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