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 <linux/completion.h>
  51 #include <uapi/linux/sched/types.h>
  52
  53 #include <drm/drm_print.h>
  54 #include <drm/gpu_scheduler.h>
  55 #include <drm/spsc_queue.h>
  56
  57 #define CREATE_TRACE_POINTS
  58 #include "gpu_scheduler_trace.h"
  59
  60 #define to_drm_sched_job(sched_job)             \
  61                 container_of((sched_job), struct drm_sched_job, queue_node)
  62
  63 /**
  64  * drm_sched_rq_init - initialize a given run queue struct
  65  *
  66  * @sched: scheduler instance to associate with this run queue
  67  * @rq: scheduler run queue
  68  *
  69  * Initializes a scheduler runqueue.
  70  */
  71 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
  72                               struct drm_sched_rq *rq)
  73 {
  74         spin_lock_init(&rq->lock);
  75         INIT_LIST_HEAD(&rq->entities);
  76         rq->current_entity = NULL;
  77         rq->sched = sched;
  78 }
  79
  80 /**
  81  * drm_sched_rq_add_entity - add an entity
  82  *
  83  * @rq: scheduler run queue
  84  * @entity: scheduler entity
  85  *
  86  * Adds a scheduler entity to the run queue.
  87  */
  88 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
  89                              struct drm_sched_entity *entity)
  90 {
  91         if (!list_empty(&entity->list))
  92                 return;
  93         spin_lock(&rq->lock);
  94         atomic_inc(&rq->sched->score);
  95         list_add_tail(&entity->list, &rq->entities);
  96         spin_unlock(&rq->lock);
  97 }
  98
  99 /**
 100  * drm_sched_rq_remove_entity - remove an entity
 101  *
 102  * @rq: scheduler run queue
 103  * @entity: scheduler entity
 104  *
 105  * Removes a scheduler entity from the run queue.
 106  */
 107 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 108                                 struct drm_sched_entity *entity)
 109 {
 110         if (list_empty(&entity->list))
 111                 return;
 112         spin_lock(&rq->lock);
 113         atomic_dec(&rq->sched->score);
 114         list_del_init(&entity->list);
 115         if (rq->current_entity == entity)
 116                 rq->current_entity = NULL;
 117         spin_unlock(&rq->lock);
 118 }
 119
 120 /**
 121  * drm_sched_rq_select_entity - Select an entity which could provide a job to run
 122  *
 123  * @rq: scheduler run queue to check.
 124  *
 125  * Try to find a ready entity, returns NULL if none found.
 126  */
 127 static struct drm_sched_entity *
 128 drm_sched_rq_select_entity(struct drm_sched_rq *rq)
 129 {
 130         struct drm_sched_entity *entity;
 131
 132         spin_lock(&rq->lock);
 133
 134         entity = rq->current_entity;
 135         if (entity) {
 136                 list_for_each_entry_continue(entity, &rq->entities, list) {
 137                         if (drm_sched_entity_is_ready(entity)) {
 138                                 rq->current_entity = entity;
 139                                 reinit_completion(&entity->entity_idle);
 140                                 spin_unlock(&rq->lock);
 141                                 return entity;
 142                         }
 143                 }
 144         }
 145
 146         list_for_each_entry(entity, &rq->entities, list) {
 147
 148                 if (drm_sched_entity_is_ready(entity)) {
 149                         rq->current_entity = entity;
 150                         reinit_completion(&entity->entity_idle);
 151                         spin_unlock(&rq->lock);
 152                         return entity;
 153                 }
 154
 155                 if (entity == rq->current_entity)
 156                         break;
 157         }
 158
 159         spin_unlock(&rq->lock);
 160
 161         return NULL;
 162 }
 163
 164 /**
 165  * drm_sched_job_done - complete a job
 166  * @s_job: pointer to the job which is done
 167  *
 168  * Finish the job's fence and wake up the worker thread.
 169  */
 170 static void drm_sched_job_done(struct drm_sched_job *s_job)
 171 {
 172         struct drm_sched_fence *s_fence = s_job->s_fence;
 173         struct drm_gpu_scheduler *sched = s_fence->sched;
 174
 175         atomic_dec(&sched->hw_rq_count);
 176         atomic_dec(&sched->score);
 177
 178         trace_drm_sched_process_job(s_fence);
 179
 180         dma_fence_get(&s_fence->finished);
 181         drm_sched_fence_finished(s_fence);
 182         dma_fence_put(&s_fence->finished);
 183         wake_up_interruptible(&sched->wake_up_worker);
 184 }
 185
 186 /**
 187  * drm_sched_job_done_cb - the callback for a done job
 188  * @f: fence
 189  * @cb: fence callbacks
 190  */
 191 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
 192 {
 193         struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
 194
 195         drm_sched_job_done(s_job);
 196 }
 197
 198 /**
 199  * drm_sched_dependency_optimized
 200  *
 201  * @fence: the dependency fence
 202  * @entity: the entity which depends on the above fence
 203  *
 204  * Returns true if the dependency can be optimized and false otherwise
 205  */
 206 bool drm_sched_dependency_optimized(struct dma_fence* fence,
 207                                     struct drm_sched_entity *entity)
 208 {
 209         struct drm_gpu_scheduler *sched = entity->rq->sched;
 210         struct drm_sched_fence *s_fence;
 211
 212         if (!fence || dma_fence_is_signaled(fence))
 213                 return false;
 214         if (fence->context == entity->fence_context)
 215                 return true;
 216         s_fence = to_drm_sched_fence(fence);
 217         if (s_fence && s_fence->sched == sched)
 218                 return true;
 219
 220         return false;
 221 }
 222 EXPORT_SYMBOL(drm_sched_dependency_optimized);
 223
 224 /**
 225  * drm_sched_start_timeout - start timeout for reset worker
 226  *
 227  * @sched: scheduler instance to start the worker for
 228  *
 229  * Start the timeout for the given scheduler.
 230  */
 231 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
 232 {
 233         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 234             !list_empty(&sched->pending_list))
 235                 schedule_delayed_work(&sched->work_tdr, sched->timeout);
 236 }
 237
 238 /**
 239  * drm_sched_fault - immediately start timeout handler
 240  *
 241  * @sched: scheduler where the timeout handling should be started.
 242  *
 243  * Start timeout handling immediately when the driver detects a hardware fault.
 244  */
 245 void drm_sched_fault(struct drm_gpu_scheduler *sched)
 246 {
 247         mod_delayed_work(system_wq, &sched->work_tdr, 0);
 248 }
 249 EXPORT_SYMBOL(drm_sched_fault);
 250
 251 /**
 252  * drm_sched_suspend_timeout - Suspend scheduler job timeout
 253  *
 254  * @sched: scheduler instance for which to suspend the timeout
 255  *
 256  * Suspend the delayed work timeout for the scheduler. This is done by
 257  * modifying the delayed work timeout to an arbitrary large value,
 258  * MAX_SCHEDULE_TIMEOUT in this case.
 259  *
 260  * Returns the timeout remaining
 261  *
 262  */
 263 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
 264 {
 265         unsigned long sched_timeout, now = jiffies;
 266
 267         sched_timeout = sched->work_tdr.timer.expires;
 268
 269         /*
 270          * Modify the timeout to an arbitrarily large value. This also prevents
 271          * the timeout to be restarted when new submissions arrive
 272          */
 273         if (mod_delayed_work(system_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
 274                         && time_after(sched_timeout, now))
 275                 return sched_timeout - now;
 276         else
 277                 return sched->timeout;
 278 }
 279 EXPORT_SYMBOL(drm_sched_suspend_timeout);
 280
 281 /**
 282  * drm_sched_resume_timeout - Resume scheduler job timeout
 283  *
 284  * @sched: scheduler instance for which to resume the timeout
 285  * @remaining: remaining timeout
 286  *
 287  * Resume the delayed work timeout for the scheduler.
 288  */
 289 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
 290                 unsigned long remaining)
 291 {
 292         spin_lock(&sched->job_list_lock);
 293
 294         if (list_empty(&sched->pending_list))
 295                 cancel_delayed_work(&sched->work_tdr);
 296         else
 297                 mod_delayed_work(system_wq, &sched->work_tdr, remaining);
 298
 299         spin_unlock(&sched->job_list_lock);
 300 }
 301 EXPORT_SYMBOL(drm_sched_resume_timeout);
 302
 303 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 304 {
 305         struct drm_gpu_scheduler *sched = s_job->sched;
 306
 307         spin_lock(&sched->job_list_lock);
 308         list_add_tail(&s_job->list, &sched->pending_list);
 309         drm_sched_start_timeout(sched);
 310         spin_unlock(&sched->job_list_lock);
 311 }
 312
 313 static void drm_sched_job_timedout(struct work_struct *work)
 314 {
 315         struct drm_gpu_scheduler *sched;
 316         struct drm_sched_job *job;
 317
 318         sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 319
 320         /* Protects against concurrent deletion in drm_sched_get_cleanup_job */
 321         spin_lock(&sched->job_list_lock);
 322         job = list_first_entry_or_null(&sched->pending_list,
 323                                        struct drm_sched_job, list);
 324
 325         if (job) {
 326                 /*
 327                  * Remove the bad job so it cannot be freed by concurrent
 328                  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
 329                  * is parked at which point it's safe.
 330                  */
 331                 list_del_init(&job->list);
 332                 spin_unlock(&sched->job_list_lock);
 333
 334                 job->sched->ops->timedout_job(job);
 335
 336                 /*
 337                  * Guilty job did complete and hence needs to be manually removed
 338                  * See drm_sched_stop doc.
 339                  */
 340                 if (sched->free_guilty) {
 341                         job->sched->ops->free_job(job);
 342                         sched->free_guilty = false;
 343                 }
 344         } else {
 345                 spin_unlock(&sched->job_list_lock);
 346         }
 347
 348         spin_lock(&sched->job_list_lock);
 349         drm_sched_start_timeout(sched);
 350         spin_unlock(&sched->job_list_lock);
 351 }
 352
 353  /**
 354   * drm_sched_increase_karma - Update sched_entity guilty flag
 355   *
 356   * @bad: The job guilty of time out
 357   *
 358   * Increment on every hang caused by the 'bad' job. If this exceeds the hang
 359   * limit of the scheduler then the respective sched entity is marked guilty and
 360   * jobs from it will not be scheduled further
 361   */
 362 void drm_sched_increase_karma(struct drm_sched_job *bad)
 363 {
 364         int i;
 365         struct drm_sched_entity *tmp;
 366         struct drm_sched_entity *entity;
 367         struct drm_gpu_scheduler *sched = bad->sched;
 368
 369         /* don't increase @bad's karma if it's from KERNEL RQ,
 370          * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
 371          * corrupt but keep in mind that kernel jobs always considered good.
 372          */
 373         if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 374                 atomic_inc(&bad->karma);
 375                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
 376                      i++) {
 377                         struct drm_sched_rq *rq = &sched->sched_rq[i];
 378
 379                         spin_lock(&rq->lock);
 380                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
 381                                 if (bad->s_fence->scheduled.context ==
 382                                     entity->fence_context) {
 383                                         if (atomic_read(&bad->karma) >
 384                                             bad->sched->hang_limit)
 385                                                 if (entity->guilty)
 386                                                         atomic_set(entity->guilty, 1);
 387                                         break;
 388                                 }
 389                         }
 390                         spin_unlock(&rq->lock);
 391                         if (&entity->list != &rq->entities)
 392                                 break;
 393                 }
 394         }
 395 }
 396 EXPORT_SYMBOL(drm_sched_increase_karma);
 397
 398 /**
 399  * drm_sched_stop - stop the scheduler
 400  *
 401  * @sched: scheduler instance
 402  * @bad: job which caused the time out
 403  *
 404  * Stop the scheduler and also removes and frees all completed jobs.
 405  * Note: bad job will not be freed as it might be used later and so it's
 406  * callers responsibility to release it manually if it's not part of the
 407  * pending list any more.
 408  *
 409  */
 410 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 411 {
 412         struct drm_sched_job *s_job, *tmp;
 413
 414         kthread_park(sched->thread);
 415
 416         /*
 417          * Reinsert back the bad job here - now it's safe as
 418          * drm_sched_get_cleanup_job cannot race against us and release the
 419          * bad job at this point - we parked (waited for) any in progress
 420          * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
 421          * now until the scheduler thread is unparked.
 422          */
 423         if (bad && bad->sched == sched)
 424                 /*
 425                  * Add at the head of the queue to reflect it was the earliest
 426                  * job extracted.
 427                  */
 428                 list_add(&bad->list, &sched->pending_list);
 429
 430         /*
 431          * Iterate the job list from later to  earlier one and either deactive
 432          * their HW callbacks or remove them from pending list if they already
 433          * signaled.
 434          * This iteration is thread safe as sched thread is stopped.
 435          */
 436         list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
 437                                          list) {
 438                 if (s_job->s_fence->parent &&
 439                     dma_fence_remove_callback(s_job->s_fence->parent,
 440                                               &s_job->cb)) {
 441                         atomic_dec(&sched->hw_rq_count);
 442                 } else {
 443                         /*
 444                          * remove job from pending_list.
 445                          * Locking here is for concurrent resume timeout
 446                          */
 447                         spin_lock(&sched->job_list_lock);
 448                         list_del_init(&s_job->list);
 449                         spin_unlock(&sched->job_list_lock);
 450
 451                         /*
 452                          * Wait for job's HW fence callback to finish using s_job
 453                          * before releasing it.
 454                          *
 455                          * Job is still alive so fence refcount at least 1
 456                          */
 457                         dma_fence_wait(&s_job->s_fence->finished, false);
 458
 459                         /*
 460                          * We must keep bad job alive for later use during
 461                          * recovery by some of the drivers but leave a hint
 462                          * that the guilty job must be released.
 463                          */
 464                         if (bad != s_job)
 465                                 sched->ops->free_job(s_job);
 466                         else
 467                                 sched->free_guilty = true;
 468                 }
 469         }
 470
 471         /*
 472          * Stop pending timer in flight as we rearm it in  drm_sched_start. This
 473          * avoids the pending timeout work in progress to fire right away after
 474          * this TDR finished and before the newly restarted jobs had a
 475          * chance to complete.
 476          */
 477         cancel_delayed_work(&sched->work_tdr);
 478 }
 479
 480 EXPORT_SYMBOL(drm_sched_stop);
 481
 482 /**
 483  * drm_sched_start - recover jobs after a reset
 484  *
 485  * @sched: scheduler instance
 486  * @full_recovery: proceed with complete sched restart
 487  *
 488  */
 489 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
 490 {
 491         struct drm_sched_job *s_job, *tmp;
 492         int r;
 493
 494         /*
 495          * Locking the list is not required here as the sched thread is parked
 496          * so no new jobs are being inserted or removed. Also concurrent
 497          * GPU recovers can't run in parallel.
 498          */
 499         list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 500                 struct dma_fence *fence = s_job->s_fence->parent;
 501
 502                 atomic_inc(&sched->hw_rq_count);
 503
 504                 if (!full_recovery)
 505                         continue;
 506
 507                 if (fence) {
 508                         r = dma_fence_add_callback(fence, &s_job->cb,
 509                                                    drm_sched_job_done_cb);
 510                         if (r == -ENOENT)
 511                                 drm_sched_job_done(s_job);
 512                         else if (r)
 513                                 DRM_ERROR("fence add callback failed (%d)\n",
 514                                           r);
 515                 } else
 516                         drm_sched_job_done(s_job);
 517         }
 518
 519         if (full_recovery) {
 520                 spin_lock(&sched->job_list_lock);
 521                 drm_sched_start_timeout(sched);
 522                 spin_unlock(&sched->job_list_lock);
 523         }
 524
 525         kthread_unpark(sched->thread);
 526 }
 527 EXPORT_SYMBOL(drm_sched_start);
 528
 529 /**
 530  * drm_sched_resubmit_jobs - helper to relunch job from pending ring list
 531  *
 532  * @sched: scheduler instance
 533  *
 534  */
 535 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
 536 {
 537         struct drm_sched_job *s_job, *tmp;
 538         uint64_t guilty_context;
 539         bool found_guilty = false;
 540         struct dma_fence *fence;
 541
 542         list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 543                 struct drm_sched_fence *s_fence = s_job->s_fence;
 544
 545                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 546                         found_guilty = true;
 547                         guilty_context = s_job->s_fence->scheduled.context;
 548                 }
 549
 550                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 551                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 552
 553                 dma_fence_put(s_job->s_fence->parent);
 554                 fence = sched->ops->run_job(s_job);
 555
 556                 if (IS_ERR_OR_NULL(fence)) {
 557                         if (IS_ERR(fence))
 558                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
 559
 560                         s_job->s_fence->parent = NULL;
 561                 } else {
 562                         s_job->s_fence->parent = fence;
 563                 }
 564
 565
 566         }
 567 }
 568 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 569
 570 /**
 571  * drm_sched_job_init - init a scheduler job
 572  *
 573  * @job: scheduler job to init
 574  * @entity: scheduler entity to use
 575  * @owner: job owner for debugging
 576  *
 577  * Refer to drm_sched_entity_push_job() documentation
 578  * for locking considerations.
 579  *
 580  * Returns 0 for success, negative error code otherwise.
 581  */
 582 int drm_sched_job_init(struct drm_sched_job *job,
 583                        struct drm_sched_entity *entity,
 584                        void *owner)
 585 {
 586         struct drm_gpu_scheduler *sched;
 587
 588         drm_sched_entity_select_rq(entity);
 589         if (!entity->rq)
 590                 return -ENOENT;
 591
 592         sched = entity->rq->sched;
 593
 594         job->sched = sched;
 595         job->entity = entity;
 596         job->s_priority = entity->rq - sched->sched_rq;
 597         job->s_fence = drm_sched_fence_create(entity, owner);
 598         if (!job->s_fence)
 599                 return -ENOMEM;
 600         job->id = atomic64_inc_return(&sched->job_id_count);
 601
 602         INIT_LIST_HEAD(&job->list);
 603
 604         return 0;
 605 }
 606 EXPORT_SYMBOL(drm_sched_job_init);
 607
 608 /**
 609  * drm_sched_job_cleanup - clean up scheduler job resources
 610  *
 611  * @job: scheduler job to clean up
 612  */
 613 void drm_sched_job_cleanup(struct drm_sched_job *job)
 614 {
 615         dma_fence_put(&job->s_fence->finished);
 616         job->s_fence = NULL;
 617 }
 618 EXPORT_SYMBOL(drm_sched_job_cleanup);
 619
 620 /**
 621  * drm_sched_ready - is the scheduler ready
 622  *
 623  * @sched: scheduler instance
 624  *
 625  * Return true if we can push more jobs to the hw, otherwise false.
 626  */
 627 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 628 {
 629         return atomic_read(&sched->hw_rq_count) <
 630                 sched->hw_submission_limit;
 631 }
 632
 633 /**
 634  * drm_sched_wakeup - Wake up the scheduler when it is ready
 635  *
 636  * @sched: scheduler instance
 637  *
 638  */
 639 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 640 {
 641         if (drm_sched_ready(sched))
 642                 wake_up_interruptible(&sched->wake_up_worker);
 643 }
 644
 645 /**
 646  * drm_sched_select_entity - Select next entity to process
 647  *
 648  * @sched: scheduler instance
 649  *
 650  * Returns the entity to process or NULL if none are found.
 651  */
 652 static struct drm_sched_entity *
 653 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 654 {
 655         struct drm_sched_entity *entity;
 656         int i;
 657
 658         if (!drm_sched_ready(sched))
 659                 return NULL;
 660
 661         /* Kernel run queue has higher priority than normal run queue*/
 662         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 663                 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
 664                 if (entity)
 665                         break;
 666         }
 667
 668         return entity;
 669 }
 670
 671 /**
 672  * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
 673  *
 674  * @sched: scheduler instance
 675  *
 676  * Returns the next finished job from the pending list (if there is one)
 677  * ready for it to be destroyed.
 678  */
 679 static struct drm_sched_job *
 680 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
 681 {
 682         struct drm_sched_job *job;
 683
 684         /*
 685          * Don't destroy jobs while the timeout worker is running  OR thread
 686          * is being parked and hence assumed to not touch pending_list
 687          */
 688         if ((sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 689             !cancel_delayed_work(&sched->work_tdr)) ||
 690             kthread_should_park())
 691                 return NULL;
 692
 693         spin_lock(&sched->job_list_lock);
 694
 695         job = list_first_entry_or_null(&sched->pending_list,
 696                                        struct drm_sched_job, list);
 697
 698         if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
 699                 /* remove job from pending_list */
 700                 list_del_init(&job->list);
 701         } else {
 702                 job = NULL;
 703                 /* queue timeout for next job */
 704                 drm_sched_start_timeout(sched);
 705         }
 706
 707         spin_unlock(&sched->job_list_lock);
 708
 709         return job;
 710 }
 711
 712 /**
 713  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
 714  * @sched_list: list of drm_gpu_schedulers
 715  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
 716  *
 717  * Returns pointer of the sched with the least load or NULL if none of the
 718  * drm_gpu_schedulers are ready
 719  */
 720 struct drm_gpu_scheduler *
 721 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
 722                      unsigned int num_sched_list)
 723 {
 724         struct drm_gpu_scheduler *sched, *picked_sched = NULL;
 725         int i;
 726         unsigned int min_score = UINT_MAX, num_score;
 727
 728         for (i = 0; i < num_sched_list; ++i) {
 729                 sched = sched_list[i];
 730
 731                 if (!sched->ready) {
 732                         DRM_WARN("scheduler %s is not ready, skipping",
 733                                  sched->name);
 734                         continue;
 735                 }
 736
 737                 num_score = atomic_read(&sched->score);
 738                 if (num_score < min_score) {
 739                         min_score = num_score;
 740                         picked_sched = sched;
 741                 }
 742         }
 743
 744         return picked_sched;
 745 }
 746 EXPORT_SYMBOL(drm_sched_pick_best);
 747
 748 /**
 749  * drm_sched_blocked - check if the scheduler is blocked
 750  *
 751  * @sched: scheduler instance
 752  *
 753  * Returns true if blocked, otherwise false.
 754  */
 755 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 756 {
 757         if (kthread_should_park()) {
 758                 kthread_parkme();
 759                 return true;
 760         }
 761
 762         return false;
 763 }
 764
 765 /**
 766  * drm_sched_main - main scheduler thread
 767  *
 768  * @param: scheduler instance
 769  *
 770  * Returns 0.
 771  */
 772 static int drm_sched_main(void *param)
 773 {
 774         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
 775         int r;
 776
 777         sched_set_fifo_low(current);
 778
 779         while (!kthread_should_stop()) {
 780                 struct drm_sched_entity *entity = NULL;
 781                 struct drm_sched_fence *s_fence;
 782                 struct drm_sched_job *sched_job;
 783                 struct dma_fence *fence;
 784                 struct drm_sched_job *cleanup_job = NULL;
 785
 786                 wait_event_interruptible(sched->wake_up_worker,
 787                                          (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
 788                                          (!drm_sched_blocked(sched) &&
 789                                           (entity = drm_sched_select_entity(sched))) ||
 790                                          kthread_should_stop());
 791
 792                 if (cleanup_job) {
 793                         sched->ops->free_job(cleanup_job);
 794                         /* queue timeout for next job */
 795                         drm_sched_start_timeout(sched);
 796                 }
 797
 798                 if (!entity)
 799                         continue;
 800
 801                 sched_job = drm_sched_entity_pop_job(entity);
 802
 803                 complete(&entity->entity_idle);
 804
 805                 if (!sched_job)
 806                         continue;
 807
 808                 s_fence = sched_job->s_fence;
 809
 810                 atomic_inc(&sched->hw_rq_count);
 811                 drm_sched_job_begin(sched_job);
 812
 813                 trace_drm_run_job(sched_job, entity);
 814                 fence = sched->ops->run_job(sched_job);
 815                 drm_sched_fence_scheduled(s_fence);
 816
 817                 if (!IS_ERR_OR_NULL(fence)) {
 818                         s_fence->parent = dma_fence_get(fence);
 819                         r = dma_fence_add_callback(fence, &sched_job->cb,
 820                                                    drm_sched_job_done_cb);
 821                         if (r == -ENOENT)
 822                                 drm_sched_job_done(sched_job);
 823                         else if (r)
 824                                 DRM_ERROR("fence add callback failed (%d)\n",
 825                                           r);
 826                         dma_fence_put(fence);
 827                 } else {
 828                         if (IS_ERR(fence))
 829                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
 830
 831                         drm_sched_job_done(sched_job);
 832                 }
 833
 834                 wake_up(&sched->job_scheduled);
 835         }
 836         return 0;
 837 }
 838
 839 /**
 840  * drm_sched_init - Init a gpu scheduler instance
 841  *
 842  * @sched: scheduler instance
 843  * @ops: backend operations for this scheduler
 844  * @hw_submission: number of hw submissions that can be in flight
 845  * @hang_limit: number of times to allow a job to hang before dropping it
 846  * @timeout: timeout value in jiffies for the scheduler
 847  * @name: name used for debugging
 848  *
 849  * Return 0 on success, otherwise error code.
 850  */
 851 int drm_sched_init(struct drm_gpu_scheduler *sched,
 852                    const struct drm_sched_backend_ops *ops,
 853                    unsigned hw_submission,
 854                    unsigned hang_limit,
 855                    long timeout,
 856                    const char *name)
 857 {
 858         int i, ret;
 859         sched->ops = ops;
 860         sched->hw_submission_limit = hw_submission;
 861         sched->name = name;
 862         sched->timeout = timeout;
 863         sched->hang_limit = hang_limit;
 864         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
 865                 drm_sched_rq_init(sched, &sched->sched_rq[i]);
 866
 867         init_waitqueue_head(&sched->wake_up_worker);
 868         init_waitqueue_head(&sched->job_scheduled);
 869         INIT_LIST_HEAD(&sched->pending_list);
 870         spin_lock_init(&sched->job_list_lock);
 871         atomic_set(&sched->hw_rq_count, 0);
 872         INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
 873         atomic_set(&sched->score, 0);
 874         atomic64_set(&sched->job_id_count, 0);
 875
 876         /* Each scheduler will run on a seperate kernel thread */
 877         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
 878         if (IS_ERR(sched->thread)) {
 879                 ret = PTR_ERR(sched->thread);
 880                 sched->thread = NULL;
 881                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 882                 return ret;
 883         }
 884
 885         sched->ready = true;
 886         return 0;
 887 }
 888 EXPORT_SYMBOL(drm_sched_init);
 889
 890 /**
 891  * drm_sched_fini - Destroy a gpu scheduler
 892  *
 893  * @sched: scheduler instance
 894  *
 895  * Tears down and cleans up the scheduler.
 896  */
 897 void drm_sched_fini(struct drm_gpu_scheduler *sched)
 898 {
 899         if (sched->thread)
 900                 kthread_stop(sched->thread);
 901
 902         /* Confirm no work left behind accessing device structures */
 903         cancel_delayed_work_sync(&sched->work_tdr);
 904
 905         sched->ready = false;
 906 }
 907 EXPORT_SYMBOL(drm_sched_fini);