drivers/gpu/drm/scheduler/sched_main.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 /**
  25  * DOC: Overview
  26  *
  27  * The GPU scheduler provides entities which allow userspace to push jobs
  28  * into software queues which are then scheduled on a hardware run queue.
  29  * The software queues have a priority among them. The scheduler selects the entities
  30  * from the run queue using a FIFO. The scheduler provides dependency handling
  31  * features among jobs. The driver is supposed to provide callback functions for
  32  * backend operations to the scheduler like submitting a job to hardware run queue,
  33  * returning the dependencies of a job etc.
  34  *
  35  * The organisation of the scheduler is the following:
  36  *
  37  * 1. Each hw run queue has one scheduler
  38  * 2. Each scheduler has multiple run queues with different priorities
  39  *    (e.g., HIGH_HW,HIGH_SW, KERNEL, NORMAL)
  40  * 3. Each scheduler run queue has a queue of entities to schedule
  41  * 4. Entities themselves maintain a queue of jobs that will be scheduled on
  42  *    the hardware.
  43  *
  44  * The jobs in a entity are always scheduled in the order that they were pushed.
  45  */
  46
  47 #include <linux/kthread.h>
  48 #include <linux/wait.h>
  49 #include <linux/sched.h>
  50 #include <uapi/linux/sched/types.h>
  51 #include <drm/drmP.h>
  52 #include <drm/gpu_scheduler.h>
  53 #include <drm/spsc_queue.h>
  54
  55 #define CREATE_TRACE_POINTS
  56 #include "gpu_scheduler_trace.h"
  57
  58 #define to_drm_sched_job(sched_job)             \
  59                 container_of((sched_job), struct drm_sched_job, queue_node)
  60
  61 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
  62
  63 /**
  64  * drm_sched_rq_init - initialize a given run queue struct
  65  *
  66  * @rq: scheduler run queue
  67  *
  68  * Initializes a scheduler runqueue.
  69  */
  70 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
  71                               struct drm_sched_rq *rq)
  72 {
  73         spin_lock_init(&rq->lock);
  74         INIT_LIST_HEAD(&rq->entities);
  75         rq->current_entity = NULL;
  76         rq->sched = sched;
  77 }
  78
  79 /**
  80  * drm_sched_rq_add_entity - add an entity
  81  *
  82  * @rq: scheduler run queue
  83  * @entity: scheduler entity
  84  *
  85  * Adds a scheduler entity to the run queue.
  86  */
  87 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
  88                              struct drm_sched_entity *entity)
  89 {
  90         if (!list_empty(&entity->list))
  91                 return;
  92         spin_lock(&rq->lock);
  93         list_add_tail(&entity->list, &rq->entities);
  94         spin_unlock(&rq->lock);
  95 }
  96
  97 /**
  98  * drm_sched_rq_remove_entity - remove an entity
  99  *
 100  * @rq: scheduler run queue
 101  * @entity: scheduler entity
 102  *
 103  * Removes a scheduler entity from the run queue.
 104  */
 105 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 106                                 struct drm_sched_entity *entity)
 107 {
 108         if (list_empty(&entity->list))
 109                 return;
 110         spin_lock(&rq->lock);
 111         list_del_init(&entity->list);
 112         if (rq->current_entity == entity)
 113                 rq->current_entity = NULL;
 114         spin_unlock(&rq->lock);
 115 }
 116
 117 /**
 118  * drm_sched_rq_select_entity - Select an entity which could provide a job to run
 119  *
 120  * @rq: scheduler run queue to check.
 121  *
 122  * Try to find a ready entity, returns NULL if none found.
 123  */
 124 static struct drm_sched_entity *
 125 drm_sched_rq_select_entity(struct drm_sched_rq *rq)
 126 {
 127         struct drm_sched_entity *entity;
 128
 129         spin_lock(&rq->lock);
 130
 131         entity = rq->current_entity;
 132         if (entity) {
 133                 list_for_each_entry_continue(entity, &rq->entities, list) {
 134                         if (drm_sched_entity_is_ready(entity)) {
 135                                 rq->current_entity = entity;
 136                                 spin_unlock(&rq->lock);
 137                                 return entity;
 138                         }
 139                 }
 140         }
 141
 142         list_for_each_entry(entity, &rq->entities, list) {
 143
 144                 if (drm_sched_entity_is_ready(entity)) {
 145                         rq->current_entity = entity;
 146                         spin_unlock(&rq->lock);
 147                         return entity;
 148                 }
 149
 150                 if (entity == rq->current_entity)
 151                         break;
 152         }
 153
 154         spin_unlock(&rq->lock);
 155
 156         return NULL;
 157 }
 158
 159 /**
 160  * drm_sched_dependency_optimized
 161  *
 162  * @fence: the dependency fence
 163  * @entity: the entity which depends on the above fence
 164  *
 165  * Returns true if the dependency can be optimized and false otherwise
 166  */
 167 bool drm_sched_dependency_optimized(struct dma_fence* fence,
 168                                     struct drm_sched_entity *entity)
 169 {
 170         struct drm_gpu_scheduler *sched = entity->rq->sched;
 171         struct drm_sched_fence *s_fence;
 172
 173         if (!fence || dma_fence_is_signaled(fence))
 174                 return false;
 175         if (fence->context == entity->fence_context)
 176                 return true;
 177         s_fence = to_drm_sched_fence(fence);
 178         if (s_fence && s_fence->sched == sched)
 179                 return true;
 180
 181         return false;
 182 }
 183 EXPORT_SYMBOL(drm_sched_dependency_optimized);
 184
 185 /**
 186  * drm_sched_start_timeout - start timeout for reset worker
 187  *
 188  * @sched: scheduler instance to start the worker for
 189  *
 190  * Start the timeout for the given scheduler.
 191  */
 192 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
 193 {
 194         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 195             !list_empty(&sched->ring_mirror_list))
 196                 schedule_delayed_work(&sched->work_tdr, sched->timeout);
 197 }
 198
 199 /**
 200  * drm_sched_fault - immediately start timeout handler
 201  *
 202  * @sched: scheduler where the timeout handling should be started.
 203  *
 204  * Start timeout handling immediately when the driver detects a hardware fault.
 205  */
 206 void drm_sched_fault(struct drm_gpu_scheduler *sched)
 207 {
 208         mod_delayed_work(system_wq, &sched->work_tdr, 0);
 209 }
 210 EXPORT_SYMBOL(drm_sched_fault);
 211
 212 /**
 213  * drm_sched_suspend_timeout - Suspend scheduler job timeout
 214  *
 215  * @sched: scheduler instance for which to suspend the timeout
 216  *
 217  * Suspend the delayed work timeout for the scheduler. This is done by
 218  * modifying the delayed work timeout to an arbitrary large value,
 219  * MAX_SCHEDULE_TIMEOUT in this case. Note that this function can be
 220  * called from an IRQ context.
 221  *
 222  * Returns the timeout remaining
 223  *
 224  */
 225 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
 226 {
 227         unsigned long sched_timeout, now = jiffies;
 228
 229         sched_timeout = sched->work_tdr.timer.expires;
 230
 231         /*
 232          * Modify the timeout to an arbitrarily large value. This also prevents
 233          * the timeout to be restarted when new submissions arrive
 234          */
 235         if (mod_delayed_work(system_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
 236                         && time_after(sched_timeout, now))
 237                 return sched_timeout - now;
 238         else
 239                 return sched->timeout;
 240 }
 241 EXPORT_SYMBOL(drm_sched_suspend_timeout);
 242
 243 /**
 244  * drm_sched_resume_timeout - Resume scheduler job timeout
 245  *
 246  * @sched: scheduler instance for which to resume the timeout
 247  * @remaining: remaining timeout
 248  *
 249  * Resume the delayed work timeout for the scheduler. Note that
 250  * this function can be called from an IRQ context.
 251  */
 252 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
 253                 unsigned long remaining)
 254 {
 255         unsigned long flags;
 256
 257         spin_lock_irqsave(&sched->job_list_lock, flags);
 258
 259         if (list_empty(&sched->ring_mirror_list))
 260                 cancel_delayed_work(&sched->work_tdr);
 261         else
 262                 mod_delayed_work(system_wq, &sched->work_tdr, remaining);
 263
 264         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 265 }
 266 EXPORT_SYMBOL(drm_sched_resume_timeout);
 267
 268 /* job_finish is called after hw fence signaled
 269  */
 270 static void drm_sched_job_finish(struct work_struct *work)
 271 {
 272         struct drm_sched_job *s_job = container_of(work, struct drm_sched_job,
 273                                                    finish_work);
 274         struct drm_gpu_scheduler *sched = s_job->sched;
 275         unsigned long flags;
 276
 277         /*
 278          * Canceling the timeout without removing our job from the ring mirror
 279          * list is safe, as we will only end up in this worker if our jobs
 280          * finished fence has been signaled. So even if some another worker
 281          * manages to find this job as the next job in the list, the fence
 282          * signaled check below will prevent the timeout to be restarted.
 283          */
 284         cancel_delayed_work_sync(&sched->work_tdr);
 285
 286         spin_lock_irqsave(&sched->job_list_lock, flags);
 287         /* queue TDR for next job */
 288         drm_sched_start_timeout(sched);
 289         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 290
 291         sched->ops->free_job(s_job);
 292 }
 293
 294 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 295 {
 296         struct drm_gpu_scheduler *sched = s_job->sched;
 297         unsigned long flags;
 298
 299         spin_lock_irqsave(&sched->job_list_lock, flags);
 300         list_add_tail(&s_job->node, &sched->ring_mirror_list);
 301         drm_sched_start_timeout(sched);
 302         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 303 }
 304
 305 static void drm_sched_job_timedout(struct work_struct *work)
 306 {
 307         struct drm_gpu_scheduler *sched;
 308         struct drm_sched_job *job;
 309         unsigned long flags;
 310
 311         sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 312         job = list_first_entry_or_null(&sched->ring_mirror_list,
 313                                        struct drm_sched_job, node);
 314
 315         if (job)
 316                 job->sched->ops->timedout_job(job);
 317
 318         spin_lock_irqsave(&sched->job_list_lock, flags);
 319         drm_sched_start_timeout(sched);
 320         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 321 }
 322
 323  /**
 324   * drm_sched_increase_karma - Update sched_entity guilty flag
 325   *
 326   * @bad: The job guilty of time out
 327   *
 328   * Increment on every hang caused by the 'bad' job. If this exceeds the hang
 329   * limit of the scheduler then the respective sched entity is marked guilty and
 330   * jobs from it will not be scheduled further
 331   */
 332 void drm_sched_increase_karma(struct drm_sched_job *bad)
 333 {
 334         int i;
 335         struct drm_sched_entity *tmp;
 336         struct drm_sched_entity *entity;
 337         struct drm_gpu_scheduler *sched = bad->sched;
 338
 339         /* don't increase @bad's karma if it's from KERNEL RQ,
 340          * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
 341          * corrupt but keep in mind that kernel jobs always considered good.
 342          */
 343         if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 344                 atomic_inc(&bad->karma);
 345                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
 346                      i++) {
 347                         struct drm_sched_rq *rq = &sched->sched_rq[i];
 348
 349                         spin_lock(&rq->lock);
 350                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
 351                                 if (bad->s_fence->scheduled.context ==
 352                                     entity->fence_context) {
 353                                         if (atomic_read(&bad->karma) >
 354                                             bad->sched->hang_limit)
 355                                                 if (entity->guilty)
 356                                                         atomic_set(entity->guilty, 1);
 357                                         break;
 358                                 }
 359                         }
 360                         spin_unlock(&rq->lock);
 361                         if (&entity->list != &rq->entities)
 362                                 break;
 363                 }
 364         }
 365 }
 366 EXPORT_SYMBOL(drm_sched_increase_karma);
 367
 368 /**
 369  * drm_sched_hw_job_reset - stop the scheduler if it contains the bad job
 370  *
 371  * @sched: scheduler instance
 372  * @bad: bad scheduler job
 373  *
 374  */
 375 void drm_sched_stop(struct drm_gpu_scheduler *sched)
 376 {
 377         struct drm_sched_job *s_job;
 378         unsigned long flags;
 379         struct dma_fence *last_fence =  NULL;
 380
 381         kthread_park(sched->thread);
 382
 383         /*
 384          * Verify all the signaled jobs in mirror list are removed from the ring
 385          * by waiting for the latest job to enter the list. This should insure that
 386          * also all the previous jobs that were in flight also already singaled
 387          * and removed from the list.
 388          */
 389         spin_lock_irqsave(&sched->job_list_lock, flags);
 390         list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
 391                 if (s_job->s_fence->parent &&
 392                     dma_fence_remove_callback(s_job->s_fence->parent,
 393                                               &s_job->cb)) {
 394                         dma_fence_put(s_job->s_fence->parent);
 395                         s_job->s_fence->parent = NULL;
 396                         atomic_dec(&sched->hw_rq_count);
 397                 } else {
 398                          last_fence = dma_fence_get(&s_job->s_fence->finished);
 399                          break;
 400                 }
 401         }
 402         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 403
 404         if (last_fence) {
 405                 dma_fence_wait(last_fence, false);
 406                 dma_fence_put(last_fence);
 407         }
 408 }
 409
 410 EXPORT_SYMBOL(drm_sched_stop);
 411
 412 /**
 413  * drm_sched_job_recovery - recover jobs after a reset
 414  *
 415  * @sched: scheduler instance
 416  *
 417  */
 418 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
 419 {
 420         struct drm_sched_job *s_job, *tmp;
 421         int r;
 422
 423         if (!full_recovery)
 424                 goto unpark;
 425
 426         /*
 427          * Locking the list is not required here as the sched thread is parked
 428          * so no new jobs are being pushed in to HW and in drm_sched_stop we
 429          * flushed all the jobs who were still in mirror list but who already
 430          * signaled and removed them self from the list. Also concurrent
 431          * GPU recovers can't run in parallel.
 432          */
 433         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 434                 struct dma_fence *fence = s_job->s_fence->parent;
 435
 436                 if (fence) {
 437                         r = dma_fence_add_callback(fence, &s_job->cb,
 438                                                    drm_sched_process_job);
 439                         if (r == -ENOENT)
 440                                 drm_sched_process_job(fence, &s_job->cb);
 441                         else if (r)
 442                                 DRM_ERROR("fence add callback failed (%d)\n",
 443                                           r);
 444                 } else
 445                         drm_sched_process_job(NULL, &s_job->cb);
 446         }
 447
 448         drm_sched_start_timeout(sched);
 449
 450 unpark:
 451         kthread_unpark(sched->thread);
 452 }
 453 EXPORT_SYMBOL(drm_sched_start);
 454
 455 /**
 456  * drm_sched_resubmit_jobs - helper to relunch job from mirror ring list
 457  *
 458  * @sched: scheduler instance
 459  *
 460  */
 461 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
 462 {
 463         struct drm_sched_job *s_job, *tmp;
 464         uint64_t guilty_context;
 465         bool found_guilty = false;
 466
 467         /*TODO DO we need spinlock here ? */
 468         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 469                 struct drm_sched_fence *s_fence = s_job->s_fence;
 470
 471                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 472                         found_guilty = true;
 473                         guilty_context = s_job->s_fence->scheduled.context;
 474                 }
 475
 476                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 477                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 478
 479                 s_job->s_fence->parent = sched->ops->run_job(s_job);
 480                 atomic_inc(&sched->hw_rq_count);
 481         }
 482 }
 483 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 484
 485 /**
 486  * drm_sched_job_init - init a scheduler job
 487  *
 488  * @job: scheduler job to init
 489  * @entity: scheduler entity to use
 490  * @owner: job owner for debugging
 491  *
 492  * Refer to drm_sched_entity_push_job() documentation
 493  * for locking considerations.
 494  *
 495  * Returns 0 for success, negative error code otherwise.
 496  */
 497 int drm_sched_job_init(struct drm_sched_job *job,
 498                        struct drm_sched_entity *entity,
 499                        void *owner)
 500 {
 501         struct drm_gpu_scheduler *sched;
 502
 503         drm_sched_entity_select_rq(entity);
 504         if (!entity->rq)
 505                 return -ENOENT;
 506
 507         sched = entity->rq->sched;
 508
 509         job->sched = sched;
 510         job->entity = entity;
 511         job->s_priority = entity->rq - sched->sched_rq;
 512         job->s_fence = drm_sched_fence_create(entity, owner);
 513         if (!job->s_fence)
 514                 return -ENOMEM;
 515         job->id = atomic64_inc_return(&sched->job_id_count);
 516
 517         INIT_WORK(&job->finish_work, drm_sched_job_finish);
 518         INIT_LIST_HEAD(&job->node);
 519
 520         return 0;
 521 }
 522 EXPORT_SYMBOL(drm_sched_job_init);
 523
 524 /**
 525  * drm_sched_job_cleanup - clean up scheduler job resources
 526  *
 527  * @job: scheduler job to clean up
 528  */
 529 void drm_sched_job_cleanup(struct drm_sched_job *job)
 530 {
 531         dma_fence_put(&job->s_fence->finished);
 532         job->s_fence = NULL;
 533 }
 534 EXPORT_SYMBOL(drm_sched_job_cleanup);
 535
 536 /**
 537  * drm_sched_ready - is the scheduler ready
 538  *
 539  * @sched: scheduler instance
 540  *
 541  * Return true if we can push more jobs to the hw, otherwise false.
 542  */
 543 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 544 {
 545         return atomic_read(&sched->hw_rq_count) <
 546                 sched->hw_submission_limit;
 547 }
 548
 549 /**
 550  * drm_sched_wakeup - Wake up the scheduler when it is ready
 551  *
 552  * @sched: scheduler instance
 553  *
 554  */
 555 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 556 {
 557         if (drm_sched_ready(sched))
 558                 wake_up_interruptible(&sched->wake_up_worker);
 559 }
 560
 561 /**
 562  * drm_sched_select_entity - Select next entity to process
 563  *
 564  * @sched: scheduler instance
 565  *
 566  * Returns the entity to process or NULL if none are found.
 567  */
 568 static struct drm_sched_entity *
 569 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 570 {
 571         struct drm_sched_entity *entity;
 572         int i;
 573
 574         if (!drm_sched_ready(sched))
 575                 return NULL;
 576
 577         /* Kernel run queue has higher priority than normal run queue*/
 578         for (i = DRM_SCHED_PRIORITY_MAX - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 579                 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
 580                 if (entity)
 581                         break;
 582         }
 583
 584         return entity;
 585 }
 586
 587 /**
 588  * drm_sched_process_job - process a job
 589  *
 590  * @f: fence
 591  * @cb: fence callbacks
 592  *
 593  * Called after job has finished execution.
 594  */
 595 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
 596 {
 597         struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
 598         struct drm_sched_fence *s_fence = s_job->s_fence;
 599         struct drm_gpu_scheduler *sched = s_fence->sched;
 600         unsigned long flags;
 601
 602         cancel_delayed_work(&sched->work_tdr);
 603
 604         atomic_dec(&sched->hw_rq_count);
 605         atomic_dec(&sched->num_jobs);
 606
 607         spin_lock_irqsave(&sched->job_list_lock, flags);
 608         /* remove job from ring_mirror_list */
 609         list_del_init(&s_job->node);
 610         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 611
 612         drm_sched_fence_finished(s_fence);
 613
 614         trace_drm_sched_process_job(s_fence);
 615         wake_up_interruptible(&sched->wake_up_worker);
 616
 617         schedule_work(&s_job->finish_work);
 618 }
 619
 620 /**
 621  * drm_sched_blocked - check if the scheduler is blocked
 622  *
 623  * @sched: scheduler instance
 624  *
 625  * Returns true if blocked, otherwise false.
 626  */
 627 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 628 {
 629         if (kthread_should_park()) {
 630                 kthread_parkme();
 631                 return true;
 632         }
 633
 634         return false;
 635 }
 636
 637 /**
 638  * drm_sched_main - main scheduler thread
 639  *
 640  * @param: scheduler instance
 641  *
 642  * Returns 0.
 643  */
 644 static int drm_sched_main(void *param)
 645 {
 646         struct sched_param sparam = {.sched_priority = 1};
 647         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
 648         int r;
 649
 650         sched_setscheduler(current, SCHED_FIFO, &sparam);
 651
 652         while (!kthread_should_stop()) {
 653                 struct drm_sched_entity *entity = NULL;
 654                 struct drm_sched_fence *s_fence;
 655                 struct drm_sched_job *sched_job;
 656                 struct dma_fence *fence;
 657
 658                 wait_event_interruptible(sched->wake_up_worker,
 659                                          (!drm_sched_blocked(sched) &&
 660                                           (entity = drm_sched_select_entity(sched))) ||
 661                                          kthread_should_stop());
 662
 663                 if (!entity)
 664                         continue;
 665
 666                 sched_job = drm_sched_entity_pop_job(entity);
 667                 if (!sched_job)
 668                         continue;
 669
 670                 s_fence = sched_job->s_fence;
 671
 672                 atomic_inc(&sched->hw_rq_count);
 673                 drm_sched_job_begin(sched_job);
 674
 675                 fence = sched->ops->run_job(sched_job);
 676                 drm_sched_fence_scheduled(s_fence);
 677
 678                 if (fence) {
 679                         s_fence->parent = dma_fence_get(fence);
 680                         r = dma_fence_add_callback(fence, &sched_job->cb,
 681                                                    drm_sched_process_job);
 682                         if (r == -ENOENT)
 683                                 drm_sched_process_job(fence, &sched_job->cb);
 684                         else if (r)
 685                                 DRM_ERROR("fence add callback failed (%d)\n",
 686                                           r);
 687                         dma_fence_put(fence);
 688                 } else
 689                         drm_sched_process_job(NULL, &sched_job->cb);
 690
 691                 wake_up(&sched->job_scheduled);
 692         }
 693         return 0;
 694 }
 695
 696 /**
 697  * drm_sched_init - Init a gpu scheduler instance
 698  *
 699  * @sched: scheduler instance
 700  * @ops: backend operations for this scheduler
 701  * @hw_submission: number of hw submissions that can be in flight
 702  * @hang_limit: number of times to allow a job to hang before dropping it
 703  * @timeout: timeout value in jiffies for the scheduler
 704  * @name: name used for debugging
 705  *
 706  * Return 0 on success, otherwise error code.
 707  */
 708 int drm_sched_init(struct drm_gpu_scheduler *sched,
 709                    const struct drm_sched_backend_ops *ops,
 710                    unsigned hw_submission,
 711                    unsigned hang_limit,
 712                    long timeout,
 713                    const char *name)
 714 {
 715         int i, ret;
 716         sched->ops = ops;
 717         sched->hw_submission_limit = hw_submission;
 718         sched->name = name;
 719         sched->timeout = timeout;
 720         sched->hang_limit = hang_limit;
 721         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++)
 722                 drm_sched_rq_init(sched, &sched->sched_rq[i]);
 723
 724         init_waitqueue_head(&sched->wake_up_worker);
 725         init_waitqueue_head(&sched->job_scheduled);
 726         INIT_LIST_HEAD(&sched->ring_mirror_list);
 727         spin_lock_init(&sched->job_list_lock);
 728         atomic_set(&sched->hw_rq_count, 0);
 729         INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
 730         atomic_set(&sched->num_jobs, 0);
 731         atomic64_set(&sched->job_id_count, 0);
 732
 733         /* Each scheduler will run on a seperate kernel thread */
 734         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
 735         if (IS_ERR(sched->thread)) {
 736                 ret = PTR_ERR(sched->thread);
 737                 sched->thread = NULL;
 738                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 739                 return ret;
 740         }
 741
 742         sched->ready = true;
 743         return 0;
 744 }
 745 EXPORT_SYMBOL(drm_sched_init);
 746
 747 /**
 748  * drm_sched_fini - Destroy a gpu scheduler
 749  *
 750  * @sched: scheduler instance
 751  *
 752  * Tears down and cleans up the scheduler.
 753  */
 754 void drm_sched_fini(struct drm_gpu_scheduler *sched)
 755 {
 756         if (sched->thread)
 757                 kthread_stop(sched->thread);
 758
 759         sched->ready = false;
 760 }
 761 EXPORT_SYMBOL(drm_sched_fini);