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  * Note that once a job was taken from the entities queue and pushed to the
  47  * hardware, i.e. the pending queue, the entity must not be referenced anymore
  48  * through the jobs entity pointer.
  49  */
  50
  51 #include <linux/kthread.h>
  52 #include <linux/wait.h>
  53 #include <linux/sched.h>
  54 #include <linux/completion.h>
  55 #include <linux/dma-resv.h>
  56 #include <uapi/linux/sched/types.h>
  57
  58 #include <drm/drm_print.h>
  59 #include <drm/drm_gem.h>
  60 #include <drm/drm_syncobj.h>
  61 #include <drm/gpu_scheduler.h>
  62 #include <drm/spsc_queue.h>
  63
  64 #define CREATE_TRACE_POINTS
  65 #include "gpu_scheduler_trace.h"
  66
  67 #define to_drm_sched_job(sched_job)             \
  68                 container_of((sched_job), struct drm_sched_job, queue_node)
  69
  70 int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
  71
  72 /**
  73  * DOC: sched_policy (int)
  74  * Used to override default entities scheduling policy in a run queue.
  75  */
  76 MODULE_PARM_DESC(sched_policy, "Specify the scheduling policy for entities on a run-queue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin, " __stringify(DRM_SCHED_POLICY_FIFO) " = FIFO (default).");
  77 module_param_named(sched_policy, drm_sched_policy, int, 0444);
  78
  79 static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
  80                                                             const struct rb_node *b)
  81 {
  82         struct drm_sched_entity *ent_a =  rb_entry((a), struct drm_sched_entity, rb_tree_node);
  83         struct drm_sched_entity *ent_b =  rb_entry((b), struct drm_sched_entity, rb_tree_node);
  84
  85         return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
  86 }
  87
  88 static inline void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity)
  89 {
  90         struct drm_sched_rq *rq = entity->rq;
  91
  92         if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
  93                 rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
  94                 RB_CLEAR_NODE(&entity->rb_tree_node);
  95         }
  96 }
  97
  98 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts)
  99 {
 100         /*
 101          * Both locks need to be grabbed, one to protect from entity->rq change
 102          * for entity from within concurrent drm_sched_entity_select_rq and the
 103          * other to update the rb tree structure.
 104          */
 105         spin_lock(&entity->rq_lock);
 106         spin_lock(&entity->rq->lock);
 107
 108         drm_sched_rq_remove_fifo_locked(entity);
 109
 110         entity->oldest_job_waiting = ts;
 111
 112         rb_add_cached(&entity->rb_tree_node, &entity->rq->rb_tree_root,
 113                       drm_sched_entity_compare_before);
 114
 115         spin_unlock(&entity->rq->lock);
 116         spin_unlock(&entity->rq_lock);
 117 }
 118
 119 /**
 120  * drm_sched_rq_init - initialize a given run queue struct
 121  *
 122  * @sched: scheduler instance to associate with this run queue
 123  * @rq: scheduler run queue
 124  *
 125  * Initializes a scheduler runqueue.
 126  */
 127 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
 128                               struct drm_sched_rq *rq)
 129 {
 130         spin_lock_init(&rq->lock);
 131         INIT_LIST_HEAD(&rq->entities);
 132         rq->rb_tree_root = RB_ROOT_CACHED;
 133         rq->current_entity = NULL;
 134         rq->sched = sched;
 135 }
 136
 137 /**
 138  * drm_sched_rq_add_entity - add an entity
 139  *
 140  * @rq: scheduler run queue
 141  * @entity: scheduler entity
 142  *
 143  * Adds a scheduler entity to the run queue.
 144  */
 145 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
 146                              struct drm_sched_entity *entity)
 147 {
 148         if (!list_empty(&entity->list))
 149                 return;
 150
 151         spin_lock(&rq->lock);
 152
 153         atomic_inc(rq->sched->score);
 154         list_add_tail(&entity->list, &rq->entities);
 155
 156         spin_unlock(&rq->lock);
 157 }
 158
 159 /**
 160  * drm_sched_rq_remove_entity - remove an entity
 161  *
 162  * @rq: scheduler run queue
 163  * @entity: scheduler entity
 164  *
 165  * Removes a scheduler entity from the run queue.
 166  */
 167 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 168                                 struct drm_sched_entity *entity)
 169 {
 170         if (list_empty(&entity->list))
 171                 return;
 172
 173         spin_lock(&rq->lock);
 174
 175         atomic_dec(rq->sched->score);
 176         list_del_init(&entity->list);
 177
 178         if (rq->current_entity == entity)
 179                 rq->current_entity = NULL;
 180
 181         if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 182                 drm_sched_rq_remove_fifo_locked(entity);
 183
 184         spin_unlock(&rq->lock);
 185 }
 186
 187 /**
 188  * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
 189  *
 190  * @rq: scheduler run queue to check.
 191  *
 192  * Try to find a ready entity, returns NULL if none found.
 193  */
 194 static struct drm_sched_entity *
 195 drm_sched_rq_select_entity_rr(struct drm_sched_rq *rq)
 196 {
 197         struct drm_sched_entity *entity;
 198
 199         spin_lock(&rq->lock);
 200
 201         entity = rq->current_entity;
 202         if (entity) {
 203                 list_for_each_entry_continue(entity, &rq->entities, list) {
 204                         if (drm_sched_entity_is_ready(entity)) {
 205                                 rq->current_entity = entity;
 206                                 reinit_completion(&entity->entity_idle);
 207                                 spin_unlock(&rq->lock);
 208                                 return entity;
 209                         }
 210                 }
 211         }
 212
 213         list_for_each_entry(entity, &rq->entities, list) {
 214
 215                 if (drm_sched_entity_is_ready(entity)) {
 216                         rq->current_entity = entity;
 217                         reinit_completion(&entity->entity_idle);
 218                         spin_unlock(&rq->lock);
 219                         return entity;
 220                 }
 221
 222                 if (entity == rq->current_entity)
 223                         break;
 224         }
 225
 226         spin_unlock(&rq->lock);
 227
 228         return NULL;
 229 }
 230
 231 /**
 232  * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
 233  *
 234  * @rq: scheduler run queue to check.
 235  *
 236  * Find oldest waiting ready entity, returns NULL if none found.
 237  */
 238 static struct drm_sched_entity *
 239 drm_sched_rq_select_entity_fifo(struct drm_sched_rq *rq)
 240 {
 241         struct rb_node *rb;
 242
 243         spin_lock(&rq->lock);
 244         for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
 245                 struct drm_sched_entity *entity;
 246
 247                 entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
 248                 if (drm_sched_entity_is_ready(entity)) {
 249                         rq->current_entity = entity;
 250                         reinit_completion(&entity->entity_idle);
 251                         break;
 252                 }
 253         }
 254         spin_unlock(&rq->lock);
 255
 256         return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
 257 }
 258
 259 /**
 260  * drm_sched_job_done - complete a job
 261  * @s_job: pointer to the job which is done
 262  *
 263  * Finish the job's fence and wake up the worker thread.
 264  */
 265 static void drm_sched_job_done(struct drm_sched_job *s_job)
 266 {
 267         struct drm_sched_fence *s_fence = s_job->s_fence;
 268         struct drm_gpu_scheduler *sched = s_fence->sched;
 269
 270         atomic_dec(&sched->hw_rq_count);
 271         atomic_dec(sched->score);
 272
 273         trace_drm_sched_process_job(s_fence);
 274
 275         dma_fence_get(&s_fence->finished);
 276         drm_sched_fence_finished(s_fence);
 277         dma_fence_put(&s_fence->finished);
 278         wake_up_interruptible(&sched->wake_up_worker);
 279 }
 280
 281 /**
 282  * drm_sched_job_done_cb - the callback for a done job
 283  * @f: fence
 284  * @cb: fence callbacks
 285  */
 286 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
 287 {
 288         struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
 289
 290         drm_sched_job_done(s_job);
 291 }
 292
 293 /**
 294  * drm_sched_start_timeout - start timeout for reset worker
 295  *
 296  * @sched: scheduler instance to start the worker for
 297  *
 298  * Start the timeout for the given scheduler.
 299  */
 300 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
 301 {
 302         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 303             !list_empty(&sched->pending_list))
 304                 queue_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
 305 }
 306
 307 /**
 308  * drm_sched_fault - immediately start timeout handler
 309  *
 310  * @sched: scheduler where the timeout handling should be started.
 311  *
 312  * Start timeout handling immediately when the driver detects a hardware fault.
 313  */
 314 void drm_sched_fault(struct drm_gpu_scheduler *sched)
 315 {
 316         mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
 317 }
 318 EXPORT_SYMBOL(drm_sched_fault);
 319
 320 /**
 321  * drm_sched_suspend_timeout - Suspend scheduler job timeout
 322  *
 323  * @sched: scheduler instance for which to suspend the timeout
 324  *
 325  * Suspend the delayed work timeout for the scheduler. This is done by
 326  * modifying the delayed work timeout to an arbitrary large value,
 327  * MAX_SCHEDULE_TIMEOUT in this case.
 328  *
 329  * Returns the timeout remaining
 330  *
 331  */
 332 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
 333 {
 334         unsigned long sched_timeout, now = jiffies;
 335
 336         sched_timeout = sched->work_tdr.timer.expires;
 337
 338         /*
 339          * Modify the timeout to an arbitrarily large value. This also prevents
 340          * the timeout to be restarted when new submissions arrive
 341          */
 342         if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
 343                         && time_after(sched_timeout, now))
 344                 return sched_timeout - now;
 345         else
 346                 return sched->timeout;
 347 }
 348 EXPORT_SYMBOL(drm_sched_suspend_timeout);
 349
 350 /**
 351  * drm_sched_resume_timeout - Resume scheduler job timeout
 352  *
 353  * @sched: scheduler instance for which to resume the timeout
 354  * @remaining: remaining timeout
 355  *
 356  * Resume the delayed work timeout for the scheduler.
 357  */
 358 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
 359                 unsigned long remaining)
 360 {
 361         spin_lock(&sched->job_list_lock);
 362
 363         if (list_empty(&sched->pending_list))
 364                 cancel_delayed_work(&sched->work_tdr);
 365         else
 366                 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
 367
 368         spin_unlock(&sched->job_list_lock);
 369 }
 370 EXPORT_SYMBOL(drm_sched_resume_timeout);
 371
 372 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 373 {
 374         struct drm_gpu_scheduler *sched = s_job->sched;
 375
 376         spin_lock(&sched->job_list_lock);
 377         list_add_tail(&s_job->list, &sched->pending_list);
 378         drm_sched_start_timeout(sched);
 379         spin_unlock(&sched->job_list_lock);
 380 }
 381
 382 static void drm_sched_job_timedout(struct work_struct *work)
 383 {
 384         struct drm_gpu_scheduler *sched;
 385         struct drm_sched_job *job;
 386         enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
 387
 388         sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 389
 390         /* Protects against concurrent deletion in drm_sched_get_cleanup_job */
 391         spin_lock(&sched->job_list_lock);
 392         job = list_first_entry_or_null(&sched->pending_list,
 393                                        struct drm_sched_job, list);
 394
 395         if (job) {
 396                 /*
 397                  * Remove the bad job so it cannot be freed by concurrent
 398                  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
 399                  * is parked at which point it's safe.
 400                  */
 401                 list_del_init(&job->list);
 402                 spin_unlock(&sched->job_list_lock);
 403
 404                 status = job->sched->ops->timedout_job(job);
 405
 406                 /*
 407                  * Guilty job did complete and hence needs to be manually removed
 408                  * See drm_sched_stop doc.
 409                  */
 410                 if (sched->free_guilty) {
 411                         job->sched->ops->free_job(job);
 412                         sched->free_guilty = false;
 413                 }
 414         } else {
 415                 spin_unlock(&sched->job_list_lock);
 416         }
 417
 418         if (status != DRM_GPU_SCHED_STAT_ENODEV) {
 419                 spin_lock(&sched->job_list_lock);
 420                 drm_sched_start_timeout(sched);
 421                 spin_unlock(&sched->job_list_lock);
 422         }
 423 }
 424
 425 /**
 426  * drm_sched_stop - stop the scheduler
 427  *
 428  * @sched: scheduler instance
 429  * @bad: job which caused the time out
 430  *
 431  * Stop the scheduler and also removes and frees all completed jobs.
 432  * Note: bad job will not be freed as it might be used later and so it's
 433  * callers responsibility to release it manually if it's not part of the
 434  * pending list any more.
 435  *
 436  */
 437 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 438 {
 439         struct drm_sched_job *s_job, *tmp;
 440
 441         kthread_park(sched->thread);
 442
 443         /*
 444          * Reinsert back the bad job here - now it's safe as
 445          * drm_sched_get_cleanup_job cannot race against us and release the
 446          * bad job at this point - we parked (waited for) any in progress
 447          * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
 448          * now until the scheduler thread is unparked.
 449          */
 450         if (bad && bad->sched == sched)
 451                 /*
 452                  * Add at the head of the queue to reflect it was the earliest
 453                  * job extracted.
 454                  */
 455                 list_add(&bad->list, &sched->pending_list);
 456
 457         /*
 458          * Iterate the job list from later to  earlier one and either deactive
 459          * their HW callbacks or remove them from pending list if they already
 460          * signaled.
 461          * This iteration is thread safe as sched thread is stopped.
 462          */
 463         list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
 464                                          list) {
 465                 if (s_job->s_fence->parent &&
 466                     dma_fence_remove_callback(s_job->s_fence->parent,
 467                                               &s_job->cb)) {
 468                         dma_fence_put(s_job->s_fence->parent);
 469                         s_job->s_fence->parent = NULL;
 470                         atomic_dec(&sched->hw_rq_count);
 471                 } else {
 472                         /*
 473                          * remove job from pending_list.
 474                          * Locking here is for concurrent resume timeout
 475                          */
 476                         spin_lock(&sched->job_list_lock);
 477                         list_del_init(&s_job->list);
 478                         spin_unlock(&sched->job_list_lock);
 479
 480                         /*
 481                          * Wait for job's HW fence callback to finish using s_job
 482                          * before releasing it.
 483                          *
 484                          * Job is still alive so fence refcount at least 1
 485                          */
 486                         dma_fence_wait(&s_job->s_fence->finished, false);
 487
 488                         /*
 489                          * We must keep bad job alive for later use during
 490                          * recovery by some of the drivers but leave a hint
 491                          * that the guilty job must be released.
 492                          */
 493                         if (bad != s_job)
 494                                 sched->ops->free_job(s_job);
 495                         else
 496                                 sched->free_guilty = true;
 497                 }
 498         }
 499
 500         /*
 501          * Stop pending timer in flight as we rearm it in  drm_sched_start. This
 502          * avoids the pending timeout work in progress to fire right away after
 503          * this TDR finished and before the newly restarted jobs had a
 504          * chance to complete.
 505          */
 506         cancel_delayed_work(&sched->work_tdr);
 507 }
 508
 509 EXPORT_SYMBOL(drm_sched_stop);
 510
 511 /**
 512  * drm_sched_start - recover jobs after a reset
 513  *
 514  * @sched: scheduler instance
 515  * @full_recovery: proceed with complete sched restart
 516  *
 517  */
 518 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
 519 {
 520         struct drm_sched_job *s_job, *tmp;
 521         int r;
 522
 523         /*
 524          * Locking the list is not required here as the sched thread is parked
 525          * so no new jobs are being inserted or removed. Also concurrent
 526          * GPU recovers can't run in parallel.
 527          */
 528         list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 529                 struct dma_fence *fence = s_job->s_fence->parent;
 530
 531                 atomic_inc(&sched->hw_rq_count);
 532
 533                 if (!full_recovery)
 534                         continue;
 535
 536                 if (fence) {
 537                         r = dma_fence_add_callback(fence, &s_job->cb,
 538                                                    drm_sched_job_done_cb);
 539                         if (r == -ENOENT)
 540                                 drm_sched_job_done(s_job);
 541                         else if (r)
 542                                 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
 543                                           r);
 544                 } else
 545                         drm_sched_job_done(s_job);
 546         }
 547
 548         if (full_recovery) {
 549                 spin_lock(&sched->job_list_lock);
 550                 drm_sched_start_timeout(sched);
 551                 spin_unlock(&sched->job_list_lock);
 552         }
 553
 554         kthread_unpark(sched->thread);
 555 }
 556 EXPORT_SYMBOL(drm_sched_start);
 557
 558 /**
 559  * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
 560  *
 561  * @sched: scheduler instance
 562  *
 563  * Re-submitting jobs was a concept AMD came up as cheap way to implement
 564  * recovery after a job timeout.
 565  *
 566  * This turned out to be not working very well. First of all there are many
 567  * problem with the dma_fence implementation and requirements. Either the
 568  * implementation is risking deadlocks with core memory management or violating
 569  * documented implementation details of the dma_fence object.
 570  *
 571  * Drivers can still save and restore their state for recovery operations, but
 572  * we shouldn't make this a general scheduler feature around the dma_fence
 573  * interface.
 574  */
 575 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
 576 {
 577         struct drm_sched_job *s_job, *tmp;
 578         uint64_t guilty_context;
 579         bool found_guilty = false;
 580         struct dma_fence *fence;
 581
 582         list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 583                 struct drm_sched_fence *s_fence = s_job->s_fence;
 584
 585                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 586                         found_guilty = true;
 587                         guilty_context = s_job->s_fence->scheduled.context;
 588                 }
 589
 590                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 591                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 592
 593                 fence = sched->ops->run_job(s_job);
 594
 595                 if (IS_ERR_OR_NULL(fence)) {
 596                         if (IS_ERR(fence))
 597                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
 598
 599                         s_job->s_fence->parent = NULL;
 600                 } else {
 601
 602                         s_job->s_fence->parent = dma_fence_get(fence);
 603
 604                         /* Drop for orignal kref_init */
 605                         dma_fence_put(fence);
 606                 }
 607         }
 608 }
 609 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 610
 611 /**
 612  * drm_sched_job_init - init a scheduler job
 613  * @job: scheduler job to init
 614  * @entity: scheduler entity to use
 615  * @owner: job owner for debugging
 616  *
 617  * Refer to drm_sched_entity_push_job() documentation
 618  * for locking considerations.
 619  *
 620  * Drivers must make sure drm_sched_job_cleanup() if this function returns
 621  * successfully, even when @job is aborted before drm_sched_job_arm() is called.
 622  *
 623  * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
 624  * has died, which can mean that there's no valid runqueue for a @entity.
 625  * This function returns -ENOENT in this case (which probably should be -EIO as
 626  * a more meanigful return value).
 627  *
 628  * Returns 0 for success, negative error code otherwise.
 629  */
 630 int drm_sched_job_init(struct drm_sched_job *job,
 631                        struct drm_sched_entity *entity,
 632                        void *owner)
 633 {
 634         if (!entity->rq)
 635                 return -ENOENT;
 636
 637         job->entity = entity;
 638         job->s_fence = drm_sched_fence_alloc(entity, owner);
 639         if (!job->s_fence)
 640                 return -ENOMEM;
 641
 642         INIT_LIST_HEAD(&job->list);
 643
 644         xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
 645
 646         return 0;
 647 }
 648 EXPORT_SYMBOL(drm_sched_job_init);
 649
 650 /**
 651  * drm_sched_job_arm - arm a scheduler job for execution
 652  * @job: scheduler job to arm
 653  *
 654  * This arms a scheduler job for execution. Specifically it initializes the
 655  * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
 656  * or other places that need to track the completion of this job.
 657  *
 658  * Refer to drm_sched_entity_push_job() documentation for locking
 659  * considerations.
 660  *
 661  * This can only be called if drm_sched_job_init() succeeded.
 662  */
 663 void drm_sched_job_arm(struct drm_sched_job *job)
 664 {
 665         struct drm_gpu_scheduler *sched;
 666         struct drm_sched_entity *entity = job->entity;
 667
 668         BUG_ON(!entity);
 669         drm_sched_entity_select_rq(entity);
 670         sched = entity->rq->sched;
 671
 672         job->sched = sched;
 673         job->s_priority = entity->rq - sched->sched_rq;
 674         job->id = atomic64_inc_return(&sched->job_id_count);
 675
 676         drm_sched_fence_init(job->s_fence, job->entity);
 677 }
 678 EXPORT_SYMBOL(drm_sched_job_arm);
 679
 680 /**
 681  * drm_sched_job_add_dependency - adds the fence as a job dependency
 682  * @job: scheduler job to add the dependencies to
 683  * @fence: the dma_fence to add to the list of dependencies.
 684  *
 685  * Note that @fence is consumed in both the success and error cases.
 686  *
 687  * Returns:
 688  * 0 on success, or an error on failing to expand the array.
 689  */
 690 int drm_sched_job_add_dependency(struct drm_sched_job *job,
 691                                  struct dma_fence *fence)
 692 {
 693         struct dma_fence *entry;
 694         unsigned long index;
 695         u32 id = 0;
 696         int ret;
 697
 698         if (!fence)
 699                 return 0;
 700
 701         /* Deduplicate if we already depend on a fence from the same context.
 702          * This lets the size of the array of deps scale with the number of
 703          * engines involved, rather than the number of BOs.
 704          */
 705         xa_for_each(&job->dependencies, index, entry) {
 706                 if (entry->context != fence->context)
 707                         continue;
 708
 709                 if (dma_fence_is_later(fence, entry)) {
 710                         dma_fence_put(entry);
 711                         xa_store(&job->dependencies, index, fence, GFP_KERNEL);
 712                 } else {
 713                         dma_fence_put(fence);
 714                 }
 715                 return 0;
 716         }
 717
 718         ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
 719         if (ret != 0)
 720                 dma_fence_put(fence);
 721
 722         return ret;
 723 }
 724 EXPORT_SYMBOL(drm_sched_job_add_dependency);
 725
 726 /**
 727  * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
 728  * @job: scheduler job to add the dependencies to
 729  * @file: drm file private pointer
 730  * @handle: syncobj handle to lookup
 731  * @point: timeline point
 732  *
 733  * This adds the fence matching the given syncobj to @job.
 734  *
 735  * Returns:
 736  * 0 on success, or an error on failing to expand the array.
 737  */
 738 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
 739                                          struct drm_file *file,
 740                                          u32 handle,
 741                                          u32 point)
 742 {
 743         struct dma_fence *fence;
 744         int ret;
 745
 746         ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
 747         if (ret)
 748                 return ret;
 749
 750         return drm_sched_job_add_dependency(job, fence);
 751 }
 752 EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
 753
 754 /**
 755  * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
 756  * @job: scheduler job to add the dependencies to
 757  * @resv: the dma_resv object to get the fences from
 758  * @usage: the dma_resv_usage to use to filter the fences
 759  *
 760  * This adds all fences matching the given usage from @resv to @job.
 761  * Must be called with the @resv lock held.
 762  *
 763  * Returns:
 764  * 0 on success, or an error on failing to expand the array.
 765  */
 766 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
 767                                         struct dma_resv *resv,
 768                                         enum dma_resv_usage usage)
 769 {
 770         struct dma_resv_iter cursor;
 771         struct dma_fence *fence;
 772         int ret;
 773
 774         dma_resv_assert_held(resv);
 775
 776         dma_resv_for_each_fence(&cursor, resv, usage, fence) {
 777                 /* Make sure to grab an additional ref on the added fence */
 778                 dma_fence_get(fence);
 779                 ret = drm_sched_job_add_dependency(job, fence);
 780                 if (ret) {
 781                         dma_fence_put(fence);
 782                         return ret;
 783                 }
 784         }
 785         return 0;
 786 }
 787 EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
 788
 789 /**
 790  * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
 791  *   dependencies
 792  * @job: scheduler job to add the dependencies to
 793  * @obj: the gem object to add new dependencies from.
 794  * @write: whether the job might write the object (so we need to depend on
 795  * shared fences in the reservation object).
 796  *
 797  * This should be called after drm_gem_lock_reservations() on your array of
 798  * GEM objects used in the job but before updating the reservations with your
 799  * own fences.
 800  *
 801  * Returns:
 802  * 0 on success, or an error on failing to expand the array.
 803  */
 804 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
 805                                             struct drm_gem_object *obj,
 806                                             bool write)
 807 {
 808         return drm_sched_job_add_resv_dependencies(job, obj->resv,
 809                                                    dma_resv_usage_rw(write));
 810 }
 811 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
 812
 813 /**
 814  * drm_sched_job_cleanup - clean up scheduler job resources
 815  * @job: scheduler job to clean up
 816  *
 817  * Cleans up the resources allocated with drm_sched_job_init().
 818  *
 819  * Drivers should call this from their error unwind code if @job is aborted
 820  * before drm_sched_job_arm() is called.
 821  *
 822  * After that point of no return @job is committed to be executed by the
 823  * scheduler, and this function should be called from the
 824  * &drm_sched_backend_ops.free_job callback.
 825  */
 826 void drm_sched_job_cleanup(struct drm_sched_job *job)
 827 {
 828         struct dma_fence *fence;
 829         unsigned long index;
 830
 831         if (kref_read(&job->s_fence->finished.refcount)) {
 832                 /* drm_sched_job_arm() has been called */
 833                 dma_fence_put(&job->s_fence->finished);
 834         } else {
 835                 /* aborted job before committing to run it */
 836                 drm_sched_fence_free(job->s_fence);
 837         }
 838
 839         job->s_fence = NULL;
 840
 841         xa_for_each(&job->dependencies, index, fence) {
 842                 dma_fence_put(fence);
 843         }
 844         xa_destroy(&job->dependencies);
 845
 846 }
 847 EXPORT_SYMBOL(drm_sched_job_cleanup);
 848
 849 /**
 850  * drm_sched_ready - is the scheduler ready
 851  *
 852  * @sched: scheduler instance
 853  *
 854  * Return true if we can push more jobs to the hw, otherwise false.
 855  */
 856 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 857 {
 858         return atomic_read(&sched->hw_rq_count) <
 859                 sched->hw_submission_limit;
 860 }
 861
 862 /**
 863  * drm_sched_wakeup - Wake up the scheduler when it is ready
 864  *
 865  * @sched: scheduler instance
 866  *
 867  */
 868 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 869 {
 870         if (drm_sched_ready(sched))
 871                 wake_up_interruptible(&sched->wake_up_worker);
 872 }
 873
 874 /**
 875  * drm_sched_select_entity - Select next entity to process
 876  *
 877  * @sched: scheduler instance
 878  *
 879  * Returns the entity to process or NULL if none are found.
 880  */
 881 static struct drm_sched_entity *
 882 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 883 {
 884         struct drm_sched_entity *entity;
 885         int i;
 886
 887         if (!drm_sched_ready(sched))
 888                 return NULL;
 889
 890         /* Kernel run queue has higher priority than normal run queue*/
 891         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 892                 entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
 893                         drm_sched_rq_select_entity_fifo(&sched->sched_rq[i]) :
 894                         drm_sched_rq_select_entity_rr(&sched->sched_rq[i]);
 895                 if (entity)
 896                         break;
 897         }
 898
 899         return entity;
 900 }
 901
 902 /**
 903  * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
 904  *
 905  * @sched: scheduler instance
 906  *
 907  * Returns the next finished job from the pending list (if there is one)
 908  * ready for it to be destroyed.
 909  */
 910 static struct drm_sched_job *
 911 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
 912 {
 913         struct drm_sched_job *job, *next;
 914
 915         spin_lock(&sched->job_list_lock);
 916
 917         job = list_first_entry_or_null(&sched->pending_list,
 918                                        struct drm_sched_job, list);
 919
 920         if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
 921                 /* remove job from pending_list */
 922                 list_del_init(&job->list);
 923
 924                 /* cancel this job's TO timer */
 925                 cancel_delayed_work(&sched->work_tdr);
 926                 /* make the scheduled timestamp more accurate */
 927                 next = list_first_entry_or_null(&sched->pending_list,
 928                                                 typeof(*next), list);
 929
 930                 if (next) {
 931                         next->s_fence->scheduled.timestamp =
 932                                 job->s_fence->finished.timestamp;
 933                         /* start TO timer for next job */
 934                         drm_sched_start_timeout(sched);
 935                 }
 936         } else {
 937                 job = NULL;
 938         }
 939
 940         spin_unlock(&sched->job_list_lock);
 941
 942         if (job) {
 943                 job->entity->elapsed_ns += ktime_to_ns(
 944                         ktime_sub(job->s_fence->finished.timestamp,
 945                                   job->s_fence->scheduled.timestamp));
 946         }
 947
 948         return job;
 949 }
 950
 951 /**
 952  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
 953  * @sched_list: list of drm_gpu_schedulers
 954  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
 955  *
 956  * Returns pointer of the sched with the least load or NULL if none of the
 957  * drm_gpu_schedulers are ready
 958  */
 959 struct drm_gpu_scheduler *
 960 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
 961                      unsigned int num_sched_list)
 962 {
 963         struct drm_gpu_scheduler *sched, *picked_sched = NULL;
 964         int i;
 965         unsigned int min_score = UINT_MAX, num_score;
 966
 967         for (i = 0; i < num_sched_list; ++i) {
 968                 sched = sched_list[i];
 969
 970                 if (!sched->ready) {
 971                         DRM_WARN("scheduler %s is not ready, skipping",
 972                                  sched->name);
 973                         continue;
 974                 }
 975
 976                 num_score = atomic_read(sched->score);
 977                 if (num_score < min_score) {
 978                         min_score = num_score;
 979                         picked_sched = sched;
 980                 }
 981         }
 982
 983         return picked_sched;
 984 }
 985 EXPORT_SYMBOL(drm_sched_pick_best);
 986
 987 /**
 988  * drm_sched_blocked - check if the scheduler is blocked
 989  *
 990  * @sched: scheduler instance
 991  *
 992  * Returns true if blocked, otherwise false.
 993  */
 994 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 995 {
 996         if (kthread_should_park()) {
 997                 kthread_parkme();
 998                 return true;
 999         }
1000
1001         return false;
1002 }
1003
1004 /**
1005  * drm_sched_main - main scheduler thread
1006  *
1007  * @param: scheduler instance
1008  *
1009  * Returns 0.
1010  */
1011 static int drm_sched_main(void *param)
1012 {
1013         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
1014         int r;
1015
1016         sched_set_fifo_low(current);
1017
1018         while (!kthread_should_stop()) {
1019                 struct drm_sched_entity *entity = NULL;
1020                 struct drm_sched_fence *s_fence;
1021                 struct drm_sched_job *sched_job;
1022                 struct dma_fence *fence;
1023                 struct drm_sched_job *cleanup_job = NULL;
1024
1025                 wait_event_interruptible(sched->wake_up_worker,
1026                                          (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
1027                                          (!drm_sched_blocked(sched) &&
1028                                           (entity = drm_sched_select_entity(sched))) ||
1029                                          kthread_should_stop());
1030
1031                 if (cleanup_job)
1032                         sched->ops->free_job(cleanup_job);
1033
1034                 if (!entity)
1035                         continue;
1036
1037                 sched_job = drm_sched_entity_pop_job(entity);
1038
1039                 if (!sched_job) {
1040                         complete_all(&entity->entity_idle);
1041                         continue;
1042                 }
1043
1044                 s_fence = sched_job->s_fence;
1045
1046                 atomic_inc(&sched->hw_rq_count);
1047                 drm_sched_job_begin(sched_job);
1048
1049                 trace_drm_run_job(sched_job, entity);
1050                 fence = sched->ops->run_job(sched_job);
1051                 complete_all(&entity->entity_idle);
1052                 drm_sched_fence_scheduled(s_fence);
1053
1054                 if (!IS_ERR_OR_NULL(fence)) {
1055                         drm_sched_fence_set_parent(s_fence, fence);
1056                         /* Drop for original kref_init of the fence */
1057                         dma_fence_put(fence);
1058
1059                         r = dma_fence_add_callback(fence, &sched_job->cb,
1060                                                    drm_sched_job_done_cb);
1061                         if (r == -ENOENT)
1062                                 drm_sched_job_done(sched_job);
1063                         else if (r)
1064                                 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
1065                                           r);
1066                 } else {
1067                         if (IS_ERR(fence))
1068                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
1069
1070                         drm_sched_job_done(sched_job);
1071                 }
1072
1073                 wake_up(&sched->job_scheduled);
1074         }
1075         return 0;
1076 }
1077
1078 /**
1079  * drm_sched_init - Init a gpu scheduler instance
1080  *
1081  * @sched: scheduler instance
1082  * @ops: backend operations for this scheduler
1083  * @hw_submission: number of hw submissions that can be in flight
1084  * @hang_limit: number of times to allow a job to hang before dropping it
1085  * @timeout: timeout value in jiffies for the scheduler
1086  * @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
1087  *              used
1088  * @score: optional score atomic shared with other schedulers
1089  * @name: name used for debugging
1090  * @dev: target &struct device
1091  *
1092  * Return 0 on success, otherwise error code.
1093  */
1094 int drm_sched_init(struct drm_gpu_scheduler *sched,
1095                    const struct drm_sched_backend_ops *ops,
1096                    unsigned hw_submission, unsigned hang_limit,
1097                    long timeout, struct workqueue_struct *timeout_wq,
1098                    atomic_t *score, const char *name, struct device *dev)
1099 {
1100         int i, ret;
1101         sched->ops = ops;
1102         sched->hw_submission_limit = hw_submission;
1103         sched->name = name;
1104         sched->timeout = timeout;
1105         sched->timeout_wq = timeout_wq ? : system_wq;
1106         sched->hang_limit = hang_limit;
1107         sched->score = score ? score : &sched->_score;
1108         sched->dev = dev;
1109         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
1110                 drm_sched_rq_init(sched, &sched->sched_rq[i]);
1111
1112         init_waitqueue_head(&sched->wake_up_worker);
1113         init_waitqueue_head(&sched->job_scheduled);
1114         INIT_LIST_HEAD(&sched->pending_list);
1115         spin_lock_init(&sched->job_list_lock);
1116         atomic_set(&sched->hw_rq_count, 0);
1117         INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1118         atomic_set(&sched->_score, 0);
1119         atomic64_set(&sched->job_id_count, 0);
1120
1121         /* Each scheduler will run on a seperate kernel thread */
1122         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
1123         if (IS_ERR(sched->thread)) {
1124                 ret = PTR_ERR(sched->thread);
1125                 sched->thread = NULL;
1126                 DRM_DEV_ERROR(sched->dev, "Failed to create scheduler for %s.\n", name);
1127                 return ret;
1128         }
1129
1130         sched->ready = true;
1131         return 0;
1132 }
1133 EXPORT_SYMBOL(drm_sched_init);
1134
1135 /**
1136  * drm_sched_fini - Destroy a gpu scheduler
1137  *
1138  * @sched: scheduler instance
1139  *
1140  * Tears down and cleans up the scheduler.
1141  */
1142 void drm_sched_fini(struct drm_gpu_scheduler *sched)
1143 {
1144         struct drm_sched_entity *s_entity;
1145         int i;
1146
1147         if (sched->thread)
1148                 kthread_stop(sched->thread);
1149
1150         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
1151                 struct drm_sched_rq *rq = &sched->sched_rq[i];
1152
1153                 if (!rq)
1154                         continue;
1155
1156                 spin_lock(&rq->lock);
1157                 list_for_each_entry(s_entity, &rq->entities, list)
1158                         /*
1159                          * Prevents reinsertion and marks job_queue as idle,
1160                          * it will removed from rq in drm_sched_entity_fini
1161                          * eventually
1162                          */
1163                         s_entity->stopped = true;
1164                 spin_unlock(&rq->lock);
1165
1166         }
1167
1168         /* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1169         wake_up_all(&sched->job_scheduled);
1170
1171         /* Confirm no work left behind accessing device structures */
1172         cancel_delayed_work_sync(&sched->work_tdr);
1173
1174         sched->ready = false;
1175 }
1176 EXPORT_SYMBOL(drm_sched_fini);
1177
1178 /**
1179  * drm_sched_increase_karma - Update sched_entity guilty flag
1180  *
1181  * @bad: The job guilty of time out
1182  *
1183  * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1184  * limit of the scheduler then the respective sched entity is marked guilty and
1185  * jobs from it will not be scheduled further
1186  */
1187 void drm_sched_increase_karma(struct drm_sched_job *bad)
1188 {
1189         int i;
1190         struct drm_sched_entity *tmp;
1191         struct drm_sched_entity *entity;
1192         struct drm_gpu_scheduler *sched = bad->sched;
1193
1194         /* don't change @bad's karma if it's from KERNEL RQ,
1195          * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1196          * corrupt but keep in mind that kernel jobs always considered good.
1197          */
1198         if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1199                 atomic_inc(&bad->karma);
1200
1201                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
1202                      i++) {
1203                         struct drm_sched_rq *rq = &sched->sched_rq[i];
1204
1205                         spin_lock(&rq->lock);
1206                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1207                                 if (bad->s_fence->scheduled.context ==
1208                                     entity->fence_context) {
1209                                         if (entity->guilty)
1210                                                 atomic_set(entity->guilty, 1);
1211                                         break;
1212                                 }
1213                         }
1214                         spin_unlock(&rq->lock);
1215                         if (&entity->list != &rq->entities)
1216                                 break;
1217                 }
1218         }
1219 }
1220 EXPORT_SYMBOL(drm_sched_increase_karma);