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