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  * @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         atomic_inc(&rq->sched->score);
  96         list_add_tail(&entity->list, &rq->entities);
  97         spin_unlock(&rq->lock);
  98 }
  99
 100 /**
 101  * drm_sched_rq_remove_entity - remove an entity
 102  *
 103  * @rq: scheduler run queue
 104  * @entity: scheduler entity
 105  *
 106  * Removes a scheduler entity from the run queue.
 107  */
 108 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 109                                 struct drm_sched_entity *entity)
 110 {
 111         if (list_empty(&entity->list))
 112                 return;
 113         spin_lock(&rq->lock);
 114         atomic_dec(&rq->sched->score);
 115         list_del_init(&entity->list);
 116         if (rq->current_entity == entity)
 117                 rq->current_entity = NULL;
 118         spin_unlock(&rq->lock);
 119 }
 120
 121 /**
 122  * drm_sched_rq_select_entity - Select an entity which could provide a job to run
 123  *
 124  * @rq: scheduler run queue to check.
 125  *
 126  * Try to find a ready entity, returns NULL if none found.
 127  */
 128 static struct drm_sched_entity *
 129 drm_sched_rq_select_entity(struct drm_sched_rq *rq)
 130 {
 131         struct drm_sched_entity *entity;
 132
 133         spin_lock(&rq->lock);
 134
 135         entity = rq->current_entity;
 136         if (entity) {
 137                 list_for_each_entry_continue(entity, &rq->entities, list) {
 138                         if (drm_sched_entity_is_ready(entity)) {
 139                                 rq->current_entity = entity;
 140                                 reinit_completion(&entity->entity_idle);
 141                                 spin_unlock(&rq->lock);
 142                                 return entity;
 143                         }
 144                 }
 145         }
 146
 147         list_for_each_entry(entity, &rq->entities, list) {
 148
 149                 if (drm_sched_entity_is_ready(entity)) {
 150                         rq->current_entity = entity;
 151                         reinit_completion(&entity->entity_idle);
 152                         spin_unlock(&rq->lock);
 153                         return entity;
 154                 }
 155
 156                 if (entity == rq->current_entity)
 157                         break;
 158         }
 159
 160         spin_unlock(&rq->lock);
 161
 162         return NULL;
 163 }
 164
 165 /**
 166  * drm_sched_dependency_optimized
 167  *
 168  * @fence: the dependency fence
 169  * @entity: the entity which depends on the above fence
 170  *
 171  * Returns true if the dependency can be optimized and false otherwise
 172  */
 173 bool drm_sched_dependency_optimized(struct dma_fence* fence,
 174                                     struct drm_sched_entity *entity)
 175 {
 176         struct drm_gpu_scheduler *sched = entity->rq->sched;
 177         struct drm_sched_fence *s_fence;
 178
 179         if (!fence || dma_fence_is_signaled(fence))
 180                 return false;
 181         if (fence->context == entity->fence_context)
 182                 return true;
 183         s_fence = to_drm_sched_fence(fence);
 184         if (s_fence && s_fence->sched == sched)
 185                 return true;
 186
 187         return false;
 188 }
 189 EXPORT_SYMBOL(drm_sched_dependency_optimized);
 190
 191 /**
 192  * drm_sched_start_timeout - start timeout for reset worker
 193  *
 194  * @sched: scheduler instance to start the worker for
 195  *
 196  * Start the timeout for the given scheduler.
 197  */
 198 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
 199 {
 200         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 201             !list_empty(&sched->ring_mirror_list))
 202                 schedule_delayed_work(&sched->work_tdr, sched->timeout);
 203 }
 204
 205 /**
 206  * drm_sched_fault - immediately start timeout handler
 207  *
 208  * @sched: scheduler where the timeout handling should be started.
 209  *
 210  * Start timeout handling immediately when the driver detects a hardware fault.
 211  */
 212 void drm_sched_fault(struct drm_gpu_scheduler *sched)
 213 {
 214         mod_delayed_work(system_wq, &sched->work_tdr, 0);
 215 }
 216 EXPORT_SYMBOL(drm_sched_fault);
 217
 218 /**
 219  * drm_sched_suspend_timeout - Suspend scheduler job timeout
 220  *
 221  * @sched: scheduler instance for which to suspend the timeout
 222  *
 223  * Suspend the delayed work timeout for the scheduler. This is done by
 224  * modifying the delayed work timeout to an arbitrary large value,
 225  * MAX_SCHEDULE_TIMEOUT in this case.
 226  *
 227  * Returns the timeout remaining
 228  *
 229  */
 230 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
 231 {
 232         unsigned long sched_timeout, now = jiffies;
 233
 234         sched_timeout = sched->work_tdr.timer.expires;
 235
 236         /*
 237          * Modify the timeout to an arbitrarily large value. This also prevents
 238          * the timeout to be restarted when new submissions arrive
 239          */
 240         if (mod_delayed_work(system_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
 241                         && time_after(sched_timeout, now))
 242                 return sched_timeout - now;
 243         else
 244                 return sched->timeout;
 245 }
 246 EXPORT_SYMBOL(drm_sched_suspend_timeout);
 247
 248 /**
 249  * drm_sched_resume_timeout - Resume scheduler job timeout
 250  *
 251  * @sched: scheduler instance for which to resume the timeout
 252  * @remaining: remaining timeout
 253  *
 254  * Resume the delayed work timeout for the scheduler.
 255  */
 256 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
 257                 unsigned long remaining)
 258 {
 259         spin_lock(&sched->job_list_lock);
 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(&sched->job_list_lock);
 267 }
 268 EXPORT_SYMBOL(drm_sched_resume_timeout);
 269
 270 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 271 {
 272         struct drm_gpu_scheduler *sched = s_job->sched;
 273
 274         spin_lock(&sched->job_list_lock);
 275         list_add_tail(&s_job->node, &sched->ring_mirror_list);
 276         drm_sched_start_timeout(sched);
 277         spin_unlock(&sched->job_list_lock);
 278 }
 279
 280 static void drm_sched_job_timedout(struct work_struct *work)
 281 {
 282         struct drm_gpu_scheduler *sched;
 283         struct drm_sched_job *job;
 284
 285         sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 286
 287         /* Protects against concurrent deletion in drm_sched_get_cleanup_job */
 288         spin_lock(&sched->job_list_lock);
 289         job = list_first_entry_or_null(&sched->ring_mirror_list,
 290                                        struct drm_sched_job, node);
 291
 292         if (job) {
 293                 /*
 294                  * Remove the bad job so it cannot be freed by concurrent
 295                  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
 296                  * is parked at which point it's safe.
 297                  */
 298                 list_del_init(&job->node);
 299                 spin_unlock(&sched->job_list_lock);
 300
 301                 job->sched->ops->timedout_job(job);
 302
 303                 /*
 304                  * Guilty job did complete and hence needs to be manually removed
 305                  * See drm_sched_stop doc.
 306                  */
 307                 if (sched->free_guilty) {
 308                         job->sched->ops->free_job(job);
 309                         sched->free_guilty = false;
 310                 }
 311         } else {
 312                 spin_unlock(&sched->job_list_lock);
 313         }
 314
 315         spin_lock(&sched->job_list_lock);
 316         drm_sched_start_timeout(sched);
 317         spin_unlock(&sched->job_list_lock);
 318 }
 319
 320  /**
 321   * drm_sched_increase_karma - Update sched_entity guilty flag
 322   *
 323   * @bad: The job guilty of time out
 324   *
 325   * Increment on every hang caused by the 'bad' job. If this exceeds the hang
 326   * limit of the scheduler then the respective sched entity is marked guilty and
 327   * jobs from it will not be scheduled further
 328   */
 329 void drm_sched_increase_karma(struct drm_sched_job *bad)
 330 {
 331         int i;
 332         struct drm_sched_entity *tmp;
 333         struct drm_sched_entity *entity;
 334         struct drm_gpu_scheduler *sched = bad->sched;
 335
 336         /* don't increase @bad's karma if it's from KERNEL RQ,
 337          * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
 338          * corrupt but keep in mind that kernel jobs always considered good.
 339          */
 340         if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 341                 atomic_inc(&bad->karma);
 342                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
 343                      i++) {
 344                         struct drm_sched_rq *rq = &sched->sched_rq[i];
 345
 346                         spin_lock(&rq->lock);
 347                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
 348                                 if (bad->s_fence->scheduled.context ==
 349                                     entity->fence_context) {
 350                                         if (atomic_read(&bad->karma) >
 351                                             bad->sched->hang_limit)
 352                                                 if (entity->guilty)
 353                                                         atomic_set(entity->guilty, 1);
 354                                         break;
 355                                 }
 356                         }
 357                         spin_unlock(&rq->lock);
 358                         if (&entity->list != &rq->entities)
 359                                 break;
 360                 }
 361         }
 362 }
 363 EXPORT_SYMBOL(drm_sched_increase_karma);
 364
 365 /**
 366  * drm_sched_stop - stop the scheduler
 367  *
 368  * @sched: scheduler instance
 369  * @bad: job which caused the time out
 370  *
 371  * Stop the scheduler and also removes and frees all completed jobs.
 372  * Note: bad job will not be freed as it might be used later and so it's
 373  * callers responsibility to release it manually if it's not part of the
 374  * mirror list any more.
 375  *
 376  */
 377 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 378 {
 379         struct drm_sched_job *s_job, *tmp;
 380
 381         kthread_park(sched->thread);
 382
 383         /*
 384          * Reinsert back the bad job here - now it's safe as
 385          * drm_sched_get_cleanup_job cannot race against us and release the
 386          * bad job at this point - we parked (waited for) any in progress
 387          * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
 388          * now until the scheduler thread is unparked.
 389          */
 390         if (bad && bad->sched == sched)
 391                 /*
 392                  * Add at the head of the queue to reflect it was the earliest
 393                  * job extracted.
 394                  */
 395                 list_add(&bad->node, &sched->ring_mirror_list);
 396
 397         /*
 398          * Iterate the job list from later to  earlier one and either deactive
 399          * their HW callbacks or remove them from mirror list if they already
 400          * signaled.
 401          * This iteration is thread safe as sched thread is stopped.
 402          */
 403         list_for_each_entry_safe_reverse(s_job, tmp, &sched->ring_mirror_list, node) {
 404                 if (s_job->s_fence->parent &&
 405                     dma_fence_remove_callback(s_job->s_fence->parent,
 406                                               &s_job->cb)) {
 407                         atomic_dec(&sched->hw_rq_count);
 408                 } else {
 409                         /*
 410                          * remove job from ring_mirror_list.
 411                          * Locking here is for concurrent resume timeout
 412                          */
 413                         spin_lock(&sched->job_list_lock);
 414                         list_del_init(&s_job->node);
 415                         spin_unlock(&sched->job_list_lock);
 416
 417                         /*
 418                          * Wait for job's HW fence callback to finish using s_job
 419                          * before releasing it.
 420                          *
 421                          * Job is still alive so fence refcount at least 1
 422                          */
 423                         dma_fence_wait(&s_job->s_fence->finished, false);
 424
 425                         /*
 426                          * We must keep bad job alive for later use during
 427                          * recovery by some of the drivers but leave a hint
 428                          * that the guilty job must be released.
 429                          */
 430                         if (bad != s_job)
 431                                 sched->ops->free_job(s_job);
 432                         else
 433                                 sched->free_guilty = true;
 434                 }
 435         }
 436
 437         /*
 438          * Stop pending timer in flight as we rearm it in  drm_sched_start. This
 439          * avoids the pending timeout work in progress to fire right away after
 440          * this TDR finished and before the newly restarted jobs had a
 441          * chance to complete.
 442          */
 443         cancel_delayed_work(&sched->work_tdr);
 444 }
 445
 446 EXPORT_SYMBOL(drm_sched_stop);
 447
 448 /**
 449  * drm_sched_job_recovery - recover jobs after a reset
 450  *
 451  * @sched: scheduler instance
 452  * @full_recovery: proceed with complete sched restart
 453  *
 454  */
 455 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
 456 {
 457         struct drm_sched_job *s_job, *tmp;
 458         int r;
 459
 460         /*
 461          * Locking the list is not required here as the sched thread is parked
 462          * so no new jobs are being inserted or removed. Also concurrent
 463          * GPU recovers can't run in parallel.
 464          */
 465         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 466                 struct dma_fence *fence = s_job->s_fence->parent;
 467
 468                 atomic_inc(&sched->hw_rq_count);
 469
 470                 if (!full_recovery)
 471                         continue;
 472
 473                 if (fence) {
 474                         r = dma_fence_add_callback(fence, &s_job->cb,
 475                                                    drm_sched_process_job);
 476                         if (r == -ENOENT)
 477                                 drm_sched_process_job(fence, &s_job->cb);
 478                         else if (r)
 479                                 DRM_ERROR("fence add callback failed (%d)\n",
 480                                           r);
 481                 } else
 482                         drm_sched_process_job(NULL, &s_job->cb);
 483         }
 484
 485         if (full_recovery) {
 486                 spin_lock(&sched->job_list_lock);
 487                 drm_sched_start_timeout(sched);
 488                 spin_unlock(&sched->job_list_lock);
 489         }
 490
 491         kthread_unpark(sched->thread);
 492 }
 493 EXPORT_SYMBOL(drm_sched_start);
 494
 495 /**
 496  * drm_sched_resubmit_jobs - helper to relunch job from mirror ring list
 497  *
 498  * @sched: scheduler instance
 499  *
 500  */
 501 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
 502 {
 503         struct drm_sched_job *s_job, *tmp;
 504         uint64_t guilty_context;
 505         bool found_guilty = false;
 506         struct dma_fence *fence;
 507
 508         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 509                 struct drm_sched_fence *s_fence = s_job->s_fence;
 510
 511                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 512                         found_guilty = true;
 513                         guilty_context = s_job->s_fence->scheduled.context;
 514                 }
 515
 516                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 517                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 518
 519                 dma_fence_put(s_job->s_fence->parent);
 520                 fence = sched->ops->run_job(s_job);
 521
 522                 if (IS_ERR_OR_NULL(fence)) {
 523                         if (IS_ERR(fence))
 524                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
 525
 526                         s_job->s_fence->parent = NULL;
 527                 } else {
 528                         s_job->s_fence->parent = fence;
 529                 }
 530
 531
 532         }
 533 }
 534 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 535
 536 /**
 537  * drm_sched_job_init - init a scheduler job
 538  *
 539  * @job: scheduler job to init
 540  * @entity: scheduler entity to use
 541  * @owner: job owner for debugging
 542  *
 543  * Refer to drm_sched_entity_push_job() documentation
 544  * for locking considerations.
 545  *
 546  * Returns 0 for success, negative error code otherwise.
 547  */
 548 int drm_sched_job_init(struct drm_sched_job *job,
 549                        struct drm_sched_entity *entity,
 550                        void *owner)
 551 {
 552         struct drm_gpu_scheduler *sched;
 553
 554         drm_sched_entity_select_rq(entity);
 555         if (!entity->rq)
 556                 return -ENOENT;
 557
 558         sched = entity->rq->sched;
 559
 560         job->sched = sched;
 561         job->entity = entity;
 562         job->s_priority = entity->rq - sched->sched_rq;
 563         job->s_fence = drm_sched_fence_create(entity, owner);
 564         if (!job->s_fence)
 565                 return -ENOMEM;
 566         job->id = atomic64_inc_return(&sched->job_id_count);
 567
 568         INIT_LIST_HEAD(&job->node);
 569
 570         return 0;
 571 }
 572 EXPORT_SYMBOL(drm_sched_job_init);
 573
 574 /**
 575  * drm_sched_job_cleanup - clean up scheduler job resources
 576  *
 577  * @job: scheduler job to clean up
 578  */
 579 void drm_sched_job_cleanup(struct drm_sched_job *job)
 580 {
 581         dma_fence_put(&job->s_fence->finished);
 582         job->s_fence = NULL;
 583 }
 584 EXPORT_SYMBOL(drm_sched_job_cleanup);
 585
 586 /**
 587  * drm_sched_ready - is the scheduler ready
 588  *
 589  * @sched: scheduler instance
 590  *
 591  * Return true if we can push more jobs to the hw, otherwise false.
 592  */
 593 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 594 {
 595         return atomic_read(&sched->hw_rq_count) <
 596                 sched->hw_submission_limit;
 597 }
 598
 599 /**
 600  * drm_sched_wakeup - Wake up the scheduler when it is ready
 601  *
 602  * @sched: scheduler instance
 603  *
 604  */
 605 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 606 {
 607         if (drm_sched_ready(sched))
 608                 wake_up_interruptible(&sched->wake_up_worker);
 609 }
 610
 611 /**
 612  * drm_sched_select_entity - Select next entity to process
 613  *
 614  * @sched: scheduler instance
 615  *
 616  * Returns the entity to process or NULL if none are found.
 617  */
 618 static struct drm_sched_entity *
 619 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 620 {
 621         struct drm_sched_entity *entity;
 622         int i;
 623
 624         if (!drm_sched_ready(sched))
 625                 return NULL;
 626
 627         /* Kernel run queue has higher priority than normal run queue*/
 628         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 629                 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
 630                 if (entity)
 631                         break;
 632         }
 633
 634         return entity;
 635 }
 636
 637 /**
 638  * drm_sched_process_job - process a job
 639  *
 640  * @f: fence
 641  * @cb: fence callbacks
 642  *
 643  * Called after job has finished execution.
 644  */
 645 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
 646 {
 647         struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
 648         struct drm_sched_fence *s_fence = s_job->s_fence;
 649         struct drm_gpu_scheduler *sched = s_fence->sched;
 650
 651         atomic_dec(&sched->hw_rq_count);
 652         atomic_dec(&sched->score);
 653
 654         trace_drm_sched_process_job(s_fence);
 655
 656         dma_fence_get(&s_fence->finished);
 657         drm_sched_fence_finished(s_fence);
 658         dma_fence_put(&s_fence->finished);
 659         wake_up_interruptible(&sched->wake_up_worker);
 660 }
 661
 662 /**
 663  * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
 664  *
 665  * @sched: scheduler instance
 666  *
 667  * Returns the next finished job from the mirror list (if there is one)
 668  * ready for it to be destroyed.
 669  */
 670 static struct drm_sched_job *
 671 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
 672 {
 673         struct drm_sched_job *job;
 674
 675         /*
 676          * Don't destroy jobs while the timeout worker is running  OR thread
 677          * is being parked and hence assumed to not touch ring_mirror_list
 678          */
 679         if ((sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 680             !cancel_delayed_work(&sched->work_tdr)) ||
 681             kthread_should_park())
 682                 return NULL;
 683
 684         spin_lock(&sched->job_list_lock);
 685
 686         job = list_first_entry_or_null(&sched->ring_mirror_list,
 687                                        struct drm_sched_job, node);
 688
 689         if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
 690                 /* remove job from ring_mirror_list */
 691                 list_del_init(&job->node);
 692         } else {
 693                 job = NULL;
 694                 /* queue timeout for next job */
 695                 drm_sched_start_timeout(sched);
 696         }
 697
 698         spin_unlock(&sched->job_list_lock);
 699
 700         return job;
 701 }
 702
 703 /**
 704  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
 705  * @sched_list: list of drm_gpu_schedulers
 706  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
 707  *
 708  * Returns pointer of the sched with the least load or NULL if none of the
 709  * drm_gpu_schedulers are ready
 710  */
 711 struct drm_gpu_scheduler *
 712 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
 713                      unsigned int num_sched_list)
 714 {
 715         struct drm_gpu_scheduler *sched, *picked_sched = NULL;
 716         int i;
 717         unsigned int min_score = UINT_MAX, num_score;
 718
 719         for (i = 0; i < num_sched_list; ++i) {
 720                 sched = sched_list[i];
 721
 722                 if (!sched->ready) {
 723                         DRM_WARN("scheduler %s is not ready, skipping",
 724                                  sched->name);
 725                         continue;
 726                 }
 727
 728                 num_score = atomic_read(&sched->score);
 729                 if (num_score < min_score) {
 730                         min_score = num_score;
 731                         picked_sched = sched;
 732                 }
 733         }
 734
 735         return picked_sched;
 736 }
 737 EXPORT_SYMBOL(drm_sched_pick_best);
 738
 739 /**
 740  * drm_sched_blocked - check if the scheduler is blocked
 741  *
 742  * @sched: scheduler instance
 743  *
 744  * Returns true if blocked, otherwise false.
 745  */
 746 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 747 {
 748         if (kthread_should_park()) {
 749                 kthread_parkme();
 750                 return true;
 751         }
 752
 753         return false;
 754 }
 755
 756 /**
 757  * drm_sched_main - main scheduler thread
 758  *
 759  * @param: scheduler instance
 760  *
 761  * Returns 0.
 762  */
 763 static int drm_sched_main(void *param)
 764 {
 765         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
 766         int r;
 767
 768         sched_set_fifo_low(current);
 769
 770         while (!kthread_should_stop()) {
 771                 struct drm_sched_entity *entity = NULL;
 772                 struct drm_sched_fence *s_fence;
 773                 struct drm_sched_job *sched_job;
 774                 struct dma_fence *fence;
 775                 struct drm_sched_job *cleanup_job = NULL;
 776
 777                 wait_event_interruptible(sched->wake_up_worker,
 778                                          (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
 779                                          (!drm_sched_blocked(sched) &&
 780                                           (entity = drm_sched_select_entity(sched))) ||
 781                                          kthread_should_stop());
 782
 783                 if (cleanup_job) {
 784                         sched->ops->free_job(cleanup_job);
 785                         /* queue timeout for next job */
 786                         drm_sched_start_timeout(sched);
 787                 }
 788
 789                 if (!entity)
 790                         continue;
 791
 792                 sched_job = drm_sched_entity_pop_job(entity);
 793
 794                 complete(&entity->entity_idle);
 795
 796                 if (!sched_job)
 797                         continue;
 798
 799                 s_fence = sched_job->s_fence;
 800
 801                 atomic_inc(&sched->hw_rq_count);
 802                 drm_sched_job_begin(sched_job);
 803
 804                 trace_drm_run_job(sched_job, entity);
 805                 fence = sched->ops->run_job(sched_job);
 806                 drm_sched_fence_scheduled(s_fence);
 807
 808                 if (!IS_ERR_OR_NULL(fence)) {
 809                         s_fence->parent = dma_fence_get(fence);
 810                         r = dma_fence_add_callback(fence, &sched_job->cb,
 811                                                    drm_sched_process_job);
 812                         if (r == -ENOENT)
 813                                 drm_sched_process_job(fence, &sched_job->cb);
 814                         else if (r)
 815                                 DRM_ERROR("fence add callback failed (%d)\n",
 816                                           r);
 817                         dma_fence_put(fence);
 818                 } else {
 819                         if (IS_ERR(fence))
 820                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
 821
 822                         drm_sched_process_job(NULL, &sched_job->cb);
 823                 }
 824
 825                 wake_up(&sched->job_scheduled);
 826         }
 827         return 0;
 828 }
 829
 830 /**
 831  * drm_sched_init - Init a gpu scheduler instance
 832  *
 833  * @sched: scheduler instance
 834  * @ops: backend operations for this scheduler
 835  * @hw_submission: number of hw submissions that can be in flight
 836  * @hang_limit: number of times to allow a job to hang before dropping it
 837  * @timeout: timeout value in jiffies for the scheduler
 838  * @name: name used for debugging
 839  *
 840  * Return 0 on success, otherwise error code.
 841  */
 842 int drm_sched_init(struct drm_gpu_scheduler *sched,
 843                    const struct drm_sched_backend_ops *ops,
 844                    unsigned hw_submission,
 845                    unsigned hang_limit,
 846                    long timeout,
 847                    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         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
 856                 drm_sched_rq_init(sched, &sched->sched_rq[i]);
 857
 858         init_waitqueue_head(&sched->wake_up_worker);
 859         init_waitqueue_head(&sched->job_scheduled);
 860         INIT_LIST_HEAD(&sched->ring_mirror_list);
 861         spin_lock_init(&sched->job_list_lock);
 862         atomic_set(&sched->hw_rq_count, 0);
 863         INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
 864         atomic_set(&sched->score, 0);
 865         atomic64_set(&sched->job_id_count, 0);
 866
 867         /* Each scheduler will run on a seperate kernel thread */
 868         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
 869         if (IS_ERR(sched->thread)) {
 870                 ret = PTR_ERR(sched->thread);
 871                 sched->thread = NULL;
 872                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 873                 return ret;
 874         }
 875
 876         sched->ready = true;
 877         return 0;
 878 }
 879 EXPORT_SYMBOL(drm_sched_init);
 880
 881 /**
 882  * drm_sched_fini - Destroy a gpu scheduler
 883  *
 884  * @sched: scheduler instance
 885  *
 886  * Tears down and cleans up the scheduler.
 887  */
 888 void drm_sched_fini(struct drm_gpu_scheduler *sched)
 889 {
 890         struct drm_sched_entity *s_entity;
 891         int i;
 892
 893         if (sched->thread)
 894                 kthread_stop(sched->thread);
 895
 896         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 897                 struct drm_sched_rq *rq = &sched->sched_rq[i];
 898
 899                 if (!rq)
 900                         continue;
 901
 902                 spin_lock(&rq->lock);
 903                 list_for_each_entry(s_entity, &rq->entities, list)
 904                         /*
 905                          * Prevents reinsertion and marks job_queue as idle,
 906                          * it will removed from rq in drm_sched_entity_fini
 907                          * eventually
 908                          */
 909                         s_entity->stopped = true;
 910                 spin_unlock(&rq->lock);
 911
 912         }
 913
 914         /* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
 915         wake_up_all(&sched->job_scheduled);
 916
 917         /* Confirm no work left behind accessing device structures */
 918         cancel_delayed_work_sync(&sched->work_tdr);
 919
 920         sched->ready = false;
 921 }
 922 EXPORT_SYMBOL(drm_sched_fini);