drivers/event/dsw/dsw_event.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Ericsson AB
   3  */
   4
   5 #include "dsw_evdev.h"
   6
   7 #ifdef DSW_SORT_DEQUEUED
   8 #include "dsw_sort.h"
   9 #endif
  10
  11 #include <stdbool.h>
  12 #include <string.h>
  13
  14 #include <rte_atomic.h>
  15 #include <rte_cycles.h>
  16 #include <rte_memcpy.h>
  17 #include <rte_random.h>
  18
  19 static bool
  20 dsw_port_acquire_credits(struct dsw_evdev *dsw, struct dsw_port *port,
  21                          int32_t credits)
  22 {
  23         int32_t inflight_credits = port->inflight_credits;
  24         int32_t missing_credits = credits - inflight_credits;
  25         int32_t total_on_loan;
  26         int32_t available;
  27         int32_t acquired_credits;
  28         int32_t new_total_on_loan;
  29
  30         if (likely(missing_credits <= 0)) {
  31                 port->inflight_credits -= credits;
  32                 return true;
  33         }
  34
  35         total_on_loan = rte_atomic32_read(&dsw->credits_on_loan);
  36         available = dsw->max_inflight - total_on_loan;
  37         acquired_credits = RTE_MAX(missing_credits, DSW_PORT_MIN_CREDITS);
  38
  39         if (available < acquired_credits)
  40                 return false;
  41
  42         /* This is a race, no locks are involved, and thus some other
  43          * thread can allocate tokens in between the check and the
  44          * allocation.
  45          */
  46         new_total_on_loan = rte_atomic32_add_return(&dsw->credits_on_loan,
  47                                                     acquired_credits);
  48
  49         if (unlikely(new_total_on_loan > dsw->max_inflight)) {
  50                 /* Some other port took the last credits */
  51                 rte_atomic32_sub(&dsw->credits_on_loan, acquired_credits);
  52                 return false;
  53         }
  54
  55         DSW_LOG_DP_PORT(DEBUG, port->id, "Acquired %d tokens from pool.\n",
  56                         acquired_credits);
  57
  58         port->inflight_credits += acquired_credits;
  59         port->inflight_credits -= credits;
  60
  61         return true;
  62 }
  63
  64 static void
  65 dsw_port_return_credits(struct dsw_evdev *dsw, struct dsw_port *port,
  66                         int32_t credits)
  67 {
  68         port->inflight_credits += credits;
  69
  70         if (unlikely(port->inflight_credits > DSW_PORT_MAX_CREDITS)) {
  71                 int32_t leave_credits = DSW_PORT_MIN_CREDITS;
  72                 int32_t return_credits =
  73                         port->inflight_credits - leave_credits;
  74
  75                 port->inflight_credits = leave_credits;
  76
  77                 rte_atomic32_sub(&dsw->credits_on_loan, return_credits);
  78
  79                 DSW_LOG_DP_PORT(DEBUG, port->id,
  80                                 "Returned %d tokens to pool.\n",
  81                                 return_credits);
  82         }
  83 }
  84
  85 static void
  86 dsw_port_enqueue_stats(struct dsw_port *port, uint16_t num_new,
  87                        uint16_t num_forward, uint16_t num_release)
  88 {
  89         port->new_enqueued += num_new;
  90         port->forward_enqueued += num_forward;
  91         port->release_enqueued += num_release;
  92 }
  93
  94 static void
  95 dsw_port_queue_enqueue_stats(struct dsw_port *source_port, uint8_t queue_id)
  96 {
  97         source_port->queue_enqueued[queue_id]++;
  98 }
  99
 100 static void
 101 dsw_port_dequeue_stats(struct dsw_port *port, uint16_t num)
 102 {
 103         port->dequeued += num;
 104 }
 105
 106 static void
 107 dsw_port_queue_dequeued_stats(struct dsw_port *source_port, uint8_t queue_id)
 108 {
 109         source_port->queue_dequeued[queue_id]++;
 110 }
 111
 112 static void
 113 dsw_port_load_record(struct dsw_port *port, unsigned int dequeued)
 114 {
 115         if (dequeued > 0 && port->busy_start == 0)
 116                 /* work period begins */
 117                 port->busy_start = rte_get_timer_cycles();
 118         else if (dequeued == 0 && port->busy_start > 0) {
 119                 /* work period ends */
 120                 uint64_t work_period =
 121                         rte_get_timer_cycles() - port->busy_start;
 122                 port->busy_cycles += work_period;
 123                 port->busy_start = 0;
 124         }
 125 }
 126
 127 static int16_t
 128 dsw_port_load_close_period(struct dsw_port *port, uint64_t now)
 129 {
 130         uint64_t passed = now - port->measurement_start;
 131         uint64_t busy_cycles = port->busy_cycles;
 132
 133         if (port->busy_start > 0) {
 134                 busy_cycles += (now - port->busy_start);
 135                 port->busy_start = now;
 136         }
 137
 138         int16_t load = (DSW_MAX_LOAD * busy_cycles) / passed;
 139
 140         port->measurement_start = now;
 141         port->busy_cycles = 0;
 142
 143         port->total_busy_cycles += busy_cycles;
 144
 145         return load;
 146 }
 147
 148 static void
 149 dsw_port_load_update(struct dsw_port *port, uint64_t now)
 150 {
 151         int16_t old_load;
 152         int16_t period_load;
 153         int16_t new_load;
 154
 155         old_load = rte_atomic16_read(&port->load);
 156
 157         period_load = dsw_port_load_close_period(port, now);
 158
 159         new_load = (period_load + old_load*DSW_OLD_LOAD_WEIGHT) /
 160                 (DSW_OLD_LOAD_WEIGHT+1);
 161
 162         rte_atomic16_set(&port->load, new_load);
 163 }
 164
 165 static void
 166 dsw_port_consider_load_update(struct dsw_port *port, uint64_t now)
 167 {
 168         if (now < port->next_load_update)
 169                 return;
 170
 171         port->next_load_update = now + port->load_update_interval;
 172
 173         dsw_port_load_update(port, now);
 174 }
 175
 176 static void
 177 dsw_port_ctl_enqueue(struct dsw_port *port, struct dsw_ctl_msg *msg)
 178 {
 179         /* there's always room on the ring */
 180         while (rte_ring_enqueue_elem(port->ctl_in_ring, msg, sizeof(*msg)) != 0)
 181                 rte_pause();
 182 }
 183
 184 static int
 185 dsw_port_ctl_dequeue(struct dsw_port *port, struct dsw_ctl_msg *msg)
 186 {
 187         return rte_ring_dequeue_elem(port->ctl_in_ring, msg, sizeof(*msg));
 188 }
 189
 190 static void
 191 dsw_port_ctl_broadcast(struct dsw_evdev *dsw, struct dsw_port *source_port,
 192                        uint8_t type, uint8_t queue_id, uint16_t flow_hash)
 193 {
 194         uint16_t port_id;
 195         struct dsw_ctl_msg msg = {
 196                 .type = type,
 197                 .originating_port_id = source_port->id,
 198                 .queue_id = queue_id,
 199                 .flow_hash = flow_hash
 200         };
 201
 202         for (port_id = 0; port_id < dsw->num_ports; port_id++)
 203                 if (port_id != source_port->id)
 204                         dsw_port_ctl_enqueue(&dsw->ports[port_id], &msg);
 205 }
 206
 207 static bool
 208 dsw_port_is_flow_paused(struct dsw_port *port, uint8_t queue_id,
 209                         uint16_t flow_hash)
 210 {
 211         uint16_t i;
 212
 213         for (i = 0; i < port->paused_flows_len; i++) {
 214                 struct dsw_queue_flow *qf = &port->paused_flows[i];
 215                 if (qf->queue_id == queue_id &&
 216                     qf->flow_hash == flow_hash)
 217                         return true;
 218         }
 219         return false;
 220 }
 221
 222 static void
 223 dsw_port_add_paused_flow(struct dsw_port *port, uint8_t queue_id,
 224                          uint16_t paused_flow_hash)
 225 {
 226         port->paused_flows[port->paused_flows_len] = (struct dsw_queue_flow) {
 227                 .queue_id = queue_id,
 228                 .flow_hash = paused_flow_hash
 229         };
 230         port->paused_flows_len++;
 231 }
 232
 233 static void
 234 dsw_port_remove_paused_flow(struct dsw_port *port, uint8_t queue_id,
 235                             uint16_t paused_flow_hash)
 236 {
 237         uint16_t i;
 238
 239         for (i = 0; i < port->paused_flows_len; i++) {
 240                 struct dsw_queue_flow *qf = &port->paused_flows[i];
 241
 242                 if (qf->queue_id == queue_id &&
 243                     qf->flow_hash == paused_flow_hash) {
 244                         uint16_t last_idx = port->paused_flows_len-1;
 245                         if (i != last_idx)
 246                                 port->paused_flows[i] =
 247                                         port->paused_flows[last_idx];
 248                         port->paused_flows_len--;
 249                         break;
 250                 }
 251         }
 252 }
 253
 254 static void
 255 dsw_port_flush_out_buffers(struct dsw_evdev *dsw, struct dsw_port *source_port);
 256
 257 static void
 258 dsw_port_handle_pause_flow(struct dsw_evdev *dsw, struct dsw_port *port,
 259                            uint8_t originating_port_id, uint8_t queue_id,
 260                            uint16_t paused_flow_hash)
 261 {
 262         struct dsw_ctl_msg cfm = {
 263                 .type = DSW_CTL_CFM,
 264                 .originating_port_id = port->id,
 265                 .queue_id = queue_id,
 266                 .flow_hash = paused_flow_hash
 267         };
 268
 269         DSW_LOG_DP_PORT(DEBUG, port->id, "Pausing queue_id %d flow_hash %d.\n",
 270                         queue_id, paused_flow_hash);
 271
 272         /* There might be already-scheduled events belonging to the
 273          * paused flow in the output buffers.
 274          */
 275         dsw_port_flush_out_buffers(dsw, port);
 276
 277         dsw_port_add_paused_flow(port, queue_id, paused_flow_hash);
 278
 279         /* Make sure any stores to the original port's in_ring is seen
 280          * before the ctl message.
 281          */
 282         rte_smp_wmb();
 283
 284         dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm);
 285 }
 286
 287 static void
 288 dsw_find_lowest_load_port(uint8_t *port_ids, uint16_t num_port_ids,
 289                           uint8_t exclude_port_id, int16_t *port_loads,
 290                           uint8_t *target_port_id, int16_t *target_load)
 291 {
 292         int16_t candidate_port_id = -1;
 293         int16_t candidate_load = DSW_MAX_LOAD;
 294         uint16_t i;
 295
 296         for (i = 0; i < num_port_ids; i++) {
 297                 uint8_t port_id = port_ids[i];
 298                 if (port_id != exclude_port_id) {
 299                         int16_t load = port_loads[port_id];
 300                         if (candidate_port_id == -1 ||
 301                             load < candidate_load) {
 302                                 candidate_port_id = port_id;
 303                                 candidate_load = load;
 304                         }
 305                 }
 306         }
 307         *target_port_id = candidate_port_id;
 308         *target_load = candidate_load;
 309 }
 310
 311 struct dsw_queue_flow_burst {
 312         struct dsw_queue_flow queue_flow;
 313         uint16_t count;
 314 };
 315
 316 static inline int
 317 dsw_cmp_burst(const void *v_burst_a, const void *v_burst_b)
 318 {
 319         const struct dsw_queue_flow_burst *burst_a = v_burst_a;
 320         const struct dsw_queue_flow_burst *burst_b = v_burst_b;
 321
 322         int a_count = burst_a->count;
 323         int b_count = burst_b->count;
 324
 325         return a_count - b_count;
 326 }
 327
 328 #define DSW_QF_TO_INT(_qf)                                      \
 329         ((int)((((_qf)->queue_id)<<16)|((_qf)->flow_hash)))
 330
 331 static inline int
 332 dsw_cmp_qf(const void *v_qf_a, const void *v_qf_b)
 333 {
 334         const struct dsw_queue_flow *qf_a = v_qf_a;
 335         const struct dsw_queue_flow *qf_b = v_qf_b;
 336
 337         return DSW_QF_TO_INT(qf_a) - DSW_QF_TO_INT(qf_b);
 338 }
 339
 340 static uint16_t
 341 dsw_sort_qfs_to_bursts(struct dsw_queue_flow *qfs, uint16_t qfs_len,
 342                        struct dsw_queue_flow_burst *bursts)
 343 {
 344         uint16_t i;
 345         struct dsw_queue_flow_burst *current_burst = NULL;
 346         uint16_t num_bursts = 0;
 347
 348         /* We don't need the stable property, and the list is likely
 349          * large enough for qsort() to outperform dsw_stable_sort(),
 350          * so we use qsort() here.
 351          */
 352         qsort(qfs, qfs_len, sizeof(qfs[0]), dsw_cmp_qf);
 353
 354         /* arrange the (now-consecutive) events into bursts */
 355         for (i = 0; i < qfs_len; i++) {
 356                 if (i == 0 ||
 357                     dsw_cmp_qf(&qfs[i], &current_burst->queue_flow) != 0) {
 358                         current_burst = &bursts[num_bursts];
 359                         current_burst->queue_flow = qfs[i];
 360                         current_burst->count = 0;
 361                         num_bursts++;
 362                 }
 363                 current_burst->count++;
 364         }
 365
 366         qsort(bursts, num_bursts, sizeof(bursts[0]), dsw_cmp_burst);
 367
 368         return num_bursts;
 369 }
 370
 371 static bool
 372 dsw_retrieve_port_loads(struct dsw_evdev *dsw, int16_t *port_loads,
 373                         int16_t load_limit)
 374 {
 375         bool below_limit = false;
 376         uint16_t i;
 377
 378         for (i = 0; i < dsw->num_ports; i++) {
 379                 int16_t load = rte_atomic16_read(&dsw->ports[i].load);
 380                 if (load < load_limit)
 381                         below_limit = true;
 382                 port_loads[i] = load;
 383         }
 384         return below_limit;
 385 }
 386
 387 static bool
 388 dsw_select_migration_target(struct dsw_evdev *dsw,
 389                             struct dsw_port *source_port,
 390                             struct dsw_queue_flow_burst *bursts,
 391                             uint16_t num_bursts, int16_t *port_loads,
 392                             int16_t max_load, struct dsw_queue_flow *target_qf,
 393                             uint8_t *target_port_id)
 394 {
 395         uint16_t source_load = port_loads[source_port->id];
 396         uint16_t i;
 397
 398         for (i = 0; i < num_bursts; i++) {
 399                 struct dsw_queue_flow *qf = &bursts[i].queue_flow;
 400
 401                 if (dsw_port_is_flow_paused(source_port, qf->queue_id,
 402                                             qf->flow_hash))
 403                         continue;
 404
 405                 struct dsw_queue *queue = &dsw->queues[qf->queue_id];
 406                 int16_t target_load;
 407
 408                 dsw_find_lowest_load_port(queue->serving_ports,
 409                                           queue->num_serving_ports,
 410                                           source_port->id, port_loads,
 411                                           target_port_id, &target_load);
 412
 413                 if (target_load < source_load &&
 414                     target_load < max_load) {
 415                         *target_qf = *qf;
 416                         return true;
 417                 }
 418         }
 419
 420         DSW_LOG_DP_PORT(DEBUG, source_port->id, "For the %d flows considered, "
 421                         "no target port found with load less than %d.\n",
 422                         num_bursts, DSW_LOAD_TO_PERCENT(max_load));
 423
 424         return false;
 425 }
 426
 427 static uint8_t
 428 dsw_schedule(struct dsw_evdev *dsw, uint8_t queue_id, uint16_t flow_hash)
 429 {
 430         struct dsw_queue *queue = &dsw->queues[queue_id];
 431         uint8_t port_id;
 432
 433         if (queue->num_serving_ports > 1)
 434                 port_id = queue->flow_to_port_map[flow_hash];
 435         else
 436                 /* A single-link queue, or atomic/ordered/parallel but
 437                  * with just a single serving port.
 438                  */
 439                 port_id = queue->serving_ports[0];
 440
 441         DSW_LOG_DP(DEBUG, "Event with queue_id %d flow_hash %d is scheduled "
 442                    "to port %d.\n", queue_id, flow_hash, port_id);
 443
 444         return port_id;
 445 }
 446
 447 static void
 448 dsw_port_transmit_buffered(struct dsw_evdev *dsw, struct dsw_port *source_port,
 449                            uint8_t dest_port_id)
 450 {
 451         struct dsw_port *dest_port = &(dsw->ports[dest_port_id]);
 452         uint16_t *buffer_len = &source_port->out_buffer_len[dest_port_id];
 453         struct rte_event *buffer = source_port->out_buffer[dest_port_id];
 454         uint16_t enqueued = 0;
 455
 456         if (*buffer_len == 0)
 457                 return;
 458
 459         /* The rings are dimensioned to fit all in-flight events (even
 460          * on a single ring), so looping will work.
 461          */
 462         do {
 463                 enqueued +=
 464                         rte_event_ring_enqueue_burst(dest_port->in_ring,
 465                                                      buffer+enqueued,
 466                                                      *buffer_len-enqueued,
 467                                                      NULL);
 468         } while (unlikely(enqueued != *buffer_len));
 469
 470         (*buffer_len) = 0;
 471 }
 472
 473 static uint16_t
 474 dsw_port_get_parallel_flow_id(struct dsw_port *port)
 475 {
 476         uint16_t flow_id = port->next_parallel_flow_id;
 477
 478         port->next_parallel_flow_id =
 479                 (port->next_parallel_flow_id + 1) % DSW_PARALLEL_FLOWS;
 480
 481         return flow_id;
 482 }
 483
 484 static void
 485 dsw_port_buffer_paused(struct dsw_port *port,
 486                        const struct rte_event *paused_event)
 487 {
 488         port->paused_events[port->paused_events_len] = *paused_event;
 489         port->paused_events_len++;
 490 }
 491
 492 static void
 493 dsw_port_buffer_non_paused(struct dsw_evdev *dsw, struct dsw_port *source_port,
 494                            uint8_t dest_port_id, const struct rte_event *event)
 495 {
 496         struct rte_event *buffer = source_port->out_buffer[dest_port_id];
 497         uint16_t *buffer_len = &source_port->out_buffer_len[dest_port_id];
 498
 499         if (*buffer_len == DSW_MAX_PORT_OUT_BUFFER)
 500                 dsw_port_transmit_buffered(dsw, source_port, dest_port_id);
 501
 502         buffer[*buffer_len] = *event;
 503
 504         (*buffer_len)++;
 505 }
 506
 507 #define DSW_FLOW_ID_BITS (24)
 508 static uint16_t
 509 dsw_flow_id_hash(uint32_t flow_id)
 510 {
 511         uint16_t hash = 0;
 512         uint16_t offset = 0;
 513
 514         do {
 515                 hash ^= ((flow_id >> offset) & DSW_MAX_FLOWS_MASK);
 516                 offset += DSW_MAX_FLOWS_BITS;
 517         } while (offset < DSW_FLOW_ID_BITS);
 518
 519         return hash;
 520 }
 521
 522 static void
 523 dsw_port_buffer_parallel(struct dsw_evdev *dsw, struct dsw_port *source_port,
 524                          struct rte_event event)
 525 {
 526         uint8_t dest_port_id;
 527
 528         event.flow_id = dsw_port_get_parallel_flow_id(source_port);
 529
 530         dest_port_id = dsw_schedule(dsw, event.queue_id,
 531                                     dsw_flow_id_hash(event.flow_id));
 532
 533         dsw_port_buffer_non_paused(dsw, source_port, dest_port_id, &event);
 534 }
 535
 536 static void
 537 dsw_port_buffer_event(struct dsw_evdev *dsw, struct dsw_port *source_port,
 538                       const struct rte_event *event)
 539 {
 540         uint16_t flow_hash;
 541         uint8_t dest_port_id;
 542
 543         if (unlikely(dsw->queues[event->queue_id].schedule_type ==
 544                      RTE_SCHED_TYPE_PARALLEL)) {
 545                 dsw_port_buffer_parallel(dsw, source_port, *event);
 546                 return;
 547         }
 548
 549         flow_hash = dsw_flow_id_hash(event->flow_id);
 550
 551         if (unlikely(dsw_port_is_flow_paused(source_port, event->queue_id,
 552                                              flow_hash))) {
 553                 dsw_port_buffer_paused(source_port, event);
 554                 return;
 555         }
 556
 557         dest_port_id = dsw_schedule(dsw, event->queue_id, flow_hash);
 558
 559         dsw_port_buffer_non_paused(dsw, source_port, dest_port_id, event);
 560 }
 561
 562 static void
 563 dsw_port_flush_paused_events(struct dsw_evdev *dsw,
 564                              struct dsw_port *source_port,
 565                              uint8_t queue_id, uint16_t paused_flow_hash)
 566 {
 567         uint16_t paused_events_len = source_port->paused_events_len;
 568         struct rte_event paused_events[paused_events_len];
 569         uint8_t dest_port_id;
 570         uint16_t i;
 571
 572         if (paused_events_len == 0)
 573                 return;
 574
 575         if (dsw_port_is_flow_paused(source_port, queue_id, paused_flow_hash))
 576                 return;
 577
 578         rte_memcpy(paused_events, source_port->paused_events,
 579                    paused_events_len * sizeof(struct rte_event));
 580
 581         source_port->paused_events_len = 0;
 582
 583         dest_port_id = dsw_schedule(dsw, queue_id, paused_flow_hash);
 584
 585         for (i = 0; i < paused_events_len; i++) {
 586                 struct rte_event *event = &paused_events[i];
 587                 uint16_t flow_hash;
 588
 589                 flow_hash = dsw_flow_id_hash(event->flow_id);
 590
 591                 if (event->queue_id == queue_id &&
 592                     flow_hash == paused_flow_hash)
 593                         dsw_port_buffer_non_paused(dsw, source_port,
 594                                                    dest_port_id, event);
 595                 else
 596                         dsw_port_buffer_paused(source_port, event);
 597         }
 598 }
 599
 600 static void
 601 dsw_port_migration_stats(struct dsw_port *port)
 602 {
 603         uint64_t migration_latency;
 604
 605         migration_latency = (rte_get_timer_cycles() - port->migration_start);
 606         port->migration_latency += migration_latency;
 607         port->migrations++;
 608 }
 609
 610 static void
 611 dsw_port_end_migration(struct dsw_evdev *dsw, struct dsw_port *port)
 612 {
 613         uint8_t queue_id = port->migration_target_qf.queue_id;
 614         uint16_t flow_hash = port->migration_target_qf.flow_hash;
 615
 616         port->migration_state = DSW_MIGRATION_STATE_IDLE;
 617         port->seen_events_len = 0;
 618
 619         dsw_port_migration_stats(port);
 620
 621         if (dsw->queues[queue_id].schedule_type != RTE_SCHED_TYPE_PARALLEL) {
 622                 dsw_port_remove_paused_flow(port, queue_id, flow_hash);
 623                 dsw_port_flush_paused_events(dsw, port, queue_id, flow_hash);
 624         }
 625
 626         DSW_LOG_DP_PORT(DEBUG, port->id, "Migration completed for queue_id "
 627                         "%d flow_hash %d.\n", queue_id, flow_hash);
 628 }
 629
 630 static void
 631 dsw_port_consider_migration(struct dsw_evdev *dsw,
 632                             struct dsw_port *source_port,
 633                             uint64_t now)
 634 {
 635         bool any_port_below_limit;
 636         struct dsw_queue_flow *seen_events = source_port->seen_events;
 637         uint16_t seen_events_len = source_port->seen_events_len;
 638         struct dsw_queue_flow_burst bursts[DSW_MAX_EVENTS_RECORDED];
 639         uint16_t num_bursts;
 640         int16_t source_port_load;
 641         int16_t port_loads[dsw->num_ports];
 642
 643         if (now < source_port->next_migration)
 644                 return;
 645
 646         if (dsw->num_ports == 1)
 647                 return;
 648
 649         DSW_LOG_DP_PORT(DEBUG, source_port->id, "Considering migration.\n");
 650
 651         /* Randomize interval to avoid having all threads considering
 652          * migration at the same in point in time, which might lead to
 653          * all choosing the same target port.
 654          */
 655         source_port->next_migration = now +
 656                 source_port->migration_interval / 2 +
 657                 rte_rand() % source_port->migration_interval;
 658
 659         if (source_port->migration_state != DSW_MIGRATION_STATE_IDLE) {
 660                 DSW_LOG_DP_PORT(DEBUG, source_port->id,
 661                                 "Migration already in progress.\n");
 662                 return;
 663         }
 664
 665         /* For simplicity, avoid migration in the unlikely case there
 666          * is still events to consume in the in_buffer (from the last
 667          * migration).
 668          */
 669         if (source_port->in_buffer_len > 0) {
 670                 DSW_LOG_DP_PORT(DEBUG, source_port->id, "There are still "
 671                                 "events in the input buffer.\n");
 672                 return;
 673         }
 674
 675         source_port_load = rte_atomic16_read(&source_port->load);
 676         if (source_port_load < DSW_MIN_SOURCE_LOAD_FOR_MIGRATION) {
 677                 DSW_LOG_DP_PORT(DEBUG, source_port->id,
 678                                 "Load %d is below threshold level %d.\n",
 679                                 DSW_LOAD_TO_PERCENT(source_port_load),
 680                        DSW_LOAD_TO_PERCENT(DSW_MIN_SOURCE_LOAD_FOR_MIGRATION));
 681                 return;
 682         }
 683
 684         /* Avoid starting any expensive operations (sorting etc), in
 685          * case of a scenario with all ports above the load limit.
 686          */
 687         any_port_below_limit =
 688                 dsw_retrieve_port_loads(dsw, port_loads,
 689                                         DSW_MAX_TARGET_LOAD_FOR_MIGRATION);
 690         if (!any_port_below_limit) {
 691                 DSW_LOG_DP_PORT(DEBUG, source_port->id,
 692                                 "Candidate target ports are all too highly "
 693                                 "loaded.\n");
 694                 return;
 695         }
 696
 697         /* Sort flows into 'bursts' to allow attempting to migrating
 698          * small (but still active) flows first - this it to avoid
 699          * having large flows moving around the worker cores too much
 700          * (to avoid cache misses, among other things). Of course, the
 701          * number of recorded events (queue+flow ids) are limited, and
 702          * provides only a snapshot, so only so many conclusions can
 703          * be drawn from this data.
 704          */
 705         num_bursts = dsw_sort_qfs_to_bursts(seen_events, seen_events_len,
 706                                             bursts);
 707         /* For non-big-little systems, there's no point in moving the
 708          * only (known) flow.
 709          */
 710         if (num_bursts < 2) {
 711                 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Only a single flow "
 712                                 "queue_id %d flow_hash %d has been seen.\n",
 713                                 bursts[0].queue_flow.queue_id,
 714                                 bursts[0].queue_flow.flow_hash);
 715                 return;
 716         }
 717
 718         /* The strategy is to first try to find a flow to move to a
 719          * port with low load (below the migration-attempt
 720          * threshold). If that fails, we try to find a port which is
 721          * below the max threshold, and also less loaded than this
 722          * port is.
 723          */
 724         if (!dsw_select_migration_target(dsw, source_port, bursts, num_bursts,
 725                                          port_loads,
 726                                          DSW_MIN_SOURCE_LOAD_FOR_MIGRATION,
 727                                          &source_port->migration_target_qf,
 728                                          &source_port->migration_target_port_id)
 729             &&
 730             !dsw_select_migration_target(dsw, source_port, bursts, num_bursts,
 731                                          port_loads,
 732                                          DSW_MAX_TARGET_LOAD_FOR_MIGRATION,
 733                                          &source_port->migration_target_qf,
 734                                        &source_port->migration_target_port_id))
 735                 return;
 736
 737         DSW_LOG_DP_PORT(DEBUG, source_port->id, "Migrating queue_id %d "
 738                         "flow_hash %d from port %d to port %d.\n",
 739                         source_port->migration_target_qf.queue_id,
 740                         source_port->migration_target_qf.flow_hash,
 741                         source_port->id, source_port->migration_target_port_id);
 742
 743         /* We have a winner. */
 744
 745         source_port->migration_state = DSW_MIGRATION_STATE_PAUSING;
 746         source_port->migration_start = rte_get_timer_cycles();
 747
 748         /* No need to go through the whole pause procedure for
 749          * parallel queues, since atomic/ordered semantics need not to
 750          * be maintained.
 751          */
 752
 753         if (dsw->queues[source_port->migration_target_qf.queue_id].schedule_type
 754             == RTE_SCHED_TYPE_PARALLEL) {
 755                 uint8_t queue_id = source_port->migration_target_qf.queue_id;
 756                 uint16_t flow_hash = source_port->migration_target_qf.flow_hash;
 757                 uint8_t dest_port_id = source_port->migration_target_port_id;
 758
 759                 /* Single byte-sized stores are always atomic. */
 760                 dsw->queues[queue_id].flow_to_port_map[flow_hash] =
 761                         dest_port_id;
 762                 rte_smp_wmb();
 763
 764                 dsw_port_end_migration(dsw, source_port);
 765
 766                 return;
 767         }
 768
 769         /* There might be 'loopback' events already scheduled in the
 770          * output buffers.
 771          */
 772         dsw_port_flush_out_buffers(dsw, source_port);
 773
 774         dsw_port_add_paused_flow(source_port,
 775                                  source_port->migration_target_qf.queue_id,
 776                                  source_port->migration_target_qf.flow_hash);
 777
 778         dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_PAUS_REQ,
 779                                source_port->migration_target_qf.queue_id,
 780                                source_port->migration_target_qf.flow_hash);
 781         source_port->cfm_cnt = 0;
 782 }
 783
 784 static void
 785 dsw_port_flush_paused_events(struct dsw_evdev *dsw,
 786                              struct dsw_port *source_port,
 787                              uint8_t queue_id, uint16_t paused_flow_hash);
 788
 789 static void
 790 dsw_port_handle_unpause_flow(struct dsw_evdev *dsw, struct dsw_port *port,
 791                              uint8_t originating_port_id, uint8_t queue_id,
 792                              uint16_t paused_flow_hash)
 793 {
 794         struct dsw_ctl_msg cfm = {
 795                 .type = DSW_CTL_CFM,
 796                 .originating_port_id = port->id,
 797                 .queue_id = queue_id,
 798                 .flow_hash = paused_flow_hash
 799         };
 800
 801         DSW_LOG_DP_PORT(DEBUG, port->id, "Un-pausing queue_id %d flow_hash %d.\n",
 802                         queue_id, paused_flow_hash);
 803
 804         dsw_port_remove_paused_flow(port, queue_id, paused_flow_hash);
 805
 806         rte_smp_rmb();
 807
 808         dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm);
 809
 810         dsw_port_flush_paused_events(dsw, port, queue_id, paused_flow_hash);
 811 }
 812
 813 #define FORWARD_BURST_SIZE (32)
 814
 815 static void
 816 dsw_port_forward_migrated_flow(struct dsw_port *source_port,
 817                                struct rte_event_ring *dest_ring,
 818                                uint8_t queue_id,
 819                                uint16_t flow_hash)
 820 {
 821         uint16_t events_left;
 822
 823         /* Control ring message should been seen before the ring count
 824          * is read on the port's in_ring.
 825          */
 826         rte_smp_rmb();
 827
 828         events_left = rte_event_ring_count(source_port->in_ring);
 829
 830         while (events_left > 0) {
 831                 uint16_t in_burst_size =
 832                         RTE_MIN(FORWARD_BURST_SIZE, events_left);
 833                 struct rte_event in_burst[in_burst_size];
 834                 uint16_t in_len;
 835                 uint16_t i;
 836
 837                 in_len = rte_event_ring_dequeue_burst(source_port->in_ring,
 838                                                       in_burst,
 839                                                       in_burst_size, NULL);
 840                 /* No need to care about bursting forwarded events (to
 841                  * the destination port's in_ring), since migration
 842                  * doesn't happen very often, and also the majority of
 843                  * the dequeued events will likely *not* be forwarded.
 844                  */
 845                 for (i = 0; i < in_len; i++) {
 846                         struct rte_event *e = &in_burst[i];
 847                         if (e->queue_id == queue_id &&
 848                             dsw_flow_id_hash(e->flow_id) == flow_hash) {
 849                                 while (rte_event_ring_enqueue_burst(dest_ring,
 850                                                                     e, 1,
 851                                                                     NULL) != 1)
 852                                         rte_pause();
 853                         } else {
 854                                 uint16_t last_idx = source_port->in_buffer_len;
 855                                 source_port->in_buffer[last_idx] = *e;
 856                                 source_port->in_buffer_len++;
 857                         }
 858                 }
 859
 860                 events_left -= in_len;
 861         }
 862 }
 863
 864 static void
 865 dsw_port_move_migrating_flow(struct dsw_evdev *dsw,
 866                              struct dsw_port *source_port)
 867 {
 868         uint8_t queue_id = source_port->migration_target_qf.queue_id;
 869         uint16_t flow_hash = source_port->migration_target_qf.flow_hash;
 870         uint8_t dest_port_id = source_port->migration_target_port_id;
 871         struct dsw_port *dest_port = &dsw->ports[dest_port_id];
 872
 873         dsw_port_flush_out_buffers(dsw, source_port);
 874
 875         rte_smp_wmb();
 876
 877         dsw->queues[queue_id].flow_to_port_map[flow_hash] =
 878                 dest_port_id;
 879
 880         dsw_port_forward_migrated_flow(source_port, dest_port->in_ring,
 881                                        queue_id, flow_hash);
 882
 883         /* Flow table update and migration destination port's enqueues
 884          * must be seen before the control message.
 885          */
 886         rte_smp_wmb();
 887
 888         dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_UNPAUS_REQ, queue_id,
 889                                flow_hash);
 890         source_port->cfm_cnt = 0;
 891         source_port->migration_state = DSW_MIGRATION_STATE_UNPAUSING;
 892 }
 893
 894 static void
 895 dsw_port_handle_confirm(struct dsw_evdev *dsw, struct dsw_port *port)
 896 {
 897         port->cfm_cnt++;
 898
 899         if (port->cfm_cnt == (dsw->num_ports-1)) {
 900                 switch (port->migration_state) {
 901                 case DSW_MIGRATION_STATE_PAUSING:
 902                         DSW_LOG_DP_PORT(DEBUG, port->id, "Going into forwarding "
 903                                         "migration state.\n");
 904                         port->migration_state = DSW_MIGRATION_STATE_FORWARDING;
 905                         break;
 906                 case DSW_MIGRATION_STATE_UNPAUSING:
 907                         dsw_port_end_migration(dsw, port);
 908                         break;
 909                 default:
 910                         RTE_ASSERT(0);
 911                         break;
 912                 }
 913         }
 914 }
 915
 916 static void
 917 dsw_port_ctl_process(struct dsw_evdev *dsw, struct dsw_port *port)
 918 {
 919         struct dsw_ctl_msg msg;
 920
 921         /* So any table loads happens before the ring dequeue, in the
 922          * case of a 'paus' message.
 923          */
 924         rte_smp_rmb();
 925
 926         if (dsw_port_ctl_dequeue(port, &msg) == 0) {
 927                 switch (msg.type) {
 928                 case DSW_CTL_PAUS_REQ:
 929                         dsw_port_handle_pause_flow(dsw, port,
 930                                                    msg.originating_port_id,
 931                                                    msg.queue_id, msg.flow_hash);
 932                         break;
 933                 case DSW_CTL_UNPAUS_REQ:
 934                         dsw_port_handle_unpause_flow(dsw, port,
 935                                                      msg.originating_port_id,
 936                                                      msg.queue_id,
 937                                                      msg.flow_hash);
 938                         break;
 939                 case DSW_CTL_CFM:
 940                         dsw_port_handle_confirm(dsw, port);
 941                         break;
 942                 }
 943         }
 944 }
 945
 946 static void
 947 dsw_port_note_op(struct dsw_port *port, uint16_t num_events)
 948 {
 949         /* To pull the control ring reasonbly often on busy ports,
 950          * each dequeued/enqueued event is considered an 'op' too.
 951          */
 952         port->ops_since_bg_task += (num_events+1);
 953 }
 954
 955 static void
 956 dsw_port_bg_process(struct dsw_evdev *dsw, struct dsw_port *port)
 957 {
 958         if (unlikely(port->migration_state == DSW_MIGRATION_STATE_FORWARDING &&
 959                      port->pending_releases == 0))
 960                 dsw_port_move_migrating_flow(dsw, port);
 961
 962         /* Polling the control ring is relatively inexpensive, and
 963          * polling it often helps bringing down migration latency, so
 964          * do this for every iteration.
 965          */
 966         dsw_port_ctl_process(dsw, port);
 967
 968         /* To avoid considering migration and flushing output buffers
 969          * on every dequeue/enqueue call, the scheduler only performs
 970          * such 'background' tasks every nth
 971          * (i.e. DSW_MAX_PORT_OPS_PER_BG_TASK) operation.
 972          */
 973         if (unlikely(port->ops_since_bg_task >= DSW_MAX_PORT_OPS_PER_BG_TASK)) {
 974                 uint64_t now;
 975
 976                 now = rte_get_timer_cycles();
 977
 978                 port->last_bg = now;
 979
 980                 /* Logic to avoid having events linger in the output
 981                  * buffer too long.
 982                  */
 983                 dsw_port_flush_out_buffers(dsw, port);
 984
 985                 dsw_port_consider_load_update(port, now);
 986
 987                 dsw_port_consider_migration(dsw, port, now);
 988
 989                 port->ops_since_bg_task = 0;
 990         }
 991 }
 992
 993 static void
 994 dsw_port_flush_out_buffers(struct dsw_evdev *dsw, struct dsw_port *source_port)
 995 {
 996         uint16_t dest_port_id;
 997
 998         for (dest_port_id = 0; dest_port_id < dsw->num_ports; dest_port_id++)
 999                 dsw_port_transmit_buffered(dsw, source_port, dest_port_id);
1000 }
1001
1002 uint16_t
1003 dsw_event_enqueue(void *port, const struct rte_event *ev)
1004 {
1005         return dsw_event_enqueue_burst(port, ev, unlikely(ev == NULL) ? 0 : 1);
1006 }
1007
1008 static __rte_always_inline uint16_t
1009 dsw_event_enqueue_burst_generic(struct dsw_port *source_port,
1010                                 const struct rte_event events[],
1011                                 uint16_t events_len, bool op_types_known,
1012                                 uint16_t num_new, uint16_t num_release,
1013                                 uint16_t num_non_release)
1014 {
1015         struct dsw_evdev *dsw = source_port->dsw;
1016         bool enough_credits;
1017         uint16_t i;
1018
1019         DSW_LOG_DP_PORT(DEBUG, source_port->id, "Attempting to enqueue %d "
1020                         "events to port %d.\n", events_len, source_port->id);
1021
1022         dsw_port_bg_process(dsw, source_port);
1023
1024         /* XXX: For performance (=ring efficiency) reasons, the
1025          * scheduler relies on internal non-ring buffers instead of
1026          * immediately sending the event to the destination ring. For
1027          * a producer that doesn't intend to produce or consume any
1028          * more events, the scheduler provides a way to flush the
1029          * buffer, by means of doing an enqueue of zero events. In
1030          * addition, a port cannot be left "unattended" (e.g. unused)
1031          * for long periods of time, since that would stall
1032          * migration. Eventdev API extensions to provide a cleaner way
1033          * to archieve both of these functions should be
1034          * considered.
1035          */
1036         if (unlikely(events_len == 0)) {
1037                 dsw_port_note_op(source_port, DSW_MAX_PORT_OPS_PER_BG_TASK);
1038                 dsw_port_flush_out_buffers(dsw, source_port);
1039                 return 0;
1040         }
1041
1042         dsw_port_note_op(source_port, events_len);
1043
1044         if (!op_types_known)
1045                 for (i = 0; i < events_len; i++) {
1046                         switch (events[i].op) {
1047                         case RTE_EVENT_OP_RELEASE:
1048                                 num_release++;
1049                                 break;
1050                         case RTE_EVENT_OP_NEW:
1051                                 num_new++;
1052                                 /* Falls through. */
1053                         default:
1054                                 num_non_release++;
1055                                 break;
1056                         }
1057                 }
1058
1059         /* Technically, we could allow the non-new events up to the
1060          * first new event in the array into the system, but for
1061          * simplicity reasons, we deny the whole burst if the port is
1062          * above the water mark.
1063          */
1064         if (unlikely(num_new > 0 && rte_atomic32_read(&dsw->credits_on_loan) >
1065                      source_port->new_event_threshold))
1066                 return 0;
1067
1068         enough_credits = dsw_port_acquire_credits(dsw, source_port,
1069                                                   num_non_release);
1070         if (unlikely(!enough_credits))
1071                 return 0;
1072
1073         source_port->pending_releases -= num_release;
1074
1075         dsw_port_enqueue_stats(source_port, num_new,
1076                                num_non_release-num_new, num_release);
1077
1078         for (i = 0; i < events_len; i++) {
1079                 const struct rte_event *event = &events[i];
1080
1081                 if (likely(num_release == 0 ||
1082                            event->op != RTE_EVENT_OP_RELEASE))
1083                         dsw_port_buffer_event(dsw, source_port, event);
1084                 dsw_port_queue_enqueue_stats(source_port, event->queue_id);
1085         }
1086
1087         DSW_LOG_DP_PORT(DEBUG, source_port->id, "%d non-release events "
1088                         "accepted.\n", num_non_release);
1089
1090         return num_non_release;
1091 }
1092
1093 uint16_t
1094 dsw_event_enqueue_burst(void *port, const struct rte_event events[],
1095                         uint16_t events_len)
1096 {
1097         struct dsw_port *source_port = port;
1098
1099         if (unlikely(events_len > source_port->enqueue_depth))
1100                 events_len = source_port->enqueue_depth;
1101
1102         return dsw_event_enqueue_burst_generic(source_port, events,
1103                                                events_len, false, 0, 0, 0);
1104 }
1105
1106 uint16_t
1107 dsw_event_enqueue_new_burst(void *port, const struct rte_event events[],
1108                             uint16_t events_len)
1109 {
1110         struct dsw_port *source_port = port;
1111
1112         if (unlikely(events_len > source_port->enqueue_depth))
1113                 events_len = source_port->enqueue_depth;
1114
1115         return dsw_event_enqueue_burst_generic(source_port, events,
1116                                                events_len, true, events_len,
1117                                                0, events_len);
1118 }
1119
1120 uint16_t
1121 dsw_event_enqueue_forward_burst(void *port, const struct rte_event events[],
1122                                 uint16_t events_len)
1123 {
1124         struct dsw_port *source_port = port;
1125
1126         if (unlikely(events_len > source_port->enqueue_depth))
1127                 events_len = source_port->enqueue_depth;
1128
1129         return dsw_event_enqueue_burst_generic(source_port, events,
1130                                                events_len, true, 0, 0,
1131                                                events_len);
1132 }
1133
1134 uint16_t
1135 dsw_event_dequeue(void *port, struct rte_event *events, uint64_t wait)
1136 {
1137         return dsw_event_dequeue_burst(port, events, 1, wait);
1138 }
1139
1140 static void
1141 dsw_port_record_seen_events(struct dsw_port *port, struct rte_event *events,
1142                             uint16_t num)
1143 {
1144         uint16_t i;
1145
1146         dsw_port_dequeue_stats(port, num);
1147
1148         for (i = 0; i < num; i++) {
1149                 uint16_t l_idx = port->seen_events_idx;
1150                 struct dsw_queue_flow *qf = &port->seen_events[l_idx];
1151                 struct rte_event *event = &events[i];
1152                 qf->queue_id = event->queue_id;
1153                 qf->flow_hash = dsw_flow_id_hash(event->flow_id);
1154
1155                 port->seen_events_idx = (l_idx+1) % DSW_MAX_EVENTS_RECORDED;
1156
1157                 dsw_port_queue_dequeued_stats(port, event->queue_id);
1158         }
1159
1160         if (unlikely(port->seen_events_len != DSW_MAX_EVENTS_RECORDED))
1161                 port->seen_events_len =
1162                         RTE_MIN(port->seen_events_len + num,
1163                                 DSW_MAX_EVENTS_RECORDED);
1164 }
1165
1166 #ifdef DSW_SORT_DEQUEUED
1167
1168 #define DSW_EVENT_TO_INT(_event)                                \
1169         ((int)((((_event)->queue_id)<<16)|((_event)->flow_id)))
1170
1171 static inline int
1172 dsw_cmp_event(const void *v_event_a, const void *v_event_b)
1173 {
1174         const struct rte_event *event_a = v_event_a;
1175         const struct rte_event *event_b = v_event_b;
1176
1177         return DSW_EVENT_TO_INT(event_a) - DSW_EVENT_TO_INT(event_b);
1178 }
1179 #endif
1180
1181 static uint16_t
1182 dsw_port_dequeue_burst(struct dsw_port *port, struct rte_event *events,
1183                        uint16_t num)
1184 {
1185         struct dsw_port *source_port = port;
1186         struct dsw_evdev *dsw = source_port->dsw;
1187
1188         dsw_port_ctl_process(dsw, source_port);
1189
1190         if (unlikely(port->in_buffer_len > 0)) {
1191                 uint16_t dequeued = RTE_MIN(num, port->in_buffer_len);
1192
1193                 rte_memcpy(events, &port->in_buffer[port->in_buffer_start],
1194                            dequeued * sizeof(struct rte_event));
1195
1196                 port->in_buffer_start += dequeued;
1197                 port->in_buffer_len -= dequeued;
1198
1199                 if (port->in_buffer_len == 0)
1200                         port->in_buffer_start = 0;
1201
1202                 return dequeued;
1203         }
1204
1205         return rte_event_ring_dequeue_burst(port->in_ring, events, num, NULL);
1206 }
1207
1208 uint16_t
1209 dsw_event_dequeue_burst(void *port, struct rte_event *events, uint16_t num,
1210                         uint64_t wait __rte_unused)
1211 {
1212         struct dsw_port *source_port = port;
1213         struct dsw_evdev *dsw = source_port->dsw;
1214         uint16_t dequeued;
1215
1216         source_port->pending_releases = 0;
1217
1218         dsw_port_bg_process(dsw, source_port);
1219
1220         if (unlikely(num > source_port->dequeue_depth))
1221                 num = source_port->dequeue_depth;
1222
1223         dequeued = dsw_port_dequeue_burst(source_port, events, num);
1224
1225         source_port->pending_releases = dequeued;
1226
1227         dsw_port_load_record(source_port, dequeued);
1228
1229         dsw_port_note_op(source_port, dequeued);
1230
1231         if (dequeued > 0) {
1232                 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Dequeued %d events.\n",
1233                                 dequeued);
1234
1235                 dsw_port_return_credits(dsw, source_port, dequeued);
1236
1237                 /* One potential optimization one might think of is to
1238                  * add a migration state (prior to 'pausing'), and
1239                  * only record seen events when the port is in this
1240                  * state (and transit to 'pausing' when enough events
1241                  * have been gathered). However, that schema doesn't
1242                  * seem to improve performance.
1243                  */
1244                 dsw_port_record_seen_events(port, events, dequeued);
1245         }
1246         /* XXX: Assuming the port can't produce any more work,
1247          *      consider flushing the output buffer, on dequeued ==
1248          *      0.
1249          */
1250
1251 #ifdef DSW_SORT_DEQUEUED
1252         dsw_stable_sort(events, dequeued, sizeof(events[0]), dsw_cmp_event);
1253 #endif
1254
1255         return dequeued;
1256 }