drivers/gpu/drm/amd/scheduler/gpu_scheduler.c

   1 /*
   2  * Copyright 2015 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  *
  23  */
  24 #include <linux/kthread.h>
  25 #include <linux/wait.h>
  26 #include <linux/sched.h>
  27 #include <drm/drmP.h>
  28 #include "gpu_scheduler.h"
  29
  30 #define CREATE_TRACE_POINTS
  31 #include "gpu_sched_trace.h"
  32
  33 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity);
  34 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched);
  35 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb);
  36
  37 /* Initialize a given run queue struct */
  38 static void amd_sched_rq_init(struct amd_sched_rq *rq)
  39 {
  40         spin_lock_init(&rq->lock);
  41         INIT_LIST_HEAD(&rq->entities);
  42         rq->current_entity = NULL;
  43 }
  44
  45 static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
  46                                     struct amd_sched_entity *entity)
  47 {
  48         if (!list_empty(&entity->list))
  49                 return;
  50         spin_lock(&rq->lock);
  51         list_add_tail(&entity->list, &rq->entities);
  52         spin_unlock(&rq->lock);
  53 }
  54
  55 static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
  56                                        struct amd_sched_entity *entity)
  57 {
  58         if (list_empty(&entity->list))
  59                 return;
  60         spin_lock(&rq->lock);
  61         list_del_init(&entity->list);
  62         if (rq->current_entity == entity)
  63                 rq->current_entity = NULL;
  64         spin_unlock(&rq->lock);
  65 }
  66
  67 /**
  68  * Select an entity which could provide a job to run
  69  *
  70  * @rq          The run queue to check.
  71  *
  72  * Try to find a ready entity, returns NULL if none found.
  73  */
  74 static struct amd_sched_entity *
  75 amd_sched_rq_select_entity(struct amd_sched_rq *rq)
  76 {
  77         struct amd_sched_entity *entity;
  78
  79         spin_lock(&rq->lock);
  80
  81         entity = rq->current_entity;
  82         if (entity) {
  83                 list_for_each_entry_continue(entity, &rq->entities, list) {
  84                         if (amd_sched_entity_is_ready(entity)) {
  85                                 rq->current_entity = entity;
  86                                 spin_unlock(&rq->lock);
  87                                 return entity;
  88                         }
  89                 }
  90         }
  91
  92         list_for_each_entry(entity, &rq->entities, list) {
  93
  94                 if (amd_sched_entity_is_ready(entity)) {
  95                         rq->current_entity = entity;
  96                         spin_unlock(&rq->lock);
  97                         return entity;
  98                 }
  99
 100                 if (entity == rq->current_entity)
 101                         break;
 102         }
 103
 104         spin_unlock(&rq->lock);
 105
 106         return NULL;
 107 }
 108
 109 /**
 110  * Init a context entity used by scheduler when submit to HW ring.
 111  *
 112  * @sched       The pointer to the scheduler
 113  * @entity      The pointer to a valid amd_sched_entity
 114  * @rq          The run queue this entity belongs
 115  * @kernel      If this is an entity for the kernel
 116  * @jobs        The max number of jobs in the job queue
 117  *
 118  * return 0 if succeed. negative error code on failure
 119 */
 120 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 121                           struct amd_sched_entity *entity,
 122                           struct amd_sched_rq *rq,
 123                           uint32_t jobs)
 124 {
 125         int r;
 126
 127         if (!(sched && entity && rq))
 128                 return -EINVAL;
 129
 130         memset(entity, 0, sizeof(struct amd_sched_entity));
 131         INIT_LIST_HEAD(&entity->list);
 132         entity->rq = rq;
 133         entity->sched = sched;
 134
 135         spin_lock_init(&entity->queue_lock);
 136         r = kfifo_alloc(&entity->job_queue, jobs * sizeof(void *), GFP_KERNEL);
 137         if (r)
 138                 return r;
 139
 140         atomic_set(&entity->fence_seq, 0);
 141         entity->fence_context = fence_context_alloc(2);
 142
 143         return 0;
 144 }
 145
 146 /**
 147  * Query if entity is initialized
 148  *
 149  * @sched       Pointer to scheduler instance
 150  * @entity      The pointer to a valid scheduler entity
 151  *
 152  * return true if entity is initialized, false otherwise
 153 */
 154 static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
 155                                             struct amd_sched_entity *entity)
 156 {
 157         return entity->sched == sched &&
 158                 entity->rq != NULL;
 159 }
 160
 161 /**
 162  * Check if entity is idle
 163  *
 164  * @entity      The pointer to a valid scheduler entity
 165  *
 166  * Return true if entity don't has any unscheduled jobs.
 167  */
 168 static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
 169 {
 170         rmb();
 171         if (kfifo_is_empty(&entity->job_queue))
 172                 return true;
 173
 174         return false;
 175 }
 176
 177 /**
 178  * Check if entity is ready
 179  *
 180  * @entity      The pointer to a valid scheduler entity
 181  *
 182  * Return true if entity could provide a job.
 183  */
 184 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity)
 185 {
 186         if (kfifo_is_empty(&entity->job_queue))
 187                 return false;
 188
 189         if (ACCESS_ONCE(entity->dependency))
 190                 return false;
 191
 192         return true;
 193 }
 194
 195 /**
 196  * Destroy a context entity
 197  *
 198  * @sched       Pointer to scheduler instance
 199  * @entity      The pointer to a valid scheduler entity
 200  *
 201  * Cleanup and free the allocated resources.
 202  */
 203 void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
 204                            struct amd_sched_entity *entity)
 205 {
 206         struct amd_sched_rq *rq = entity->rq;
 207
 208         if (!amd_sched_entity_is_initialized(sched, entity))
 209                 return;
 210
 211         /**
 212          * The client will not queue more IBs during this fini, consume existing
 213          * queued IBs
 214         */
 215         wait_event(sched->job_scheduled, amd_sched_entity_is_idle(entity));
 216
 217         amd_sched_rq_remove_entity(rq, entity);
 218         kfifo_free(&entity->job_queue);
 219 }
 220
 221 static void amd_sched_entity_wakeup(struct fence *f, struct fence_cb *cb)
 222 {
 223         struct amd_sched_entity *entity =
 224                 container_of(cb, struct amd_sched_entity, cb);
 225         entity->dependency = NULL;
 226         fence_put(f);
 227         amd_sched_wakeup(entity->sched);
 228 }
 229
 230 static void amd_sched_entity_clear_dep(struct fence *f, struct fence_cb *cb)
 231 {
 232         struct amd_sched_entity *entity =
 233                 container_of(cb, struct amd_sched_entity, cb);
 234         entity->dependency = NULL;
 235         fence_put(f);
 236 }
 237
 238 static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
 239 {
 240         struct amd_gpu_scheduler *sched = entity->sched;
 241         struct fence * fence = entity->dependency;
 242         struct amd_sched_fence *s_fence;
 243
 244         if (fence->context == entity->fence_context) {
 245                 /* We can ignore fences from ourself */
 246                 fence_put(entity->dependency);
 247                 return false;
 248         }
 249
 250         s_fence = to_amd_sched_fence(fence);
 251         if (s_fence && s_fence->sched == sched) {
 252
 253                 /*
 254                  * Fence is from the same scheduler, only need to wait for
 255                  * it to be scheduled
 256                  */
 257                 fence = fence_get(&s_fence->scheduled);
 258                 fence_put(entity->dependency);
 259                 entity->dependency = fence;
 260                 if (!fence_add_callback(fence, &entity->cb,
 261                                         amd_sched_entity_clear_dep))
 262                         return true;
 263
 264                 /* Ignore it when it is already scheduled */
 265                 fence_put(fence);
 266                 return false;
 267         }
 268
 269         if (!fence_add_callback(entity->dependency, &entity->cb,
 270                                 amd_sched_entity_wakeup))
 271                 return true;
 272
 273         fence_put(entity->dependency);
 274         return false;
 275 }
 276
 277 static struct amd_sched_job *
 278 amd_sched_entity_pop_job(struct amd_sched_entity *entity)
 279 {
 280         struct amd_gpu_scheduler *sched = entity->sched;
 281         struct amd_sched_job *sched_job;
 282
 283         if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job)))
 284                 return NULL;
 285
 286         while ((entity->dependency = sched->ops->dependency(sched_job)))
 287                 if (amd_sched_entity_add_dependency_cb(entity))
 288                         return NULL;
 289
 290         return sched_job;
 291 }
 292
 293 /**
 294  * Helper to submit a job to the job queue
 295  *
 296  * @sched_job           The pointer to job required to submit
 297  *
 298  * Returns true if we could submit the job.
 299  */
 300 static bool amd_sched_entity_in(struct amd_sched_job *sched_job)
 301 {
 302         struct amd_gpu_scheduler *sched = sched_job->sched;
 303         struct amd_sched_entity *entity = sched_job->s_entity;
 304         bool added, first = false;
 305
 306         spin_lock(&entity->queue_lock);
 307         added = kfifo_in(&entity->job_queue, &sched_job,
 308                         sizeof(sched_job)) == sizeof(sched_job);
 309
 310         if (added && kfifo_len(&entity->job_queue) == sizeof(sched_job))
 311                 first = true;
 312
 313         spin_unlock(&entity->queue_lock);
 314
 315         /* first job wakes up scheduler */
 316         if (first) {
 317                 /* Add the entity to the run queue */
 318                 amd_sched_rq_add_entity(entity->rq, entity);
 319                 amd_sched_wakeup(sched);
 320         }
 321         return added;
 322 }
 323
 324 /* job_finish is called after hw fence signaled, and
 325  * the job had already been deleted from ring_mirror_list
 326  */
 327 static void amd_sched_job_finish(struct work_struct *work)
 328 {
 329         struct amd_sched_job *s_job = container_of(work, struct amd_sched_job,
 330                                                    finish_work);
 331         struct amd_gpu_scheduler *sched = s_job->sched;
 332
 333         /* remove job from ring_mirror_list */
 334         spin_lock(&sched->job_list_lock);
 335         list_del_init(&s_job->node);
 336         if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
 337                 struct amd_sched_job *next;
 338
 339                 spin_unlock(&sched->job_list_lock);
 340                 cancel_delayed_work_sync(&s_job->work_tdr);
 341                 spin_lock(&sched->job_list_lock);
 342
 343                 /* queue TDR for next job */
 344                 next = list_first_entry_or_null(&sched->ring_mirror_list,
 345                                                 struct amd_sched_job, node);
 346
 347                 if (next)
 348                         schedule_delayed_work(&next->work_tdr, sched->timeout);
 349         }
 350         spin_unlock(&sched->job_list_lock);
 351         sched->ops->free_job(s_job);
 352 }
 353
 354 static void amd_sched_job_finish_cb(struct fence *f, struct fence_cb *cb)
 355 {
 356         struct amd_sched_job *job = container_of(cb, struct amd_sched_job,
 357                                                  finish_cb);
 358         schedule_work(&job->finish_work);
 359 }
 360
 361 static void amd_sched_job_begin(struct amd_sched_job *s_job)
 362 {
 363         struct amd_gpu_scheduler *sched = s_job->sched;
 364
 365         spin_lock(&sched->job_list_lock);
 366         list_add_tail(&s_job->node, &sched->ring_mirror_list);
 367         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 368             list_first_entry_or_null(&sched->ring_mirror_list,
 369                                      struct amd_sched_job, node) == s_job)
 370                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
 371         spin_unlock(&sched->job_list_lock);
 372 }
 373
 374 static void amd_sched_job_timedout(struct work_struct *work)
 375 {
 376         struct amd_sched_job *job = container_of(work, struct amd_sched_job,
 377                                                  work_tdr.work);
 378
 379         job->sched->ops->timedout_job(job);
 380 }
 381
 382 void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched)
 383 {
 384         struct amd_sched_job *s_job;
 385
 386         spin_lock(&sched->job_list_lock);
 387         list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
 388                 if (fence_remove_callback(s_job->s_fence->parent, &s_job->s_fence->cb)) {
 389                         fence_put(s_job->s_fence->parent);
 390                         s_job->s_fence->parent = NULL;
 391                 }
 392         }
 393         atomic_set(&sched->hw_rq_count, 0);
 394         spin_unlock(&sched->job_list_lock);
 395 }
 396
 397 void amd_sched_job_recovery(struct amd_gpu_scheduler *sched)
 398 {
 399         struct amd_sched_job *s_job, *tmp;
 400         int r;
 401
 402         spin_lock(&sched->job_list_lock);
 403         s_job = list_first_entry_or_null(&sched->ring_mirror_list,
 404                                          struct amd_sched_job, node);
 405         if (s_job && sched->timeout != MAX_SCHEDULE_TIMEOUT)
 406                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
 407
 408         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 409                 struct amd_sched_fence *s_fence = s_job->s_fence;
 410                 struct fence *fence;
 411
 412                 spin_unlock(&sched->job_list_lock);
 413                 fence = sched->ops->run_job(s_job);
 414                 atomic_inc(&sched->hw_rq_count);
 415                 if (fence) {
 416                         s_fence->parent = fence_get(fence);
 417                         r = fence_add_callback(fence, &s_fence->cb,
 418                                                amd_sched_process_job);
 419                         if (r == -ENOENT)
 420                                 amd_sched_process_job(fence, &s_fence->cb);
 421                         else if (r)
 422                                 DRM_ERROR("fence add callback failed (%d)\n",
 423                                           r);
 424                         fence_put(fence);
 425                 } else {
 426                         DRM_ERROR("Failed to run job!\n");
 427                         amd_sched_process_job(NULL, &s_fence->cb);
 428                 }
 429                 spin_lock(&sched->job_list_lock);
 430         }
 431         spin_unlock(&sched->job_list_lock);
 432 }
 433
 434 /**
 435  * Submit a job to the job queue
 436  *
 437  * @sched_job           The pointer to job required to submit
 438  *
 439  * Returns 0 for success, negative error code otherwise.
 440  */
 441 void amd_sched_entity_push_job(struct amd_sched_job *sched_job)
 442 {
 443         struct amd_sched_entity *entity = sched_job->s_entity;
 444
 445         trace_amd_sched_job(sched_job);
 446         fence_add_callback(&sched_job->s_fence->finished, &sched_job->finish_cb,
 447                            amd_sched_job_finish_cb);
 448         wait_event(entity->sched->job_scheduled,
 449                    amd_sched_entity_in(sched_job));
 450 }
 451
 452 /* init a sched_job with basic field */
 453 int amd_sched_job_init(struct amd_sched_job *job,
 454                        struct amd_gpu_scheduler *sched,
 455                        struct amd_sched_entity *entity,
 456                        void *owner)
 457 {
 458         job->sched = sched;
 459         job->s_entity = entity;
 460         job->s_fence = amd_sched_fence_create(entity, owner);
 461         if (!job->s_fence)
 462                 return -ENOMEM;
 463
 464         INIT_WORK(&job->finish_work, amd_sched_job_finish);
 465         INIT_LIST_HEAD(&job->node);
 466         INIT_DELAYED_WORK(&job->work_tdr, amd_sched_job_timedout);
 467
 468         return 0;
 469 }
 470
 471 /**
 472  * Return ture if we can push more jobs to the hw.
 473  */
 474 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
 475 {
 476         return atomic_read(&sched->hw_rq_count) <
 477                 sched->hw_submission_limit;
 478 }
 479
 480 /**
 481  * Wake up the scheduler when it is ready
 482  */
 483 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched)
 484 {
 485         if (amd_sched_ready(sched))
 486                 wake_up_interruptible(&sched->wake_up_worker);
 487 }
 488
 489 /**
 490  * Select next entity to process
 491 */
 492 static struct amd_sched_entity *
 493 amd_sched_select_entity(struct amd_gpu_scheduler *sched)
 494 {
 495         struct amd_sched_entity *entity;
 496         int i;
 497
 498         if (!amd_sched_ready(sched))
 499                 return NULL;
 500
 501         /* Kernel run queue has higher priority than normal run queue*/
 502         for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++) {
 503                 entity = amd_sched_rq_select_entity(&sched->sched_rq[i]);
 504                 if (entity)
 505                         break;
 506         }
 507
 508         return entity;
 509 }
 510
 511 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 512 {
 513         struct amd_sched_fence *s_fence =
 514                 container_of(cb, struct amd_sched_fence, cb);
 515         struct amd_gpu_scheduler *sched = s_fence->sched;
 516
 517         atomic_dec(&sched->hw_rq_count);
 518         amd_sched_fence_finished(s_fence);
 519
 520         trace_amd_sched_process_job(s_fence);
 521         fence_put(&s_fence->finished);
 522         wake_up_interruptible(&sched->wake_up_worker);
 523 }
 524
 525 static bool amd_sched_blocked(struct amd_gpu_scheduler *sched)
 526 {
 527         if (kthread_should_park()) {
 528                 kthread_parkme();
 529                 return true;
 530         }
 531
 532         return false;
 533 }
 534
 535 static int amd_sched_main(void *param)
 536 {
 537         struct sched_param sparam = {.sched_priority = 1};
 538         struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
 539         int r, count;
 540
 541         sched_setscheduler(current, SCHED_FIFO, &sparam);
 542
 543         while (!kthread_should_stop()) {
 544                 struct amd_sched_entity *entity = NULL;
 545                 struct amd_sched_fence *s_fence;
 546                 struct amd_sched_job *sched_job;
 547                 struct fence *fence;
 548
 549                 wait_event_interruptible(sched->wake_up_worker,
 550                                          (!amd_sched_blocked(sched) &&
 551                                           (entity = amd_sched_select_entity(sched))) ||
 552                                          kthread_should_stop());
 553
 554                 if (!entity)
 555                         continue;
 556
 557                 sched_job = amd_sched_entity_pop_job(entity);
 558                 if (!sched_job)
 559                         continue;
 560
 561                 s_fence = sched_job->s_fence;
 562
 563                 atomic_inc(&sched->hw_rq_count);
 564                 amd_sched_job_begin(sched_job);
 565
 566                 fence = sched->ops->run_job(sched_job);
 567                 amd_sched_fence_scheduled(s_fence);
 568                 if (fence) {
 569                         s_fence->parent = fence_get(fence);
 570                         r = fence_add_callback(fence, &s_fence->cb,
 571                                                amd_sched_process_job);
 572                         if (r == -ENOENT)
 573                                 amd_sched_process_job(fence, &s_fence->cb);
 574                         else if (r)
 575                                 DRM_ERROR("fence add callback failed (%d)\n",
 576                                           r);
 577                         fence_put(fence);
 578                 } else {
 579                         DRM_ERROR("Failed to run job!\n");
 580                         amd_sched_process_job(NULL, &s_fence->cb);
 581                 }
 582
 583                 count = kfifo_out(&entity->job_queue, &sched_job,
 584                                 sizeof(sched_job));
 585                 WARN_ON(count != sizeof(sched_job));
 586                 wake_up(&sched->job_scheduled);
 587         }
 588         return 0;
 589 }
 590
 591 /**
 592  * Init a gpu scheduler instance
 593  *
 594  * @sched               The pointer to the scheduler
 595  * @ops                 The backend operations for this scheduler.
 596  * @hw_submissions      Number of hw submissions to do.
 597  * @name                Name used for debugging
 598  *
 599  * Return 0 on success, otherwise error code.
 600 */
 601 int amd_sched_init(struct amd_gpu_scheduler *sched,
 602                    const struct amd_sched_backend_ops *ops,
 603                    unsigned hw_submission, long timeout, const char *name)
 604 {
 605         int i;
 606         sched->ops = ops;
 607         sched->hw_submission_limit = hw_submission;
 608         sched->name = name;
 609         sched->timeout = timeout;
 610         for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++)
 611                 amd_sched_rq_init(&sched->sched_rq[i]);
 612
 613         init_waitqueue_head(&sched->wake_up_worker);
 614         init_waitqueue_head(&sched->job_scheduled);
 615         INIT_LIST_HEAD(&sched->ring_mirror_list);
 616         spin_lock_init(&sched->job_list_lock);
 617         atomic_set(&sched->hw_rq_count, 0);
 618
 619         /* Each scheduler will run on a seperate kernel thread */
 620         sched->thread = kthread_run(amd_sched_main, sched, sched->name);
 621         if (IS_ERR(sched->thread)) {
 622                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 623                 return PTR_ERR(sched->thread);
 624         }
 625
 626         return 0;
 627 }
 628
 629 /**
 630  * Destroy a gpu scheduler
 631  *
 632  * @sched       The pointer to the scheduler
 633  */
 634 void amd_sched_fini(struct amd_gpu_scheduler *sched)
 635 {
 636         if (sched->thread)
 637                 kthread_stop(sched->thread);
 638 }