 403         eventdev_stop_flush_t flush;
 404         unsigned int i;
 405         uint8_t dev_id;
 406         void *arg;
 407
 408         flush = dev->dev_ops->dev_stop_flush;
 409         dev_id = dev->data->dev_id;
 410         arg = dev->data->dev_stop_flush_arg;
 411
 412         for (i = 0; i < sw->port_count; i++) {
 413                 struct rte_event ev;
 414
 415                 while (rte_event_dequeue_burst(dev_id, i, &ev, 1, 0)) {
 416                         if (flush)
 417                                 flush(dev_id, ev, arg);
 418
 419                         ev.op = RTE_EVENT_OP_RELEASE;
 420                         rte_event_enqueue_burst(dev_id, i, &ev, 1);
 421                 }
 422         }
 423 }
 424
 425 static void
 426 sw_drain_queue(struct rte_eventdev *dev, struct sw_iq *iq)
 427 {
 428         struct sw_evdev *sw = sw_pmd_priv(dev);
 429         eventdev_stop_flush_t flush;
 430         uint8_t dev_id;
 431         void *arg;
 432
 433         flush = dev->dev_ops->dev_stop_flush;
 434         dev_id = dev->data->dev_id;
 435         arg = dev->data->dev_stop_flush_arg;
 436
 437         while (iq_count(iq) > 0) {
 438                 struct rte_event ev;
 439
 440                 iq_dequeue_burst(sw, iq, &ev, 1);
 441
 442                 if (flush)
 443                         flush(dev_id, ev, arg);
 444         }
 445 }
 446
 447 static void
 448 sw_drain_queues(struct rte_eventdev *dev)
 449 {
 450         struct sw_evdev *sw = sw_pmd_priv(dev);
 451         unsigned int i, j;
 452
 453         for (i = 0; i < sw->qid_count; i++) {
 454                 for (j = 0; j < SW_IQS_MAX; j++)
 455                         sw_drain_queue(dev, &sw->qids[i].iq[j]);
 456         }
 457 }
 458
 459 static void
 460 sw_clean_qid_iqs(struct rte_eventdev *dev)
 461 {
 462         struct sw_evdev *sw = sw_pmd_priv(dev);
 463         int i, j;
 464
 465         /* Release the IQ memory of all configured qids */
 466         for (i = 0; i < RTE_EVENT_MAX_QUEUES_PER_DEV; i++) {
 467                 struct sw_qid *qid = &sw->qids[i];
 468
 469                 for (j = 0; j < SW_IQS_MAX; j++) {
 470                         if (!qid->iq[j].head)
 471                                 continue;
 472                         iq_free_chunk_list(sw, qid->iq[j].head);
 473                         qid->iq[j].head = NULL;
 474                 }
 475         }
 476 }
 477
 478 static void
 479 sw_queue_def_conf(struct rte_eventdev *dev, uint8_t queue_id,
 480                                  struct rte_event_queue_conf *conf)
 481 {
 482         RTE_SET_USED(dev);
 483         RTE_SET_USED(queue_id);
 484
 485         static const struct rte_event_queue_conf default_conf = {
 486                 .nb_atomic_flows = 4096,
 487                 .nb_atomic_order_sequences = 1,
 488                 .schedule_type = RTE_SCHED_TYPE_ATOMIC,
 489                 .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 490         };
 491
 492         *conf = default_conf;
 493 }
 494
 495 static void
 496 sw_port_def_conf(struct rte_eventdev *dev, uint8_t port_id,
 497                  struct rte_event_port_conf *port_conf)
 498 {
 499         RTE_SET_USED(dev);
 500         RTE_SET_USED(port_id);
 501
 502         port_conf->new_event_threshold = 1024;
 503         port_conf->dequeue_depth = 16;
 504         port_conf->enqueue_depth = 16;
 505         port_conf->event_port_cfg = 0;
 506 }
 507
 508 static int
 509 sw_dev_configure(const struct rte_eventdev *dev)
 510 {
 511         struct sw_evdev *sw = sw_pmd_priv(dev);
 512         const struct rte_eventdev_data *data = dev->data;
 513         const struct rte_event_dev_config *conf = &data->dev_conf;
 514         int num_chunks, i;
 515
 516         sw->qid_count = conf->nb_event_queues;
 517         sw->port_count = conf->nb_event_ports;
 518         sw->nb_events_limit = conf->nb_events_limit;
 519         rte_atomic32_set(&sw->inflights, 0);
 520
 521         /* Number of chunks sized for worst-case spread of events across IQs */
 522         num_chunks = ((SW_INFLIGHT_EVENTS_TOTAL/SW_EVS_PER_Q_CHUNK)+1) +
 523                         sw->qid_count*SW_IQS_MAX*2;
 524
 525         /* If this is a reconfiguration, free the previous IQ allocation. All
 526          * IQ chunk references were cleaned out of the QIDs in sw_stop(), and
 527          * will be reinitialized in sw_start().
 528          */
 529         if (sw->chunks)
 530                 rte_free(sw->chunks);
 531
 532         sw->chunks = rte_malloc_socket(NULL,
 533                                        sizeof(struct sw_queue_chunk) *
 534                                        num_chunks,
 535                                        0,
 536                                        sw->data->socket_id);
 537         if (!sw->chunks)
 538                 return -ENOMEM;
 539
 540         sw->chunk_list_head = NULL;
 541         for (i = 0; i < num_chunks; i++)
 542                 iq_free_chunk(sw, &sw->chunks[i]);
 543
 544         if (conf->event_dev_cfg & RTE_EVENT_DEV_CFG_PER_DEQUEUE_TIMEOUT)
 545                 return -ENOTSUP;
 546
 547         return 0;
 548 }
 549
 550 struct rte_eth_dev;
 551
 552 static int
 553 sw_eth_rx_adapter_caps_get(const struct rte_eventdev *dev,
 554                         const struct rte_eth_dev *eth_dev,
 555                         uint32_t *caps)
 556 {
 557         RTE_SET_USED(dev);
 558         RTE_SET_USED(eth_dev);
 559         *caps = RTE_EVENT_ETH_RX_ADAPTER_SW_CAP;
 560         return 0;
 561 }
 562
 563 static int
 564 sw_timer_adapter_caps_get(const struct rte_eventdev *dev,
 565                           uint64_t flags,
 566                           uint32_t *caps,
 567                           const struct rte_event_timer_adapter_ops **ops)
 568 {
 569         RTE_SET_USED(dev);
 570         RTE_SET_USED(flags);
 571         *caps = 0;
 572
 573         /* Use default SW ops */
 574         *ops = NULL;
 575
 576         return 0;
 577 }
 578
 579 static int
 580 sw_crypto_adapter_caps_get(const struct rte_eventdev *dev,
 581                            const struct rte_cryptodev *cdev,
 582                            uint32_t *caps)
 583 {
 584         RTE_SET_USED(dev);
 585         RTE_SET_USED(cdev);
 586         *caps = RTE_EVENT_CRYPTO_ADAPTER_SW_CAP;
 587         return 0;
 588 }
 589
 590 static void
 591 sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info)
 592 {
 593         RTE_SET_USED(dev);
 594
 595         static const struct rte_event_dev_info evdev_sw_info = {
 596                         .driver_name = SW_PMD_NAME,
 597                         .max_event_queues = RTE_EVENT_MAX_QUEUES_PER_DEV,
 598                         .max_event_queue_flows = SW_QID_NUM_FIDS,
 599                         .max_event_queue_priority_levels = SW_Q_PRIORITY_MAX,
 600                         .max_event_priority_levels = SW_IQS_MAX,
 601                         .max_event_ports = SW_PORTS_MAX,
 602                         .max_event_port_dequeue_depth = MAX_SW_CONS_Q_DEPTH,
 603                         .max_event_port_enqueue_depth = MAX_SW_PROD_Q_DEPTH,
 604                         .max_num_events = SW_INFLIGHT_EVENTS_TOTAL,
 605                         .event_dev_cap = (
 606                                 RTE_EVENT_DEV_CAP_QUEUE_QOS |
 607                                 RTE_EVENT_DEV_CAP_BURST_MODE |
 608                                 RTE_EVENT_DEV_CAP_EVENT_QOS |
 609                                 RTE_EVENT_DEV_CAP_IMPLICIT_RELEASE_DISABLE|
 610                                 RTE_EVENT_DEV_CAP_RUNTIME_PORT_LINK |
 611                                 RTE_EVENT_DEV_CAP_MULTIPLE_QUEUE_PORT |
 612                                 RTE_EVENT_DEV_CAP_NONSEQ_MODE |
 613                                 RTE_EVENT_DEV_CAP_CARRY_FLOW_ID),
 614         };
 615
 616         *info = evdev_sw_info;
 617 }
 618
 619 static void
 620 sw_dump(struct rte_eventdev *dev, FILE *f)
 621 {
 622         const struct sw_evdev *sw = sw_pmd_priv(dev);
 623
 624         static const char * const q_type_strings[] = {
 625                         "Ordered", "Atomic", "Parallel", "Directed"
 626         };
 627         uint32_t i;
 628         fprintf(f, "EventDev %s: ports %d, qids %d\n", "todo-fix-name",
 629                         sw->port_count, sw->qid_count);
 630
 631         fprintf(f, "\trx   %"PRIu64"\n\tdrop %"PRIu64"\n\ttx   %"PRIu64"\n",
 632                 sw->stats.rx_pkts, sw->stats.rx_dropped, sw->stats.tx_pkts);
 633         fprintf(f, "\tsched calls: %"PRIu64"\n", sw->sched_called);
 634         fprintf(f, "\tsched cq/qid call: %"PRIu64"\n", sw->sched_cq_qid_called);
 635         fprintf(f, "\tsched no IQ enq: %"PRIu64"\n", sw->sched_no_iq_enqueues);
 636         fprintf(f, "\tsched no CQ enq: %"PRIu64"\n", sw->sched_no_cq_enqueues);
 637         uint32_t inflights = rte_atomic32_read(&sw->inflights);
 638         uint32_t credits = sw->nb_events_limit - inflights;
 639         fprintf(f, "\tinflight %d, credits: %d\n", inflights, credits);
 640
 641 #define COL_RED "\x1b[31m"
 642 #define COL_RESET "\x1b[0m"
 643
 644         for (i = 0; i < sw->port_count; i++) {
 645                 int max, j;
 646                 const struct sw_port *p = &sw->ports[i];
 647                 if (!p->initialized) {
 648                         fprintf(f, "  %sPort %d not initialized.%s\n",
 649                                 COL_RED, i, COL_RESET);
 650                         continue;
 651                 }
 652                 fprintf(f, "  Port %d %s\n", i,
 653                         p->is_directed ? " (SingleCons)" : "");
 654                 fprintf(f, "\trx   %"PRIu64"\tdrop %"PRIu64"\ttx   %"PRIu64
 655                         "\t%sinflight %d%s\n", sw->ports[i].stats.rx_pkts,
 656                         sw->ports[i].stats.rx_dropped,
 657                         sw->ports[i].stats.tx_pkts,
 658                         (p->inflights == p->inflight_max) ?
 659                                 COL_RED : COL_RESET,
 660                         sw->ports[i].inflights, COL_RESET);
 661
 662                 fprintf(f, "\tMax New: %u"
 663                         "\tAvg cycles PP: %"PRIu64"\tCredits: %u\n",
 664                         sw->ports[i].inflight_max,
 665                         sw->ports[i].avg_pkt_ticks,
 666                         sw->ports[i].inflight_credits);
 667                 fprintf(f, "\tReceive burst distribution:\n");
 668                 float zp_percent = p->zero_polls * 100.0 / p->total_polls;
 669                 fprintf(f, zp_percent < 10 ? "\t\t0:%.02f%% " : "\t\t0:%.0f%% ",
 670                                 zp_percent);
 671                 for (max = (int)RTE_DIM(p->poll_buckets); max-- > 0;)
 672                         if (p->poll_buckets[max] != 0)
 673                                 break;
 674                 for (j = 0; j <= max; j++) {
 675                         if (p->poll_buckets[j] != 0) {
 676                                 float poll_pc = p->poll_buckets[j] * 100.0 /
 677                                         p->total_polls;
 678                                 fprintf(f, "%u-%u:%.02f%% ",
 679                                         ((j << SW_DEQ_STAT_BUCKET_SHIFT) + 1),
 680                                         ((j+1) << SW_DEQ_STAT_BUCKET_SHIFT),
 681                                         poll_pc);
 682                         }
 683                 }
 684                 fprintf(f, "\n");
 685
 686                 if (p->rx_worker_ring) {
 687                         uint64_t used = rte_event_ring_count(p->rx_worker_ring);
 688                         uint64_t space = rte_event_ring_free_count(
 689                                         p->rx_worker_ring);
 690                         const char *col = (space == 0) ? COL_RED : COL_RESET;
 691                         fprintf(f, "\t%srx ring used: %4"PRIu64"\tfree: %4"
 692                                         PRIu64 COL_RESET"\n", col, used, space);
 693                 } else
 694                         fprintf(f, "\trx ring not initialized.\n");
 695
 696                 if (p->cq_worker_ring) {
 697                         uint64_t used = rte_event_ring_count(p->cq_worker_ring);
 698                         uint64_t space = rte_event_ring_free_count(
 699                                         p->cq_worker_ring);
 700                         const char *col = (space == 0) ? COL_RED : COL_RESET;
 701                         fprintf(f, "\t%scq ring used: %4"PRIu64"\tfree: %4"
 702                                         PRIu64 COL_RESET"\n", col, used, space);
 703                 } else
 704                         fprintf(f, "\tcq ring not initialized.\n");
 705         }
 706
 707         for (i = 0; i < sw->qid_count; i++) {
 708                 const struct sw_qid *qid = &sw->qids[i];
 709                 if (!qid->initialized) {
 710                         fprintf(f, "  %sQueue %d not initialized.%s\n",
 711                                 COL_RED, i, COL_RESET);
 712                         continue;
 713                 }
 714                 int affinities_per_port[SW_PORTS_MAX] = {0};
 715                 uint32_t inflights = 0;
 716
 717                 fprintf(f, "  Queue %d (%s)\n", i, q_type_strings[qid->type]);
 718                 fprintf(f, "\trx   %"PRIu64"\tdrop %"PRIu64"\ttx   %"PRIu64"\n",
 719                         qid->stats.rx_pkts, qid->stats.rx_dropped,
 720                         qid->stats.tx_pkts);
 721                 if (qid->type == RTE_SCHED_TYPE_ORDERED) {
 722                         struct rob_ring *rob_buf_free =
 723                                 qid->reorder_buffer_freelist;
 724                         if (rob_buf_free)
 725                                 fprintf(f, "\tReorder entries in use: %u\n",
 726                                         rob_ring_free_count(rob_buf_free));
 727                         else
 728                                 fprintf(f,
 729                                         "\tReorder buffer not initialized\n");
 730                 }
 731
 732                 uint32_t flow;
 733                 for (flow = 0; flow < RTE_DIM(qid->fids); flow++)
 734                         if (qid->fids[flow].cq != -1) {
 735                                 affinities_per_port[qid->fids[flow].cq]++;
 736                                 inflights += qid->fids[flow].pcount;
 737                         }
 738
 739                 uint32_t port;
 740                 fprintf(f, "\tPer Port Stats:\n");
 741                 for (port = 0; port < sw->port_count; port++) {
 742                         fprintf(f, "\t  Port %d: Pkts: %"PRIu64, port,
 743                                         qid->to_port[port]);
 744                         fprintf(f, "\tFlows: %d\n", affinities_per_port[port]);
 745                 }
 746
 747                 uint32_t iq;
 748                 uint32_t iq_printed = 0;
 749                 for (iq = 0; iq < SW_IQS_MAX; iq++) {
 750                         if (!qid->iq[iq].head) {
 751                                 fprintf(f, "\tiq %d is not initialized.\n", iq);
 752                                 iq_printed = 1;
 753                                 continue;
 754                         }
 755                         uint32_t used = iq_count(&qid->iq[iq]);
 756                         const char *col = COL_RESET;
 757                         if (used > 0) {
 758                                 fprintf(f, "\t%siq %d: Used %d"
 759                                         COL_RESET"\n", col, iq, used);
 760                                 iq_printed = 1;
 761                         }
 762                 }
 763                 if (iq_printed == 0)
 764                         fprintf(f, "\t-- iqs empty --\n");
 765         }
 766 }
 767
 768 static int
 769 sw_start(struct rte_eventdev *dev)
 770 {
 771         unsigned int i, j;
 772         struct sw_evdev *sw = sw_pmd_priv(dev);
 773
 774         rte_service_component_runstate_set(sw->service_id, 1);
 775
 776         /* check a service core is mapped to this service */
 777         if (!rte_service_runstate_get(sw->service_id)) {
 778                 SW_LOG_ERR("Warning: No Service core enabled on service %s\n",
 779                                 sw->service_name);
 780                 return -ENOENT;
 781         }
 782
 783         /* check all ports are set up */
 784         for (i = 0; i < sw->port_count; i++)
 785                 if (sw->ports[i].rx_worker_ring == NULL) {
 786                         SW_LOG_ERR("Port %d not configured\n", i);
 787                         return -ESTALE;
 788                 }
 789
 790         /* check all queues are configured and mapped to ports*/
 791         for (i = 0; i < sw->qid_count; i++)
 792                 if (!sw->qids[i].initialized ||
 793                     sw->qids[i].cq_num_mapped_cqs == 0) {
 794                         SW_LOG_ERR("Queue %d not configured\n", i);
 795                         return -ENOLINK;
 796                 }
 797
 798         /* build up our prioritized array of qids */
 799         /* We don't use qsort here, as if all/multiple entries have the same
 800          * priority, the result is non-deterministic. From "man 3 qsort":
 801          * "If two members compare as equal, their order in the sorted
 802          * array is undefined."
 803          */
 804         uint32_t qidx = 0;
 805         for (j = 0; j <= RTE_EVENT_DEV_PRIORITY_LOWEST; j++) {
 806                 for (i = 0; i < sw->qid_count; i++) {
 807                         if (sw->qids[i].priority == j) {
 808                                 sw->qids_prioritized[qidx] = &sw->qids[i];
 809                                 qidx++;
 810                         }
 811                 }
 812         }
 813
 814         sw_init_qid_iqs(sw);
 815
 816         if (sw_xstats_init(sw) < 0)
 817                 return -EINVAL;
 818
 819         rte_smp_wmb();
 820         sw->started = 1;
 821
 822         return 0;
 823 }
 824
 825 static void
 826 sw_stop(struct rte_eventdev *dev)
 827 {
 828         struct sw_evdev *sw = sw_pmd_priv(dev);
 829         int32_t runstate;
 830
 831         /* Stop the scheduler if it's running */
 832         runstate = rte_service_runstate_get(sw->service_id);
 833         if (runstate == 1)
 834                 rte_service_runstate_set(sw->service_id, 0);
 835
 836         while (rte_service_may_be_active(sw->service_id))
 837                 rte_pause();
 838
 839         /* Flush all events out of the device */
 840         while (!(sw_qids_empty(sw) && sw_ports_empty(sw))) {
 841                 sw_event_schedule(dev);
 842                 sw_drain_ports(dev);
 843                 sw_drain_queues(dev);
 844         }
 845
 846         sw_clean_qid_iqs(dev);
 847         sw_xstats_uninit(sw);
 848         sw->started = 0;
 849         rte_smp_wmb();
 850
 851         if (runstate == 1)
 852                 rte_service_runstate_set(sw->service_id, 1);
 853 }
 854
 855 static int
 856 sw_close(struct rte_eventdev *dev)
 857 {
 858         struct sw_evdev *sw = sw_pmd_priv(dev);
 859         uint32_t i;
 860
 861         for (i = 0; i < sw->qid_count; i++)
 862                 sw_queue_release(dev, i);
 863         sw->qid_count = 0;
 864
 865         for (i = 0; i < sw->port_count; i++)
 866                 sw_port_release(&sw->ports[i]);
 867         sw->port_count = 0;
 868
 869         memset(&sw->stats, 0, sizeof(sw->stats));
 870         sw->sched_called = 0;
 871         sw->sched_no_iq_enqueues = 0;
 872         sw->sched_no_cq_enqueues = 0;
 873         sw->sched_cq_qid_called = 0;
 874
 875         return 0;
 876 }
 877
 878 static int
 879 assign_numa_node(const char *key __rte_unused, const char *value, void *opaque)
 880 {
 881         int *socket_id = opaque;
 882         *socket_id = atoi(value);
 883         if (*socket_id >= RTE_MAX_NUMA_NODES)
 884                 return -1;
 885         return 0;
 886 }
 887
 888 static int
 889 set_sched_quanta(const char *key __rte_unused, const char *value, void *opaque)
 890 {
 891         int *quanta = opaque;
 892         *quanta = atoi(value);
 893         if (*quanta < 0 || *quanta >= 4096)
 894                 return -1;
 895         return 0;
 896 }
 897
 898 static int
 899 set_credit_quanta(const char *key __rte_unused, const char *value, void *opaque)
 900 {
 901         int *credit = opaque;
 902         *credit = atoi(value);
 903         if (*credit < 0 || *credit >= 128)
 904                 return -1;
 905         return 0;
 906 }
 907
 908 static int
 909 set_deq_burst_sz(const char *key __rte_unused, const char *value, void *opaque)
 910 {
 911         int *deq_burst_sz = opaque;
 912         *deq_burst_sz = atoi(value);
 913         if (*deq_burst_sz < 0 || *deq_burst_sz > SCHED_DEQUEUE_MAX_BURST_SIZE)
 914                 return -1;
 915         return 0;
 916 }
 917
 918 static int
 919 set_min_burst_sz(const char *key __rte_unused, const char *value, void *opaque)
 920 {
 921         int *min_burst_sz = opaque;
 922         *min_burst_sz = atoi(value);
 923         if (*min_burst_sz < 0 || *min_burst_sz > SCHED_DEQUEUE_MAX_BURST_SIZE)
 924                 return -1;
 925         return 0;
 926 }
 927
 928 static int
 929 set_refill_once(const char *key __rte_unused, const char *value, void *opaque)
 930 {
 931         int *refill_once_per_call = opaque;
 932         *refill_once_per_call = atoi(value);
 933         if (*refill_once_per_call < 0 || *refill_once_per_call > 1)
 934                 return -1;
 935         return 0;
 936 }
 937
 938 static int32_t sw_sched_service_func(void *args)
 939 {
 940         struct rte_eventdev *dev = args;
 941         sw_event_schedule(dev);
 942         return 0;
 943 }
 944
 945 static int
 946 sw_probe(struct rte_vdev_device *vdev)
 947 {
 948         static struct rte_eventdev_ops evdev_sw_ops = {
 949                         .dev_configure = sw_dev_configure,
 950                         .dev_infos_get = sw_info_get,
 951                         .dev_close = sw_close,
 952                         .dev_start = sw_start,
 953                         .dev_stop = sw_stop,
 954                         .dump = sw_dump,
 955
 956                         .queue_def_conf = sw_queue_def_conf,
 957                         .queue_setup = sw_queue_setup,
 958                         .queue_release = sw_queue_release,
 959                         .port_def_conf = sw_port_def_conf,
 960                         .port_setup = sw_port_setup,
 961                         .port_release = sw_port_release,
 962                         .port_link = sw_port_link,
 963                         .port_unlink = sw_port_unlink,
 964                         .port_unlinks_in_progress = sw_port_unlinks_in_progress,
 965
 966                         .eth_rx_adapter_caps_get = sw_eth_rx_adapter_caps_get,
 967
 968                         .timer_adapter_caps_get = sw_timer_adapter_caps_get,
 969
 970                         .crypto_adapter_caps_get = sw_crypto_adapter_caps_get,
 971
 972                         .xstats_get = sw_xstats_get,
 973                         .xstats_get_names = sw_xstats_get_names,
 974                         .xstats_get_by_name = sw_xstats_get_by_name,
 975                         .xstats_reset = sw_xstats_reset,
 976
 977                         .dev_selftest = test_sw_eventdev,
 978         };
 979
 980         static const char *const args[] = {
 981                 NUMA_NODE_ARG,
 982                 SCHED_QUANTA_ARG,
 983                 CREDIT_QUANTA_ARG,
 984                 MIN_BURST_SIZE_ARG,
 985                 DEQ_BURST_SIZE_ARG,
 986                 REFIL_ONCE_ARG,
 987                 NULL
 988         };
 989         const char *name;
 990         const char *params;
 991         struct rte_eventdev *dev;
 992         struct sw_evdev *sw;
 993         int socket_id = rte_socket_id();
 994         int sched_quanta  = SW_DEFAULT_SCHED_QUANTA;
 995         int credit_quanta = SW_DEFAULT_CREDIT_QUANTA;
 996         int min_burst_size = 1;
 997         int deq_burst_size = SCHED_DEQUEUE_DEFAULT_BURST_SIZE;
 998         int refill_once = 0;
 999
1000         name = rte_vdev_device_name(vdev);
1001         params = rte_vdev_device_args(vdev);
1002         if (params != NULL && params[0] != '\0') {
1003                 struct rte_kvargs *kvlist = rte_kvargs_parse(params, args);
1004
1005                 if (!kvlist) {
1006                         SW_LOG_INFO(
1007                                 "Ignoring unsupported parameters when creating device '%s'\n",
1008                                 name);
1009                 } else {
1010                         int ret = rte_kvargs_process(kvlist, NUMA_NODE_ARG,
1011                                         assign_numa_node, &socket_id);
1012                         if (ret != 0) {
1013                                 SW_LOG_ERR(
1014                                         "%s: Error parsing numa node parameter",
1015                                         name);
1016                                 rte_kvargs_free(kvlist);
1017                                 return ret;
1018                         }
1019
1020                         ret = rte_kvargs_process(kvlist, SCHED_QUANTA_ARG,
1021                                         set_sched_quanta, &sched_quanta);
1022                         if (ret != 0) {
1023                                 SW_LOG_ERR(
1024                                         "%s: Error parsing sched quanta parameter",
1025                                         name);
1026                                 rte_kvargs_free(kvlist);
1027                                 return ret;
1028                         }
1029
1030                         ret = rte_kvargs_process(kvlist, CREDIT_QUANTA_ARG,
1031                                         set_credit_quanta, &credit_quanta);
1032                         if (ret != 0) {
1033                                 SW_LOG_ERR(
1034                                         "%s: Error parsing credit quanta parameter",
1035                                         name);
1036                                 rte_kvargs_free(kvlist);
1037                                 return ret;
1038                         }
1039
1040                         ret = rte_kvargs_process(kvlist, MIN_BURST_SIZE_ARG,
1041                                         set_min_burst_sz, &min_burst_size);
1042                         if (ret != 0) {
1043                                 SW_LOG_ERR(
1044                                         "%s: Error parsing minimum burst size parameter",
1045                                         name);
1046                                 rte_kvargs_free(kvlist);
1047                                 return ret;
1048                         }
1049
1050                         ret = rte_kvargs_process(kvlist, DEQ_BURST_SIZE_ARG,
1051                                         set_deq_burst_sz, &deq_burst_size);
1052                         if (ret != 0) {
1053                                 SW_LOG_ERR(
1054                                         "%s: Error parsing dequeue burst size parameter",
1055                                         name);
1056                                 rte_kvargs_free(kvlist);
1057                                 return ret;
1058                         }
1059
1060                         ret = rte_kvargs_process(kvlist, REFIL_ONCE_ARG,
1061                                         set_refill_once, &refill_once);
1062                         if (ret != 0) {
1063                                 SW_LOG_ERR(
1064                                         "%s: Error parsing refill once per call switch",
1065                                         name);
1066                                 rte_kvargs_free(kvlist);
1067                                 return ret;
1068                         }
1069
1070                         rte_kvargs_free(kvlist);
1071                 }
1072         }
1073
1074         SW_LOG_INFO(
1075                         "Creating eventdev sw device %s, numa_node=%d, "
1076                         "sched_quanta=%d, credit_quanta=%d "
1077                         "min_burst=%d, deq_burst=%d, refill_once=%d\n",
1078                         name, socket_id, sched_quanta, credit_quanta,
1079                         min_burst_size, deq_burst_size, refill_once);
1080
1081         dev = rte_event_pmd_vdev_init(name,
1082                         sizeof(struct sw_evdev), socket_id);
1083         if (dev == NULL) {
1084                 SW_LOG_ERR("eventdev vdev init() failed");
1085                 return -EFAULT;
1086         }
1087         dev->dev_ops = &evdev_sw_ops;
1088         dev->enqueue = sw_event_enqueue;
1089         dev->enqueue_burst = sw_event_enqueue_burst;
1090         dev->enqueue_new_burst = sw_event_enqueue_burst;
1091         dev->enqueue_forward_burst = sw_event_enqueue_burst;
1092         dev->dequeue = sw_event_dequeue;
1093         dev->dequeue_burst = sw_event_dequeue_burst;
1094
1095         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1096                 return 0;
1097
1098         sw = dev->data->dev_private;
1099         sw->data = dev->data;
1100
1101         /* copy values passed from vdev command line to instance */
1102         sw->credit_update_quanta = credit_quanta;
1103         sw->sched_quanta = sched_quanta;
1104         sw->sched_min_burst_size = min_burst_size;
1105         sw->sched_deq_burst_size = deq_burst_size;
1106         sw->refill_once_per_iter = refill_once;
1107
1108         /* register service with EAL */
1109         struct rte_service_spec service;
1110         memset(&service, 0, sizeof(struct rte_service_spec));
1111         snprintf(service.name, sizeof(service.name), "%s_service", name);
1112         snprintf(sw->service_name, sizeof(sw->service_name), "%s_service",
1113                         name);
1114         service.socket_id = socket_id;
1115         service.callback = sw_sched_service_func;
1116         service.callback_userdata = (void *)dev;
1117
1118         int32_t ret = rte_service_component_register(&service, &sw->service_id);
1119         if (ret) {
1120                 SW_LOG_ERR("service register() failed");
1121                 return -ENOEXEC;
1122         }
1123
1124         dev->data->service_inited = 1;
1125         dev->data->service_id = sw->service_id;
1126
1127         return 0;
1128 }
1129
1130 static int
1131 sw_remove(struct rte_vdev_device *vdev)
1132 {
1133         const char *name;
1134
1135         name = rte_vdev_device_name(vdev);
1136         if (name == NULL)
1137                 return -EINVAL;
1138
1139         SW_LOG_INFO("Closing eventdev sw device %s\n", name);
1140
1141         return rte_event_pmd_vdev_uninit(name);
1142 }
1143
1144 static struct rte_vdev_driver evdev_sw_pmd_drv = {
1145         .probe = sw_probe,
1146         .remove = sw_remove
1147 };
1148
1149 RTE_PMD_REGISTER_VDEV(EVENTDEV_NAME_SW_PMD, evdev_sw_pmd_drv);
1150 RTE_PMD_REGISTER_PARAM_STRING(event_sw, NUMA_NODE_ARG "=<int> "
1151                 SCHED_QUANTA_ARG "=<int>" CREDIT_QUANTA_ARG "=<int>"
1152                 MIN_BURST_SIZE_ARG "=<int>" DEQ_BURST_SIZE_ARG "=<int>"
1153                 REFIL_ONCE_ARG "=<int>");
1154 RTE_LOG_REGISTER(eventdev_sw_log_level, pmd.event.sw, NOTICE);