drivers/event/sw/sw_evdev.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2016-2017 Intel Corporation. All rights reserved.
   5  *
   6  *   Redistribution and use in source and binary forms, with or without
   7  *   modification, are permitted provided that the following conditions
   8  *   are met:
   9  *
  10  *     * Redistributions of source code must retain the above copyright
  11  *       notice, this list of conditions and the following disclaimer.
  12  *     * Redistributions in binary form must reproduce the above copyright
  13  *       notice, this list of conditions and the following disclaimer in
  14  *       the documentation and/or other materials provided with the
  15  *       distribution.
  16  *     * Neither the name of Intel Corporation nor the names of its
  17  *       contributors may be used to endorse or promote products derived
  18  *       from this software without specific prior written permission.
  19  *
  20  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  24  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  25  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  26  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  27  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  28  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  29  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  30  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  31  */
  32
  33 #include <inttypes.h>
  34 #include <string.h>
  35
  36 #include <rte_vdev.h>
  37 #include <rte_memzone.h>
  38 #include <rte_kvargs.h>
  39 #include <rte_ring.h>
  40 #include <rte_errno.h>
  41 #include <rte_event_ring.h>
  42
  43 #include "sw_evdev.h"
  44 #include "iq_ring.h"
  45
  46 #define EVENTDEV_NAME_SW_PMD event_sw
  47 #define NUMA_NODE_ARG "numa_node"
  48 #define SCHED_QUANTA_ARG "sched_quanta"
  49 #define CREDIT_QUANTA_ARG "credit_quanta"
  50
  51 static void
  52 sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info);
  53
  54 static int
  55 sw_port_link(struct rte_eventdev *dev, void *port, const uint8_t queues[],
  56                 const uint8_t priorities[], uint16_t num)
  57 {
  58         struct sw_port *p = port;
  59         struct sw_evdev *sw = sw_pmd_priv(dev);
  60         int i;
  61
  62         RTE_SET_USED(priorities);
  63         for (i = 0; i < num; i++) {
  64                 struct sw_qid *q = &sw->qids[queues[i]];
  65
  66                 /* check for qid map overflow */
  67                 if (q->cq_num_mapped_cqs >= RTE_DIM(q->cq_map)) {
  68                         rte_errno = -EDQUOT;
  69                         break;
  70                 }
  71
  72                 if (p->is_directed && p->num_qids_mapped > 0) {
  73                         rte_errno = -EDQUOT;
  74                         break;
  75                 }
  76
  77                 if (q->type == SW_SCHED_TYPE_DIRECT) {
  78                         /* check directed qids only map to one port */
  79                         if (p->num_qids_mapped > 0) {
  80                                 rte_errno = -EDQUOT;
  81                                 break;
  82                         }
  83                         /* check port only takes a directed flow */
  84                         if (num > 1) {
  85                                 rte_errno = -EDQUOT;
  86                                 break;
  87                         }
  88
  89                         p->is_directed = 1;
  90                         p->num_qids_mapped = 1;
  91                 } else if (q->type == RTE_SCHED_TYPE_ORDERED) {
  92                         p->num_ordered_qids++;
  93                         p->num_qids_mapped++;
  94                 } else if (q->type == RTE_SCHED_TYPE_ATOMIC ||
  95                                 q->type == RTE_SCHED_TYPE_PARALLEL) {
  96                         p->num_qids_mapped++;
  97                 }
  98
  99                 q->cq_map[q->cq_num_mapped_cqs] = p->id;
 100                 rte_smp_wmb();
 101                 q->cq_num_mapped_cqs++;
 102         }
 103         return i;
 104 }
 105
 106 static int
 107 sw_port_unlink(struct rte_eventdev *dev, void *port, uint8_t queues[],
 108                 uint16_t nb_unlinks)
 109 {
 110         struct sw_port *p = port;
 111         struct sw_evdev *sw = sw_pmd_priv(dev);
 112         unsigned int i, j;
 113
 114         int unlinked = 0;
 115         for (i = 0; i < nb_unlinks; i++) {
 116                 struct sw_qid *q = &sw->qids[queues[i]];
 117                 for (j = 0; j < q->cq_num_mapped_cqs; j++) {
 118                         if (q->cq_map[j] == p->id) {
 119                                 q->cq_map[j] =
 120                                         q->cq_map[q->cq_num_mapped_cqs - 1];
 121                                 rte_smp_wmb();
 122                                 q->cq_num_mapped_cqs--;
 123                                 unlinked++;
 124
 125                                 p->num_qids_mapped--;
 126
 127                                 if (q->type == RTE_SCHED_TYPE_ORDERED)
 128                                         p->num_ordered_qids--;
 129
 130                                 continue;
 131                         }
 132                 }
 133         }
 134         return unlinked;
 135 }
 136
 137 static int
 138 sw_port_setup(struct rte_eventdev *dev, uint8_t port_id,
 139                 const struct rte_event_port_conf *conf)
 140 {
 141         struct sw_evdev *sw = sw_pmd_priv(dev);
 142         struct sw_port *p = &sw->ports[port_id];
 143         char buf[RTE_RING_NAMESIZE];
 144         unsigned int i;
 145
 146         struct rte_event_dev_info info;
 147         sw_info_get(dev, &info);
 148
 149         /* detect re-configuring and return credits to instance if needed */
 150         if (p->initialized) {
 151                 /* taking credits from pool is done one quanta at a time, and
 152                  * credits may be spend (counted in p->inflights) or still
 153                  * available in the port (p->inflight_credits). We must return
 154                  * the sum to no leak credits
 155                  */
 156                 int possible_inflights = p->inflight_credits + p->inflights;
 157                 rte_atomic32_sub(&sw->inflights, possible_inflights);
 158         }
 159
 160         *p = (struct sw_port){0}; /* zero entire structure */
 161         p->id = port_id;
 162         p->sw = sw;
 163
 164         /* check to see if rings exists - port_setup() can be called multiple
 165          * times legally (assuming device is stopped). If ring exists, free it
 166          * to so it gets re-created with the correct size
 167          */
 168         snprintf(buf, sizeof(buf), "sw%d_p%u_%s", dev->data->dev_id,
 169                         port_id, "rx_worker_ring");
 170         struct rte_event_ring *existing_ring = rte_event_ring_lookup(buf);
 171         if (existing_ring)
 172                 rte_event_ring_free(existing_ring);
 173
 174         p->rx_worker_ring = rte_event_ring_create(buf, MAX_SW_PROD_Q_DEPTH,
 175                         dev->data->socket_id,
 176                         RING_F_SP_ENQ | RING_F_SC_DEQ | RING_F_EXACT_SZ);
 177         if (p->rx_worker_ring == NULL) {
 178                 SW_LOG_ERR("Error creating RX worker ring for port %d\n",
 179                                 port_id);
 180                 return -1;
 181         }
 182
 183         p->inflight_max = conf->new_event_threshold;
 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_RING_NAMESIZE];
 235         struct sw_qid *qid = &sw->qids[idx];
 236
 237         for (i = 0; i < SW_IQS_MAX; i++) {
 238                 snprintf(buf, sizeof(buf), "q_%u_iq_%d", idx, i);
 239                 qid->iq[i] = iq_ring_create(buf, socket_id);
 240                 if (!qid->iq[i]) {
 241                         SW_LOG_DBG("ring create failed");
 242                         goto cleanup;
 243                 }
 244         }
 245
 246         /* Initialize the FID structures to no pinning (-1), and zero packets */
 247         const struct sw_fid_t fid = {.cq = -1, .pcount = 0};
 248         for (i = 0; i < RTE_DIM(qid->fids); i++)
 249                 qid->fids[i] = fid;
 250
 251         qid->id = idx;
 252         qid->type = type;
 253         qid->priority = queue_conf->priority;
 254
 255         if (qid->type == RTE_SCHED_TYPE_ORDERED) {
 256                 char ring_name[RTE_RING_NAMESIZE];
 257                 uint32_t window_size;
 258
 259                 /* rte_ring and window_size_mask require require window_size to
 260                  * be a power-of-2.
 261                  */
 262                 window_size = rte_align32pow2(
 263                                 queue_conf->nb_atomic_order_sequences);
 264
 265                 qid->window_size = window_size - 1;
 266
 267                 if (!window_size) {
 268                         SW_LOG_DBG(
 269                                 "invalid reorder_window_size for ordered queue\n"
 270                                 );
 271                         goto cleanup;
 272                 }
 273
 274                 snprintf(buf, sizeof(buf), "sw%d_iq_%d_rob", dev_id, i);
 275                 qid->reorder_buffer = rte_zmalloc_socket(buf,
 276                                 window_size * sizeof(qid->reorder_buffer[0]),
 277                                 0, socket_id);
 278                 if (!qid->reorder_buffer) {
 279                         SW_LOG_DBG("reorder_buffer malloc failed\n");
 280                         goto cleanup;
 281                 }
 282
 283                 memset(&qid->reorder_buffer[0],
 284                        0,
 285                        window_size * sizeof(qid->reorder_buffer[0]));
 286
 287                 snprintf(ring_name, sizeof(ring_name), "sw%d_q%d_freelist",
 288                                 dev_id, idx);
 289
 290                 /* lookup the ring, and if it already exists, free it */
 291                 struct rte_ring *cleanup = rte_ring_lookup(ring_name);
 292                 if (cleanup)
 293                         rte_ring_free(cleanup);
 294
 295                 qid->reorder_buffer_freelist = rte_ring_create(ring_name,
 296                                 window_size,
 297                                 socket_id,
 298                                 RING_F_SP_ENQ | RING_F_SC_DEQ);
 299                 if (!qid->reorder_buffer_freelist) {
 300                         SW_LOG_DBG("freelist ring create failed");
 301                         goto cleanup;
 302                 }
 303
 304                 /* Populate the freelist with reorder buffer entries. Enqueue
 305                  * 'window_size - 1' entries because the rte_ring holds only
 306                  * that many.
 307                  */
 308                 for (i = 0; i < window_size - 1; i++) {
 309                         if (rte_ring_sp_enqueue(qid->reorder_buffer_freelist,
 310                                                 &qid->reorder_buffer[i]) < 0)
 311                                 goto cleanup;
 312                 }
 313
 314                 qid->reorder_buffer_index = 0;
 315                 qid->cq_next_tx = 0;
 316         }
 317
 318         qid->initialized = 1;
 319
 320         return 0;
 321
 322 cleanup:
 323         for (i = 0; i < SW_IQS_MAX; i++) {
 324                 if (qid->iq[i])
 325                         iq_ring_destroy(qid->iq[i]);
 326         }
 327
 328         if (qid->reorder_buffer) {
 329                 rte_free(qid->reorder_buffer);
 330                 qid->reorder_buffer = NULL;
 331         }
 332
 333         if (qid->reorder_buffer_freelist) {
 334                 rte_ring_free(qid->reorder_buffer_freelist);
 335                 qid->reorder_buffer_freelist = NULL;
 336         }
 337
 338         return -EINVAL;
 339 }
 340
 341 static int
 342 sw_queue_setup(struct rte_eventdev *dev, uint8_t queue_id,
 343                 const struct rte_event_queue_conf *conf)
 344 {
 345         int type;
 346
 347         /* SINGLE_LINK can be OR-ed with other types, so handle first */
 348         if (RTE_EVENT_QUEUE_CFG_SINGLE_LINK & conf->event_queue_cfg) {
 349                 type = SW_SCHED_TYPE_DIRECT;
 350         } else {
 351                 switch (conf->event_queue_cfg) {
 352                 case RTE_EVENT_QUEUE_CFG_ATOMIC_ONLY:
 353                         type = RTE_SCHED_TYPE_ATOMIC;
 354                         break;
 355                 case RTE_EVENT_QUEUE_CFG_ORDERED_ONLY:
 356                         type = RTE_SCHED_TYPE_ORDERED;
 357                         break;
 358                 case RTE_EVENT_QUEUE_CFG_PARALLEL_ONLY:
 359                         type = RTE_SCHED_TYPE_PARALLEL;
 360                         break;
 361                 case RTE_EVENT_QUEUE_CFG_ALL_TYPES:
 362                         SW_LOG_ERR("QUEUE_CFG_ALL_TYPES not supported\n");
 363                         return -ENOTSUP;
 364                 default:
 365                         SW_LOG_ERR("Unknown queue type %d requested\n",
 366                                    conf->event_queue_cfg);
 367                         return -EINVAL;
 368                 }
 369         }
 370
 371         struct sw_evdev *sw = sw_pmd_priv(dev);
 372         return qid_init(sw, queue_id, type, conf);
 373 }
 374
 375 static void
 376 sw_queue_release(struct rte_eventdev *dev, uint8_t id)
 377 {
 378         struct sw_evdev *sw = sw_pmd_priv(dev);
 379         struct sw_qid *qid = &sw->qids[id];
 380         uint32_t i;
 381
 382         for (i = 0; i < SW_IQS_MAX; i++)
 383                 iq_ring_destroy(qid->iq[i]);
 384
 385         if (qid->type == RTE_SCHED_TYPE_ORDERED) {
 386                 rte_free(qid->reorder_buffer);
 387                 rte_ring_free(qid->reorder_buffer_freelist);
 388         }
 389         memset(qid, 0, sizeof(*qid));
 390 }
 391
 392 static void
 393 sw_queue_def_conf(struct rte_eventdev *dev, uint8_t queue_id,
 394                                  struct rte_event_queue_conf *conf)
 395 {
 396         RTE_SET_USED(dev);
 397         RTE_SET_USED(queue_id);
 398
 399         static const struct rte_event_queue_conf default_conf = {
 400                 .nb_atomic_flows = 4096,
 401                 .nb_atomic_order_sequences = 1,
 402                 .event_queue_cfg = RTE_EVENT_QUEUE_CFG_ATOMIC_ONLY,
 403                 .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 404         };
 405
 406         *conf = default_conf;
 407 }
 408
 409 static void
 410 sw_port_def_conf(struct rte_eventdev *dev, uint8_t port_id,
 411                  struct rte_event_port_conf *port_conf)
 412 {
 413         RTE_SET_USED(dev);
 414         RTE_SET_USED(port_id);
 415
 416         port_conf->new_event_threshold = 1024;
 417         port_conf->dequeue_depth = 16;
 418         port_conf->enqueue_depth = 16;
 419 }
 420
 421 static int
 422 sw_dev_configure(const struct rte_eventdev *dev)
 423 {
 424         struct sw_evdev *sw = sw_pmd_priv(dev);
 425         const struct rte_eventdev_data *data = dev->data;
 426         const struct rte_event_dev_config *conf = &data->dev_conf;
 427
 428         sw->qid_count = conf->nb_event_queues;
 429         sw->port_count = conf->nb_event_ports;
 430         sw->nb_events_limit = conf->nb_events_limit;
 431         rte_atomic32_set(&sw->inflights, 0);
 432
 433         if (conf->event_dev_cfg & RTE_EVENT_DEV_CFG_PER_DEQUEUE_TIMEOUT)
 434                 return -ENOTSUP;
 435
 436         return 0;
 437 }
 438
 439 static void
 440 sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info)
 441 {
 442         RTE_SET_USED(dev);
 443
 444         static const struct rte_event_dev_info evdev_sw_info = {
 445                         .driver_name = SW_PMD_NAME,
 446                         .max_event_queues = RTE_EVENT_MAX_QUEUES_PER_DEV,
 447                         .max_event_queue_flows = SW_QID_NUM_FIDS,
 448                         .max_event_queue_priority_levels = SW_Q_PRIORITY_MAX,
 449                         .max_event_priority_levels = SW_IQS_MAX,
 450                         .max_event_ports = SW_PORTS_MAX,
 451                         .max_event_port_dequeue_depth = MAX_SW_CONS_Q_DEPTH,
 452                         .max_event_port_enqueue_depth = MAX_SW_PROD_Q_DEPTH,
 453                         .max_num_events = SW_INFLIGHT_EVENTS_TOTAL,
 454                         .event_dev_cap = (RTE_EVENT_DEV_CAP_QUEUE_QOS |
 455                                         RTE_EVENT_DEV_CAP_BURST_MODE |
 456                                         RTE_EVENT_DEV_CAP_EVENT_QOS),
 457         };
 458
 459         *info = evdev_sw_info;
 460 }
 461
 462 static void
 463 sw_dump(struct rte_eventdev *dev, FILE *f)
 464 {
 465         const struct sw_evdev *sw = sw_pmd_priv(dev);
 466
 467         static const char * const q_type_strings[] = {
 468                         "Ordered", "Atomic", "Parallel", "Directed"
 469         };
 470         uint32_t i;
 471         fprintf(f, "EventDev %s: ports %d, qids %d\n", "todo-fix-name",
 472                         sw->port_count, sw->qid_count);
 473
 474         fprintf(f, "\trx   %"PRIu64"\n\tdrop %"PRIu64"\n\ttx   %"PRIu64"\n",
 475                 sw->stats.rx_pkts, sw->stats.rx_dropped, sw->stats.tx_pkts);
 476         fprintf(f, "\tsched calls: %"PRIu64"\n", sw->sched_called);
 477         fprintf(f, "\tsched cq/qid call: %"PRIu64"\n", sw->sched_cq_qid_called);
 478         fprintf(f, "\tsched no IQ enq: %"PRIu64"\n", sw->sched_no_iq_enqueues);
 479         fprintf(f, "\tsched no CQ enq: %"PRIu64"\n", sw->sched_no_cq_enqueues);
 480         uint32_t inflights = rte_atomic32_read(&sw->inflights);
 481         uint32_t credits = sw->nb_events_limit - inflights;
 482         fprintf(f, "\tinflight %d, credits: %d\n", inflights, credits);
 483
 484 #define COL_RED "\x1b[31m"
 485 #define COL_RESET "\x1b[0m"
 486
 487         for (i = 0; i < sw->port_count; i++) {
 488                 int max, j;
 489                 const struct sw_port *p = &sw->ports[i];
 490                 if (!p->initialized) {
 491                         fprintf(f, "  %sPort %d not initialized.%s\n",
 492                                 COL_RED, i, COL_RESET);
 493                         continue;
 494                 }
 495                 fprintf(f, "  Port %d %s\n", i,
 496                         p->is_directed ? " (SingleCons)" : "");
 497                 fprintf(f, "\trx   %"PRIu64"\tdrop %"PRIu64"\ttx   %"PRIu64
 498                         "\t%sinflight %d%s\n", sw->ports[i].stats.rx_pkts,
 499                         sw->ports[i].stats.rx_dropped,
 500                         sw->ports[i].stats.tx_pkts,
 501                         (p->inflights == p->inflight_max) ?
 502                                 COL_RED : COL_RESET,
 503                         sw->ports[i].inflights, COL_RESET);
 504
 505                 fprintf(f, "\tMax New: %u"
 506                         "\tAvg cycles PP: %"PRIu64"\tCredits: %u\n",
 507                         sw->ports[i].inflight_max,
 508                         sw->ports[i].avg_pkt_ticks,
 509                         sw->ports[i].inflight_credits);
 510                 fprintf(f, "\tReceive burst distribution:\n");
 511                 float zp_percent = p->zero_polls * 100.0 / p->total_polls;
 512                 fprintf(f, zp_percent < 10 ? "\t\t0:%.02f%% " : "\t\t0:%.0f%% ",
 513                                 zp_percent);
 514                 for (max = (int)RTE_DIM(p->poll_buckets); max-- > 0;)
 515                         if (p->poll_buckets[max] != 0)
 516                                 break;
 517                 for (j = 0; j <= max; j++) {
 518                         if (p->poll_buckets[j] != 0) {
 519                                 float poll_pc = p->poll_buckets[j] * 100.0 /
 520                                         p->total_polls;
 521                                 fprintf(f, "%u-%u:%.02f%% ",
 522                                         ((j << SW_DEQ_STAT_BUCKET_SHIFT) + 1),
 523                                         ((j+1) << SW_DEQ_STAT_BUCKET_SHIFT),
 524                                         poll_pc);
 525                         }
 526                 }
 527                 fprintf(f, "\n");
 528
 529                 if (p->rx_worker_ring) {
 530                         uint64_t used = rte_event_ring_count(p->rx_worker_ring);
 531                         uint64_t space = rte_event_ring_free_count(
 532                                         p->rx_worker_ring);
 533                         const char *col = (space == 0) ? COL_RED : COL_RESET;
 534                         fprintf(f, "\t%srx ring used: %4"PRIu64"\tfree: %4"
 535                                         PRIu64 COL_RESET"\n", col, used, space);
 536                 } else
 537                         fprintf(f, "\trx ring not initialized.\n");
 538
 539                 if (p->cq_worker_ring) {
 540                         uint64_t used = rte_event_ring_count(p->cq_worker_ring);
 541                         uint64_t space = rte_event_ring_free_count(
 542                                         p->cq_worker_ring);
 543                         const char *col = (space == 0) ? COL_RED : COL_RESET;
 544                         fprintf(f, "\t%scq ring used: %4"PRIu64"\tfree: %4"
 545                                         PRIu64 COL_RESET"\n", col, used, space);
 546                 } else
 547                         fprintf(f, "\tcq ring not initialized.\n");
 548         }
 549
 550         for (i = 0; i < sw->qid_count; i++) {
 551                 const struct sw_qid *qid = &sw->qids[i];
 552                 if (!qid->initialized) {
 553                         fprintf(f, "  %sQueue %d not initialized.%s\n",
 554                                 COL_RED, i, COL_RESET);
 555                         continue;
 556                 }
 557                 int affinities_per_port[SW_PORTS_MAX] = {0};
 558                 uint32_t inflights = 0;
 559
 560                 fprintf(f, "  Queue %d (%s)\n", i, q_type_strings[qid->type]);
 561                 fprintf(f, "\trx   %"PRIu64"\tdrop %"PRIu64"\ttx   %"PRIu64"\n",
 562                         qid->stats.rx_pkts, qid->stats.rx_dropped,
 563                         qid->stats.tx_pkts);
 564                 if (qid->type == RTE_SCHED_TYPE_ORDERED) {
 565                         struct rte_ring *rob_buf_free =
 566                                 qid->reorder_buffer_freelist;
 567                         if (rob_buf_free)
 568                                 fprintf(f, "\tReorder entries in use: %u\n",
 569                                         rte_ring_free_count(rob_buf_free));
 570                         else
 571                                 fprintf(f,
 572                                         "\tReorder buffer not initialized\n");
 573                 }
 574
 575                 uint32_t flow;
 576                 for (flow = 0; flow < RTE_DIM(qid->fids); flow++)
 577                         if (qid->fids[flow].cq != -1) {
 578                                 affinities_per_port[qid->fids[flow].cq]++;
 579                                 inflights += qid->fids[flow].pcount;
 580                         }
 581
 582                 uint32_t port;
 583                 fprintf(f, "\tPer Port Stats:\n");
 584                 for (port = 0; port < sw->port_count; port++) {
 585                         fprintf(f, "\t  Port %d: Pkts: %"PRIu64, port,
 586                                         qid->to_port[port]);
 587                         fprintf(f, "\tFlows: %d\n", affinities_per_port[port]);
 588                 }
 589
 590                 uint32_t iq;
 591                 uint32_t iq_printed = 0;
 592                 for (iq = 0; iq < SW_IQS_MAX; iq++) {
 593                         if (!qid->iq[iq]) {
 594                                 fprintf(f, "\tiq %d is not initialized.\n", iq);
 595                                 iq_printed = 1;
 596                                 continue;
 597                         }
 598                         uint32_t used = iq_ring_count(qid->iq[iq]);
 599                         uint32_t free = iq_ring_free_count(qid->iq[iq]);
 600                         const char *col = (free == 0) ? COL_RED : COL_RESET;
 601                         if (used > 0) {
 602                                 fprintf(f, "\t%siq %d: Used %d\tFree %d"
 603                                         COL_RESET"\n", col, iq, used, free);
 604                                 iq_printed = 1;
 605                         }
 606                 }
 607                 if (iq_printed == 0)
 608                         fprintf(f, "\t-- iqs empty --\n");
 609         }
 610 }
 611
 612 static int
 613 sw_start(struct rte_eventdev *dev)
 614 {
 615         unsigned int i, j;
 616         struct sw_evdev *sw = sw_pmd_priv(dev);
 617         /* check all ports are set up */
 618         for (i = 0; i < sw->port_count; i++)
 619                 if (sw->ports[i].rx_worker_ring == NULL) {
 620                         SW_LOG_ERR("Port %d not configured\n", i);
 621                         return -ESTALE;
 622                 }
 623
 624         /* check all queues are configured and mapped to ports*/
 625         for (i = 0; i < sw->qid_count; i++)
 626                 if (sw->qids[i].iq[0] == NULL ||
 627                                 sw->qids[i].cq_num_mapped_cqs == 0) {
 628                         SW_LOG_ERR("Queue %d not configured\n", i);
 629                         return -ENOLINK;
 630                 }
 631
 632         /* build up our prioritized array of qids */
 633         /* We don't use qsort here, as if all/multiple entries have the same
 634          * priority, the result is non-deterministic. From "man 3 qsort":
 635          * "If two members compare as equal, their order in the sorted
 636          * array is undefined."
 637          */
 638         uint32_t qidx = 0;
 639         for (j = 0; j <= RTE_EVENT_DEV_PRIORITY_LOWEST; j++) {
 640                 for (i = 0; i < sw->qid_count; i++) {
 641                         if (sw->qids[i].priority == j) {
 642                                 sw->qids_prioritized[qidx] = &sw->qids[i];
 643                                 qidx++;
 644                         }
 645                 }
 646         }
 647
 648         if (sw_xstats_init(sw) < 0)
 649                 return -EINVAL;
 650
 651         rte_smp_wmb();
 652         sw->started = 1;
 653
 654         return 0;
 655 }
 656
 657 static void
 658 sw_stop(struct rte_eventdev *dev)
 659 {
 660         struct sw_evdev *sw = sw_pmd_priv(dev);
 661         sw_xstats_uninit(sw);
 662         sw->started = 0;
 663         rte_smp_wmb();
 664 }
 665
 666 static int
 667 sw_close(struct rte_eventdev *dev)
 668 {
 669         struct sw_evdev *sw = sw_pmd_priv(dev);
 670         uint32_t i;
 671
 672         for (i = 0; i < sw->qid_count; i++)
 673                 sw_queue_release(dev, i);
 674         sw->qid_count = 0;
 675
 676         for (i = 0; i < sw->port_count; i++)
 677                 sw_port_release(&sw->ports[i]);
 678         sw->port_count = 0;
 679
 680         memset(&sw->stats, 0, sizeof(sw->stats));
 681         sw->sched_called = 0;
 682         sw->sched_no_iq_enqueues = 0;
 683         sw->sched_no_cq_enqueues = 0;
 684         sw->sched_cq_qid_called = 0;
 685
 686         return 0;
 687 }
 688
 689 static int
 690 assign_numa_node(const char *key __rte_unused, const char *value, void *opaque)
 691 {
 692         int *socket_id = opaque;
 693         *socket_id = atoi(value);
 694         if (*socket_id >= RTE_MAX_NUMA_NODES)
 695                 return -1;
 696         return 0;
 697 }
 698
 699 static int
 700 set_sched_quanta(const char *key __rte_unused, const char *value, void *opaque)
 701 {
 702         int *quanta = opaque;
 703         *quanta = atoi(value);
 704         if (*quanta < 0 || *quanta >= 4096)
 705                 return -1;
 706         return 0;
 707 }
 708
 709 static int
 710 set_credit_quanta(const char *key __rte_unused, const char *value, void *opaque)
 711 {
 712         int *credit = opaque;
 713         *credit = atoi(value);
 714         if (*credit < 0 || *credit >= 128)
 715                 return -1;
 716         return 0;
 717 }
 718
 719 static int
 720 sw_probe(struct rte_vdev_device *vdev)
 721 {
 722         static const struct rte_eventdev_ops evdev_sw_ops = {
 723                         .dev_configure = sw_dev_configure,
 724                         .dev_infos_get = sw_info_get,
 725                         .dev_close = sw_close,
 726                         .dev_start = sw_start,
 727                         .dev_stop = sw_stop,
 728                         .dump = sw_dump,
 729
 730                         .queue_def_conf = sw_queue_def_conf,
 731                         .queue_setup = sw_queue_setup,
 732                         .queue_release = sw_queue_release,
 733                         .port_def_conf = sw_port_def_conf,
 734                         .port_setup = sw_port_setup,
 735                         .port_release = sw_port_release,
 736                         .port_link = sw_port_link,
 737                         .port_unlink = sw_port_unlink,
 738
 739                         .xstats_get = sw_xstats_get,
 740                         .xstats_get_names = sw_xstats_get_names,
 741                         .xstats_get_by_name = sw_xstats_get_by_name,
 742                         .xstats_reset = sw_xstats_reset,
 743         };
 744
 745         static const char *const args[] = {
 746                 NUMA_NODE_ARG,
 747                 SCHED_QUANTA_ARG,
 748                 CREDIT_QUANTA_ARG,
 749                 NULL
 750         };
 751         const char *name;
 752         const char *params;
 753         struct rte_eventdev *dev;
 754         struct sw_evdev *sw;
 755         int socket_id = rte_socket_id();
 756         int sched_quanta  = SW_DEFAULT_SCHED_QUANTA;
 757         int credit_quanta = SW_DEFAULT_CREDIT_QUANTA;
 758
 759         name = rte_vdev_device_name(vdev);
 760         params = rte_vdev_device_args(vdev);
 761         if (params != NULL && params[0] != '\0') {
 762                 struct rte_kvargs *kvlist = rte_kvargs_parse(params, args);
 763
 764                 if (!kvlist) {
 765                         SW_LOG_INFO(
 766                                 "Ignoring unsupported parameters when creating device '%s'\n",
 767                                 name);
 768                 } else {
 769                         int ret = rte_kvargs_process(kvlist, NUMA_NODE_ARG,
 770                                         assign_numa_node, &socket_id);
 771                         if (ret != 0) {
 772                                 SW_LOG_ERR(
 773                                         "%s: Error parsing numa node parameter",
 774                                         name);
 775                                 rte_kvargs_free(kvlist);
 776                                 return ret;
 777                         }
 778
 779                         ret = rte_kvargs_process(kvlist, SCHED_QUANTA_ARG,
 780                                         set_sched_quanta, &sched_quanta);
 781                         if (ret != 0) {
 782                                 SW_LOG_ERR(
 783                                         "%s: Error parsing sched quanta parameter",
 784                                         name);
 785                                 rte_kvargs_free(kvlist);
 786                                 return ret;
 787                         }
 788
 789                         ret = rte_kvargs_process(kvlist, CREDIT_QUANTA_ARG,
 790                                         set_credit_quanta, &credit_quanta);
 791                         if (ret != 0) {
 792                                 SW_LOG_ERR(
 793                                         "%s: Error parsing credit quanta parameter",
 794                                         name);
 795                                 rte_kvargs_free(kvlist);
 796                                 return ret;
 797                         }
 798
 799                         rte_kvargs_free(kvlist);
 800                 }
 801         }
 802
 803         SW_LOG_INFO(
 804                         "Creating eventdev sw device %s, numa_node=%d, sched_quanta=%d, credit_quanta=%d\n",
 805                         name, socket_id, sched_quanta, credit_quanta);
 806
 807         dev = rte_event_pmd_vdev_init(name,
 808                         sizeof(struct sw_evdev), socket_id);
 809         if (dev == NULL) {
 810                 SW_LOG_ERR("eventdev vdev init() failed");
 811                 return -EFAULT;
 812         }
 813         dev->dev_ops = &evdev_sw_ops;
 814         dev->enqueue = sw_event_enqueue;
 815         dev->enqueue_burst = sw_event_enqueue_burst;
 816         dev->enqueue_new_burst = sw_event_enqueue_burst;
 817         dev->enqueue_forward_burst = sw_event_enqueue_burst;
 818         dev->dequeue = sw_event_dequeue;
 819         dev->dequeue_burst = sw_event_dequeue_burst;
 820         dev->schedule = sw_event_schedule;
 821
 822         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 823                 return 0;
 824
 825         sw = dev->data->dev_private;
 826         sw->data = dev->data;
 827
 828         /* copy values passed from vdev command line to instance */
 829         sw->credit_update_quanta = credit_quanta;
 830         sw->sched_quanta = sched_quanta;
 831
 832         return 0;
 833 }
 834
 835 static int
 836 sw_remove(struct rte_vdev_device *vdev)
 837 {
 838         const char *name;
 839
 840         name = rte_vdev_device_name(vdev);
 841         if (name == NULL)
 842                 return -EINVAL;
 843
 844         SW_LOG_INFO("Closing eventdev sw device %s\n", name);
 845
 846         return rte_event_pmd_vdev_uninit(name);
 847 }
 848
 849 static struct rte_vdev_driver evdev_sw_pmd_drv = {
 850         .probe = sw_probe,
 851         .remove = sw_remove
 852 };
 853
 854 RTE_PMD_REGISTER_VDEV(EVENTDEV_NAME_SW_PMD, evdev_sw_pmd_drv);
 855 RTE_PMD_REGISTER_PARAM_STRING(event_sw, NUMA_NODE_ARG "=<int> "
 856                 SCHED_QUANTA_ARG "=<int>" CREDIT_QUANTA_ARG "=<int>");