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 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 269 {
 270         struct drm_gpu_scheduler *sched = s_job->sched;
 271         unsigned long flags;
 272
 273         spin_lock_irqsave(&sched->job_list_lock, flags);
 274         list_add_tail(&s_job->node, &sched->ring_mirror_list);
 275         drm_sched_start_timeout(sched);
 276         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 277 }
 278
 279 static void drm_sched_job_timedout(struct work_struct *work)
 280 {
 281         struct drm_gpu_scheduler *sched;
 282         struct drm_sched_job *job;
 283         unsigned long flags;
 284
 285         sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 286         job = list_first_entry_or_null(&sched->ring_mirror_list,
 287                                        struct drm_sched_job, node);
 288
 289         if (job)
 290                 job->sched->ops->timedout_job(job);
 291
 292         /*
 293          * Guilty job did complete and hence needs to be manually removed
 294          * See drm_sched_stop doc.
 295          */
 296         if (list_empty(&job->node))
 297                 job->sched->ops->free_job(job);
 298
 299         spin_lock_irqsave(&sched->job_list_lock, flags);
 300         drm_sched_start_timeout(sched);
 301         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 302 }
 303
 304  /**
 305   * drm_sched_increase_karma - Update sched_entity guilty flag
 306   *
 307   * @bad: The job guilty of time out
 308   *
 309   * Increment on every hang caused by the 'bad' job. If this exceeds the hang
 310   * limit of the scheduler then the respective sched entity is marked guilty and
 311   * jobs from it will not be scheduled further
 312   */
 313 void drm_sched_increase_karma(struct drm_sched_job *bad)
 314 {
 315         int i;
 316         struct drm_sched_entity *tmp;
 317         struct drm_sched_entity *entity;
 318         struct drm_gpu_scheduler *sched = bad->sched;
 319
 320         /* don't increase @bad's karma if it's from KERNEL RQ,
 321          * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
 322          * corrupt but keep in mind that kernel jobs always considered good.
 323          */
 324         if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 325                 atomic_inc(&bad->karma);
 326                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
 327                      i++) {
 328                         struct drm_sched_rq *rq = &sched->sched_rq[i];
 329
 330                         spin_lock(&rq->lock);
 331                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
 332                                 if (bad->s_fence->scheduled.context ==
 333                                     entity->fence_context) {
 334                                         if (atomic_read(&bad->karma) >
 335                                             bad->sched->hang_limit)
 336                                                 if (entity->guilty)
 337                                                         atomic_set(entity->guilty, 1);
 338                                         break;
 339                                 }
 340                         }
 341                         spin_unlock(&rq->lock);
 342                         if (&entity->list != &rq->entities)
 343                                 break;
 344                 }
 345         }
 346 }
 347 EXPORT_SYMBOL(drm_sched_increase_karma);
 348
 349 /**
 350  * drm_sched_hw_job_reset - stop the scheduler if it contains the bad job
 351  *
 352  * @sched: scheduler instance
 353  * @bad: bad scheduler job
 354  *
 355  * Stop the scheduler and also removes and frees all completed jobs.
 356  * Note: bad job will not be freed as it might be used later and so it's
 357  * callers responsibility to release it manually if it's not part of the
 358  * mirror list any more.
 359  *
 360  */
 361 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 362 {
 363         struct drm_sched_job *s_job, *tmp;
 364         unsigned long flags;
 365
 366         kthread_park(sched->thread);
 367
 368         /*
 369          * Iterate the job list from later to  earlier one and either deactive
 370          * their HW callbacks or remove them from mirror list if they already
 371          * signaled.
 372          * This iteration is thread safe as sched thread is stopped.
 373          */
 374         list_for_each_entry_safe_reverse(s_job, tmp, &sched->ring_mirror_list, node) {
 375                 if (s_job->s_fence->parent &&
 376                     dma_fence_remove_callback(s_job->s_fence->parent,
 377                                               &s_job->cb)) {
 378                         atomic_dec(&sched->hw_rq_count);
 379                 } else {
 380                         /*
 381                          * remove job from ring_mirror_list.
 382                          * Locking here is for concurrent resume timeout
 383                          */
 384                         spin_lock_irqsave(&sched->job_list_lock, flags);
 385                         list_del_init(&s_job->node);
 386                         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 387
 388                         /*
 389                          * Wait for job's HW fence callback to finish using s_job
 390                          * before releasing it.
 391                          *
 392                          * Job is still alive so fence refcount at least 1
 393                          */
 394                         dma_fence_wait(&s_job->s_fence->finished, false);
 395
 396                         /*
 397                          * We must keep bad job alive for later use during
 398                          * recovery by some of the drivers
 399                          */
 400                         if (bad != s_job)
 401                                 sched->ops->free_job(s_job);
 402                 }
 403         }
 404
 405         /*
 406          * Stop pending timer in flight as we rearm it in  drm_sched_start. This
 407          * avoids the pending timeout work in progress to fire right away after
 408          * this TDR finished and before the newly restarted jobs had a
 409          * chance to complete.
 410          */
 411         cancel_delayed_work(&sched->work_tdr);
 412 }
 413
 414 EXPORT_SYMBOL(drm_sched_stop);
 415
 416 /**
 417  * drm_sched_job_recovery - recover jobs after a reset
 418  *
 419  * @sched: scheduler instance
 420  *
 421  */
 422 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
 423 {
 424         struct drm_sched_job *s_job, *tmp;
 425         unsigned long flags;
 426         int r;
 427
 428         /*
 429          * Locking the list is not required here as the sched thread is parked
 430          * so no new jobs are being inserted or removed. 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                 atomic_inc(&sched->hw_rq_count);
 437
 438                 if (!full_recovery)
 439                         continue;
 440
 441                 if (fence) {
 442                         r = dma_fence_add_callback(fence, &s_job->cb,
 443                                                    drm_sched_process_job);
 444                         if (r == -ENOENT)
 445                                 drm_sched_process_job(fence, &s_job->cb);
 446                         else if (r)
 447                                 DRM_ERROR("fence add callback failed (%d)\n",
 448                                           r);
 449                 } else
 450                         drm_sched_process_job(NULL, &s_job->cb);
 451         }
 452
 453         if (full_recovery) {
 454                 spin_lock_irqsave(&sched->job_list_lock, flags);
 455                 drm_sched_start_timeout(sched);
 456                 spin_unlock_irqrestore(&sched->job_list_lock, flags);
 457         }
 458
 459         kthread_unpark(sched->thread);
 460 }
 461 EXPORT_SYMBOL(drm_sched_start);
 462
 463 /**
 464  * drm_sched_resubmit_jobs - helper to relunch job from mirror ring list
 465  *
 466  * @sched: scheduler instance
 467  *
 468  */
 469 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
 470 {
 471         struct drm_sched_job *s_job, *tmp;
 472         uint64_t guilty_context;
 473         bool found_guilty = false;
 474
 475         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 476                 struct drm_sched_fence *s_fence = s_job->s_fence;
 477
 478                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 479                         found_guilty = true;
 480                         guilty_context = s_job->s_fence->scheduled.context;
 481                 }
 482
 483                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 484                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 485
 486                 dma_fence_put(s_job->s_fence->parent);
 487                 s_job->s_fence->parent = sched->ops->run_job(s_job);
 488         }
 489 }
 490 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 491
 492 /**
 493  * drm_sched_job_init - init a scheduler job
 494  *
 495  * @job: scheduler job to init
 496  * @entity: scheduler entity to use
 497  * @owner: job owner for debugging
 498  *
 499  * Refer to drm_sched_entity_push_job() documentation
 500  * for locking considerations.
 501  *
 502  * Returns 0 for success, negative error code otherwise.
 503  */
 504 int drm_sched_job_init(struct drm_sched_job *job,
 505                        struct drm_sched_entity *entity,
 506                        void *owner)
 507 {
 508         struct drm_gpu_scheduler *sched;
 509
 510         drm_sched_entity_select_rq(entity);
 511         if (!entity->rq)
 512                 return -ENOENT;
 513
 514         sched = entity->rq->sched;
 515
 516         job->sched = sched;
 517         job->entity = entity;
 518         job->s_priority = entity->rq - sched->sched_rq;
 519         job->s_fence = drm_sched_fence_create(entity, owner);
 520         if (!job->s_fence)
 521                 return -ENOMEM;
 522         job->id = atomic64_inc_return(&sched->job_id_count);
 523
 524         INIT_LIST_HEAD(&job->node);
 525
 526         return 0;
 527 }
 528 EXPORT_SYMBOL(drm_sched_job_init);
 529
 530 /**
 531  * drm_sched_job_cleanup - clean up scheduler job resources
 532  *
 533  * @job: scheduler job to clean up
 534  */
 535 void drm_sched_job_cleanup(struct drm_sched_job *job)
 536 {
 537         dma_fence_put(&job->s_fence->finished);
 538         job->s_fence = NULL;
 539 }
 540 EXPORT_SYMBOL(drm_sched_job_cleanup);
 541
 542 /**
 543  * drm_sched_ready - is the scheduler ready
 544  *
 545  * @sched: scheduler instance
 546  *
 547  * Return true if we can push more jobs to the hw, otherwise false.
 548  */
 549 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 550 {
 551         return atomic_read(&sched->hw_rq_count) <
 552                 sched->hw_submission_limit;
 553 }
 554
 555 /**
 556  * drm_sched_wakeup - Wake up the scheduler when it is ready
 557  *
 558  * @sched: scheduler instance
 559  *
 560  */
 561 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 562 {
 563         if (drm_sched_ready(sched))
 564                 wake_up_interruptible(&sched->wake_up_worker);
 565 }
 566
 567 /**
 568  * drm_sched_select_entity - Select next entity to process
 569  *
 570  * @sched: scheduler instance
 571  *
 572  * Returns the entity to process or NULL if none are found.
 573  */
 574 static struct drm_sched_entity *
 575 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 576 {
 577         struct drm_sched_entity *entity;
 578         int i;
 579
 580         if (!drm_sched_ready(sched))
 581                 return NULL;
 582
 583         /* Kernel run queue has higher priority than normal run queue*/
 584         for (i = DRM_SCHED_PRIORITY_MAX - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 585                 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
 586                 if (entity)
 587                         break;
 588         }
 589
 590         return entity;
 591 }
 592
 593 /**
 594  * drm_sched_process_job - process a job
 595  *
 596  * @f: fence
 597  * @cb: fence callbacks
 598  *
 599  * Called after job has finished execution.
 600  */
 601 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
 602 {
 603         struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
 604         struct drm_sched_fence *s_fence = s_job->s_fence;
 605         struct drm_gpu_scheduler *sched = s_fence->sched;
 606
 607         atomic_dec(&sched->hw_rq_count);
 608         atomic_dec(&sched->num_jobs);
 609
 610         trace_drm_sched_process_job(s_fence);
 611
 612         drm_sched_fence_finished(s_fence);
 613         wake_up_interruptible(&sched->wake_up_worker);
 614 }
 615
 616 /**
 617  * drm_sched_cleanup_jobs - destroy finished jobs
 618  *
 619  * @sched: scheduler instance
 620  *
 621  * Remove all finished jobs from the mirror list and destroy them.
 622  */
 623 static void drm_sched_cleanup_jobs(struct drm_gpu_scheduler *sched)
 624 {
 625         unsigned long flags;
 626
 627         /* Don't destroy jobs while the timeout worker is running */
 628         if (!cancel_delayed_work(&sched->work_tdr))
 629                 return;
 630
 631
 632         while (!list_empty(&sched->ring_mirror_list)) {
 633                 struct drm_sched_job *job;
 634
 635                 job = list_first_entry(&sched->ring_mirror_list,
 636                                        struct drm_sched_job, node);
 637                 if (!dma_fence_is_signaled(&job->s_fence->finished))
 638                         break;
 639
 640                 spin_lock_irqsave(&sched->job_list_lock, flags);
 641                 /* remove job from ring_mirror_list */
 642                 list_del_init(&job->node);
 643                 spin_unlock_irqrestore(&sched->job_list_lock, flags);
 644
 645                 sched->ops->free_job(job);
 646         }
 647
 648         /* queue timeout for next job */
 649         spin_lock_irqsave(&sched->job_list_lock, flags);
 650         drm_sched_start_timeout(sched);
 651         spin_unlock_irqrestore(&sched->job_list_lock, flags);
 652
 653 }
 654
 655 /**
 656  * drm_sched_blocked - check if the scheduler is blocked
 657  *
 658  * @sched: scheduler instance
 659  *
 660  * Returns true if blocked, otherwise false.
 661  */
 662 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 663 {
 664         if (kthread_should_park()) {
 665                 kthread_parkme();
 666                 return true;
 667         }
 668
 669         return false;
 670 }
 671
 672 /**
 673  * drm_sched_main - main scheduler thread
 674  *
 675  * @param: scheduler instance
 676  *
 677  * Returns 0.
 678  */
 679 static int drm_sched_main(void *param)
 680 {
 681         struct sched_param sparam = {.sched_priority = 1};
 682         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
 683         int r;
 684
 685         sched_setscheduler(current, SCHED_FIFO, &sparam);
 686
 687         while (!kthread_should_stop()) {
 688                 struct drm_sched_entity *entity = NULL;
 689                 struct drm_sched_fence *s_fence;
 690                 struct drm_sched_job *sched_job;
 691                 struct dma_fence *fence;
 692
 693                 wait_event_interruptible(sched->wake_up_worker,
 694                                          (drm_sched_cleanup_jobs(sched),
 695                                          (!drm_sched_blocked(sched) &&
 696                                           (entity = drm_sched_select_entity(sched))) ||
 697                                          kthread_should_stop()));
 698
 699                 if (!entity)
 700                         continue;
 701
 702                 sched_job = drm_sched_entity_pop_job(entity);
 703                 if (!sched_job)
 704                         continue;
 705
 706                 s_fence = sched_job->s_fence;
 707
 708                 atomic_inc(&sched->hw_rq_count);
 709                 drm_sched_job_begin(sched_job);
 710
 711                 fence = sched->ops->run_job(sched_job);
 712                 drm_sched_fence_scheduled(s_fence);
 713
 714                 if (fence) {
 715                         s_fence->parent = dma_fence_get(fence);
 716                         r = dma_fence_add_callback(fence, &sched_job->cb,
 717                                                    drm_sched_process_job);
 718                         if (r == -ENOENT)
 719                                 drm_sched_process_job(fence, &sched_job->cb);
 720                         else if (r)
 721                                 DRM_ERROR("fence add callback failed (%d)\n",
 722                                           r);
 723                         dma_fence_put(fence);
 724                 } else
 725                         drm_sched_process_job(NULL, &sched_job->cb);
 726
 727                 wake_up(&sched->job_scheduled);
 728         }
 729         return 0;
 730 }
 731
 732 /**
 733  * drm_sched_init - Init a gpu scheduler instance
 734  *
 735  * @sched: scheduler instance
 736  * @ops: backend operations for this scheduler
 737  * @hw_submission: number of hw submissions that can be in flight
 738  * @hang_limit: number of times to allow a job to hang before dropping it
 739  * @timeout: timeout value in jiffies for the scheduler
 740  * @name: name used for debugging
 741  *
 742  * Return 0 on success, otherwise error code.
 743  */
 744 int drm_sched_init(struct drm_gpu_scheduler *sched,
 745                    const struct drm_sched_backend_ops *ops,
 746                    unsigned hw_submission,
 747                    unsigned hang_limit,
 748                    long timeout,
 749                    const char *name)
 750 {
 751         int i, ret;
 752         sched->ops = ops;
 753         sched->hw_submission_limit = hw_submission;
 754         sched->name = name;
 755         sched->timeout = timeout;
 756         sched->hang_limit = hang_limit;
 757         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++)
 758                 drm_sched_rq_init(sched, &sched->sched_rq[i]);
 759
 760         init_waitqueue_head(&sched->wake_up_worker);
 761         init_waitqueue_head(&sched->job_scheduled);
 762         INIT_LIST_HEAD(&sched->ring_mirror_list);
 763         spin_lock_init(&sched->job_list_lock);
 764         atomic_set(&sched->hw_rq_count, 0);
 765         INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
 766         atomic_set(&sched->num_jobs, 0);
 767         atomic64_set(&sched->job_id_count, 0);
 768
 769         /* Each scheduler will run on a seperate kernel thread */
 770         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
 771         if (IS_ERR(sched->thread)) {
 772                 ret = PTR_ERR(sched->thread);
 773                 sched->thread = NULL;
 774                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 775                 return ret;
 776         }
 777
 778         sched->ready = true;
 779         return 0;
 780 }
 781 EXPORT_SYMBOL(drm_sched_init);
 782
 783 /**
 784  * drm_sched_fini - Destroy a gpu scheduler
 785  *
 786  * @sched: scheduler instance
 787  *
 788  * Tears down and cleans up the scheduler.
 789  */
 790 void drm_sched_fini(struct drm_gpu_scheduler *sched)
 791 {
 792         if (sched->thread)
 793                 kthread_stop(sched->thread);
 794
 795         sched->ready = false;
 796 }
 797 EXPORT_SYMBOL(drm_sched_fini);