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         drm_sched_increase_karma_ext(bad, 1);
 365 }
 366 EXPORT_SYMBOL(drm_sched_increase_karma);
 367
 368 void drm_sched_reset_karma(struct drm_sched_job *bad)
 369 {
 370         drm_sched_increase_karma_ext(bad, 0);
 371 }
 372 EXPORT_SYMBOL(drm_sched_reset_karma);
 373
 374 /**
 375  * drm_sched_stop - stop the scheduler
 376  *
 377  * @sched: scheduler instance
 378  * @bad: job which caused the time out
 379  *
 380  * Stop the scheduler and also removes and frees all completed jobs.
 381  * Note: bad job will not be freed as it might be used later and so it's
 382  * callers responsibility to release it manually if it's not part of the
 383  * pending list any more.
 384  *
 385  */
 386 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 387 {
 388         struct drm_sched_job *s_job, *tmp;
 389
 390         kthread_park(sched->thread);
 391
 392         /*
 393          * Reinsert back the bad job here - now it's safe as
 394          * drm_sched_get_cleanup_job cannot race against us and release the
 395          * bad job at this point - we parked (waited for) any in progress
 396          * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
 397          * now until the scheduler thread is unparked.
 398          */
 399         if (bad && bad->sched == sched)
 400                 /*
 401                  * Add at the head of the queue to reflect it was the earliest
 402                  * job extracted.
 403                  */
 404                 list_add(&bad->list, &sched->pending_list);
 405
 406         /*
 407          * Iterate the job list from later to  earlier one and either deactive
 408          * their HW callbacks or remove them from pending list if they already
 409          * signaled.
 410          * This iteration is thread safe as sched thread is stopped.
 411          */
 412         list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
 413                                          list) {
 414                 if (s_job->s_fence->parent &&
 415                     dma_fence_remove_callback(s_job->s_fence->parent,
 416                                               &s_job->cb)) {
 417                         atomic_dec(&sched->hw_rq_count);
 418                 } else {
 419                         /*
 420                          * remove job from pending_list.
 421                          * Locking here is for concurrent resume timeout
 422                          */
 423                         spin_lock(&sched->job_list_lock);
 424                         list_del_init(&s_job->list);
 425                         spin_unlock(&sched->job_list_lock);
 426
 427                         /*
 428                          * Wait for job's HW fence callback to finish using s_job
 429                          * before releasing it.
 430                          *
 431                          * Job is still alive so fence refcount at least 1
 432                          */
 433                         dma_fence_wait(&s_job->s_fence->finished, false);
 434
 435                         /*
 436                          * We must keep bad job alive for later use during
 437                          * recovery by some of the drivers but leave a hint
 438                          * that the guilty job must be released.
 439                          */
 440                         if (bad != s_job)
 441                                 sched->ops->free_job(s_job);
 442                         else
 443                                 sched->free_guilty = true;
 444                 }
 445         }
 446
 447         /*
 448          * Stop pending timer in flight as we rearm it in  drm_sched_start. This
 449          * avoids the pending timeout work in progress to fire right away after
 450          * this TDR finished and before the newly restarted jobs had a
 451          * chance to complete.
 452          */
 453         cancel_delayed_work(&sched->work_tdr);
 454 }
 455
 456 EXPORT_SYMBOL(drm_sched_stop);
 457
 458 /**
 459  * drm_sched_start - recover jobs after a reset
 460  *
 461  * @sched: scheduler instance
 462  * @full_recovery: proceed with complete sched restart
 463  *
 464  */
 465 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
 466 {
 467         struct drm_sched_job *s_job, *tmp;
 468         int r;
 469
 470         /*
 471          * Locking the list is not required here as the sched thread is parked
 472          * so no new jobs are being inserted or removed. Also concurrent
 473          * GPU recovers can't run in parallel.
 474          */
 475         list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 476                 struct dma_fence *fence = s_job->s_fence->parent;
 477
 478                 atomic_inc(&sched->hw_rq_count);
 479
 480                 if (!full_recovery)
 481                         continue;
 482
 483                 if (fence) {
 484                         r = dma_fence_add_callback(fence, &s_job->cb,
 485                                                    drm_sched_job_done_cb);
 486                         if (r == -ENOENT)
 487                                 drm_sched_job_done(s_job);
 488                         else if (r)
 489                                 DRM_ERROR("fence add callback failed (%d)\n",
 490                                           r);
 491                 } else
 492                         drm_sched_job_done(s_job);
 493         }
 494
 495         if (full_recovery) {
 496                 spin_lock(&sched->job_list_lock);
 497                 drm_sched_start_timeout(sched);
 498                 spin_unlock(&sched->job_list_lock);
 499         }
 500
 501         kthread_unpark(sched->thread);
 502 }
 503 EXPORT_SYMBOL(drm_sched_start);
 504
 505 /**
 506  * drm_sched_resubmit_jobs - helper to relaunch jobs from the pending list
 507  *
 508  * @sched: scheduler instance
 509  *
 510  */
 511 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
 512 {
 513         drm_sched_resubmit_jobs_ext(sched, INT_MAX);
 514 }
 515 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 516
 517 /**
 518  * drm_sched_resubmit_jobs_ext - helper to relunch certain number of jobs from mirror ring list
 519  *
 520  * @sched: scheduler instance
 521  * @max: job numbers to relaunch
 522  *
 523  */
 524 void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)
 525 {
 526         struct drm_sched_job *s_job, *tmp;
 527         uint64_t guilty_context;
 528         bool found_guilty = false;
 529         struct dma_fence *fence;
 530         int i = 0;
 531
 532         list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 533                 struct drm_sched_fence *s_fence = s_job->s_fence;
 534
 535                 if (i >= max)
 536                         break;
 537
 538                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 539                         found_guilty = true;
 540                         guilty_context = s_job->s_fence->scheduled.context;
 541                 }
 542
 543                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 544                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 545
 546                 dma_fence_put(s_job->s_fence->parent);
 547                 fence = sched->ops->run_job(s_job);
 548                 i++;
 549
 550                 if (IS_ERR_OR_NULL(fence)) {
 551                         if (IS_ERR(fence))
 552                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
 553
 554                         s_job->s_fence->parent = NULL;
 555                 } else {
 556                         s_job->s_fence->parent = fence;
 557                 }
 558         }
 559 }
 560 EXPORT_SYMBOL(drm_sched_resubmit_jobs_ext);
 561
 562 /**
 563  * drm_sched_job_init - init a scheduler job
 564  *
 565  * @job: scheduler job to init
 566  * @entity: scheduler entity to use
 567  * @owner: job owner for debugging
 568  *
 569  * Refer to drm_sched_entity_push_job() documentation
 570  * for locking considerations.
 571  *
 572  * Returns 0 for success, negative error code otherwise.
 573  */
 574 int drm_sched_job_init(struct drm_sched_job *job,
 575                        struct drm_sched_entity *entity,
 576                        void *owner)
 577 {
 578         struct drm_gpu_scheduler *sched;
 579
 580         drm_sched_entity_select_rq(entity);
 581         if (!entity->rq)
 582                 return -ENOENT;
 583
 584         sched = entity->rq->sched;
 585
 586         job->sched = sched;
 587         job->entity = entity;
 588         job->s_priority = entity->rq - sched->sched_rq;
 589         job->s_fence = drm_sched_fence_create(entity, owner);
 590         if (!job->s_fence)
 591                 return -ENOMEM;
 592         job->id = atomic64_inc_return(&sched->job_id_count);
 593
 594         INIT_LIST_HEAD(&job->list);
 595
 596         return 0;
 597 }
 598 EXPORT_SYMBOL(drm_sched_job_init);
 599
 600 /**
 601  * drm_sched_job_cleanup - clean up scheduler job resources
 602  *
 603  * @job: scheduler job to clean up
 604  */
 605 void drm_sched_job_cleanup(struct drm_sched_job *job)
 606 {
 607         dma_fence_put(&job->s_fence->finished);
 608         job->s_fence = NULL;
 609 }
 610 EXPORT_SYMBOL(drm_sched_job_cleanup);
 611
 612 /**
 613  * drm_sched_ready - is the scheduler ready
 614  *
 615  * @sched: scheduler instance
 616  *
 617  * Return true if we can push more jobs to the hw, otherwise false.
 618  */
 619 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 620 {
 621         return atomic_read(&sched->hw_rq_count) <
 622                 sched->hw_submission_limit;
 623 }
 624
 625 /**
 626  * drm_sched_wakeup - Wake up the scheduler when it is ready
 627  *
 628  * @sched: scheduler instance
 629  *
 630  */
 631 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 632 {
 633         if (drm_sched_ready(sched))
 634                 wake_up_interruptible(&sched->wake_up_worker);
 635 }
 636
 637 /**
 638  * drm_sched_select_entity - Select next entity to process
 639  *
 640  * @sched: scheduler instance
 641  *
 642  * Returns the entity to process or NULL if none are found.
 643  */
 644 static struct drm_sched_entity *
 645 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 646 {
 647         struct drm_sched_entity *entity;
 648         int i;
 649
 650         if (!drm_sched_ready(sched))
 651                 return NULL;
 652
 653         /* Kernel run queue has higher priority than normal run queue*/
 654         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 655                 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
 656                 if (entity)
 657                         break;
 658         }
 659
 660         return entity;
 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 pending 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 pending_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->pending_list,
 688                                        struct drm_sched_job, list);
 689
 690         if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
 691                 /* remove job from pending_list */
 692                 list_del_init(&job->list);
 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_job_done_cb);
 813                         if (r == -ENOENT)
 814                                 drm_sched_job_done(sched_job);
 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_job_done(sched_job);
 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  * @score: optional score atomic shared with other schedulers
 840  * @name: name used for debugging
 841  *
 842  * Return 0 on success, otherwise error code.
 843  */
 844 int drm_sched_init(struct drm_gpu_scheduler *sched,
 845                    const struct drm_sched_backend_ops *ops,
 846                    unsigned hw_submission, unsigned hang_limit, long timeout,
 847                    atomic_t *score, const char *name)
 848 {
 849         int i, ret;
 850         sched->ops = ops;
 851         sched->hw_submission_limit = hw_submission;
 852         sched->name = name;
 853         sched->timeout = timeout;
 854         sched->hang_limit = hang_limit;
 855         sched->score = score ? score : &sched->_score;
 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->pending_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         /* Confirm no work left behind accessing device structures */
 895         cancel_delayed_work_sync(&sched->work_tdr);
 896
 897         sched->ready = false;
 898 }
 899 EXPORT_SYMBOL(drm_sched_fini);
 900
 901 /**
 902  * drm_sched_increase_karma_ext - Update sched_entity guilty flag
 903  *
 904  * @bad: The job guilty of time out
 905  * @type: type for increase/reset karma
 906  *
 907  */
 908 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type)
 909 {
 910         int i;
 911         struct drm_sched_entity *tmp;
 912         struct drm_sched_entity *entity;
 913         struct drm_gpu_scheduler *sched = bad->sched;
 914
 915         /* don't change @bad's karma if it's from KERNEL RQ,
 916          * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
 917          * corrupt but keep in mind that kernel jobs always considered good.
 918          */
 919         if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 920                 if (type == 0)
 921                         atomic_set(&bad->karma, 0);
 922                 else if (type == 1)
 923                         atomic_inc(&bad->karma);
 924
 925                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
 926                      i++) {
 927                         struct drm_sched_rq *rq = &sched->sched_rq[i];
 928
 929                         spin_lock(&rq->lock);
 930                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
 931                                 if (bad->s_fence->scheduled.context ==
 932                                     entity->fence_context) {
 933                                         if (entity->guilty)
 934                                                 atomic_set(entity->guilty, type);
 935                                         break;
 936                                 }
 937                         }
 938                         spin_unlock(&rq->lock);
 939                         if (&entity->list != &rq->entities)
 940                                 break;
 941                 }
 942         }
 943 }
 944 EXPORT_SYMBOL(drm_sched_increase_karma_ext);