Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
  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 */
 71static 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 */
 88void 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 */
107void 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 */
127static struct drm_sched_entity *
128drm_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 */
170static 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 */
191static 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 */
206bool 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}
222EXPORT_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 */
231static 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 */
245void drm_sched_fault(struct drm_gpu_scheduler *sched)
246{
247	mod_delayed_work(system_wq, &sched->work_tdr, 0);
248}
249EXPORT_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 */
263unsigned 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}
279EXPORT_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 */
289void 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}
301EXPORT_SYMBOL(drm_sched_resume_timeout);
302
303static 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
313static void drm_sched_job_timedout(struct work_struct *work)
314{
315	struct drm_gpu_scheduler *sched;
316	struct drm_sched_job *job;
317	enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
318
319	sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
320
321	/* Protects against concurrent deletion in drm_sched_get_cleanup_job */
322	spin_lock(&sched->job_list_lock);
323	job = list_first_entry_or_null(&sched->pending_list,
324				       struct drm_sched_job, list);
325
326	if (job) {
327		/*
328		 * Remove the bad job so it cannot be freed by concurrent
329		 * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
330		 * is parked at which point it's safe.
331		 */
332		list_del_init(&job->list);
333		spin_unlock(&sched->job_list_lock);
334
335		status = job->sched->ops->timedout_job(job);
336
337		/*
338		 * Guilty job did complete and hence needs to be manually removed
339		 * See drm_sched_stop doc.
340		 */
341		if (sched->free_guilty) {
342			job->sched->ops->free_job(job);
343			sched->free_guilty = false;
344		}
345	} else {
346		spin_unlock(&sched->job_list_lock);
347	}
348
349	if (status != DRM_GPU_SCHED_STAT_ENODEV) {
350		spin_lock(&sched->job_list_lock);
351		drm_sched_start_timeout(sched);
352		spin_unlock(&sched->job_list_lock);
353	}
354}
355
356 /**
357  * drm_sched_increase_karma - Update sched_entity guilty flag
358  *
359  * @bad: The job guilty of time out
360  *
361  * Increment on every hang caused by the 'bad' job. If this exceeds the hang
362  * limit of the scheduler then the respective sched entity is marked guilty and
363  * jobs from it will not be scheduled further
364  */
365void drm_sched_increase_karma(struct drm_sched_job *bad)
366{
367	drm_sched_increase_karma_ext(bad, 1);
368}
369EXPORT_SYMBOL(drm_sched_increase_karma);
370
371void drm_sched_reset_karma(struct drm_sched_job *bad)
372{
373	drm_sched_increase_karma_ext(bad, 0);
374}
375EXPORT_SYMBOL(drm_sched_reset_karma);
376
377/**
378 * drm_sched_stop - stop the scheduler
379 *
380 * @sched: scheduler instance
381 * @bad: job which caused the time out
382 *
383 * Stop the scheduler and also removes and frees all completed jobs.
384 * Note: bad job will not be freed as it might be used later and so it's
385 * callers responsibility to release it manually if it's not part of the
386 * pending list any more.
387 *
388 */
389void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
390{
391	struct drm_sched_job *s_job, *tmp;
392
393	kthread_park(sched->thread);
394
395	/*
396	 * Reinsert back the bad job here - now it's safe as
397	 * drm_sched_get_cleanup_job cannot race against us and release the
398	 * bad job at this point - we parked (waited for) any in progress
399	 * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
400	 * now until the scheduler thread is unparked.
401	 */
402	if (bad && bad->sched == sched)
403		/*
404		 * Add at the head of the queue to reflect it was the earliest
405		 * job extracted.
406		 */
407		list_add(&bad->list, &sched->pending_list);
408
409	/*
410	 * Iterate the job list from later to  earlier one and either deactive
411	 * their HW callbacks or remove them from pending list if they already
412	 * signaled.
413	 * This iteration is thread safe as sched thread is stopped.
414	 */
415	list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
416					 list) {
417		if (s_job->s_fence->parent &&
418		    dma_fence_remove_callback(s_job->s_fence->parent,
419					      &s_job->cb)) {
420			atomic_dec(&sched->hw_rq_count);
421		} else {
422			/*
423			 * remove job from pending_list.
424			 * Locking here is for concurrent resume timeout
425			 */
426			spin_lock(&sched->job_list_lock);
427			list_del_init(&s_job->list);
428			spin_unlock(&sched->job_list_lock);
429
430			/*
431			 * Wait for job's HW fence callback to finish using s_job
432			 * before releasing it.
433			 *
434			 * Job is still alive so fence refcount at least 1
435			 */
436			dma_fence_wait(&s_job->s_fence->finished, false);
437
438			/*
439			 * We must keep bad job alive for later use during
440			 * recovery by some of the drivers but leave a hint
441			 * that the guilty job must be released.
442			 */
443			if (bad != s_job)
444				sched->ops->free_job(s_job);
445			else
446				sched->free_guilty = true;
447		}
448	}
449
450	/*
451	 * Stop pending timer in flight as we rearm it in  drm_sched_start. This
452	 * avoids the pending timeout work in progress to fire right away after
453	 * this TDR finished and before the newly restarted jobs had a
454	 * chance to complete.
455	 */
456	cancel_delayed_work(&sched->work_tdr);
457}
458
459EXPORT_SYMBOL(drm_sched_stop);
460
461/**
462 * drm_sched_start - recover jobs after a reset
463 *
464 * @sched: scheduler instance
465 * @full_recovery: proceed with complete sched restart
466 *
467 */
468void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
469{
470	struct drm_sched_job *s_job, *tmp;
471	int r;
472
473	/*
474	 * Locking the list is not required here as the sched thread is parked
475	 * so no new jobs are being inserted or removed. Also concurrent
476	 * GPU recovers can't run in parallel.
477	 */
478	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
479		struct dma_fence *fence = s_job->s_fence->parent;
480
481		atomic_inc(&sched->hw_rq_count);
482
483		if (!full_recovery)
484			continue;
485
486		if (fence) {
487			r = dma_fence_add_callback(fence, &s_job->cb,
488						   drm_sched_job_done_cb);
489			if (r == -ENOENT)
490				drm_sched_job_done(s_job);
491			else if (r)
492				DRM_ERROR("fence add callback failed (%d)\n",
493					  r);
494		} else
495			drm_sched_job_done(s_job);
496	}
497
498	if (full_recovery) {
499		spin_lock(&sched->job_list_lock);
500		drm_sched_start_timeout(sched);
501		spin_unlock(&sched->job_list_lock);
502	}
503
504	kthread_unpark(sched->thread);
505}
506EXPORT_SYMBOL(drm_sched_start);
507
508/**
509 * drm_sched_resubmit_jobs - helper to relaunch jobs from the pending list
510 *
511 * @sched: scheduler instance
512 *
513 */
514void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
515{
516	drm_sched_resubmit_jobs_ext(sched, INT_MAX);
517}
518EXPORT_SYMBOL(drm_sched_resubmit_jobs);
519
520/**
521 * drm_sched_resubmit_jobs_ext - helper to relunch certain number of jobs from mirror ring list
522 *
523 * @sched: scheduler instance
524 * @max: job numbers to relaunch
525 *
526 */
527void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)
528{
529	struct drm_sched_job *s_job, *tmp;
530	uint64_t guilty_context;
531	bool found_guilty = false;
532	struct dma_fence *fence;
533	int i = 0;
534
535	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
536		struct drm_sched_fence *s_fence = s_job->s_fence;
537
538		if (i >= max)
539			break;
540
541		if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
542			found_guilty = true;
543			guilty_context = s_job->s_fence->scheduled.context;
544		}
545
546		if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
547			dma_fence_set_error(&s_fence->finished, -ECANCELED);
548
549		dma_fence_put(s_job->s_fence->parent);
550		fence = sched->ops->run_job(s_job);
551		i++;
552
553		if (IS_ERR_OR_NULL(fence)) {
554			if (IS_ERR(fence))
555				dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
556
557			s_job->s_fence->parent = NULL;
558		} else {
559			s_job->s_fence->parent = fence;
560		}
561	}
562}
563EXPORT_SYMBOL(drm_sched_resubmit_jobs_ext);
564
565/**
566 * drm_sched_job_init - init a scheduler job
567 *
568 * @job: scheduler job to init
569 * @entity: scheduler entity to use
570 * @owner: job owner for debugging
571 *
572 * Refer to drm_sched_entity_push_job() documentation
573 * for locking considerations.
574 *
575 * Returns 0 for success, negative error code otherwise.
576 */
577int drm_sched_job_init(struct drm_sched_job *job,
578		       struct drm_sched_entity *entity,
579		       void *owner)
580{
581	struct drm_gpu_scheduler *sched;
582
583	drm_sched_entity_select_rq(entity);
584	if (!entity->rq)
585		return -ENOENT;
586
587	sched = entity->rq->sched;
588
589	job->sched = sched;
590	job->entity = entity;
591	job->s_priority = entity->rq - sched->sched_rq;
592	job->s_fence = drm_sched_fence_create(entity, owner);
593	if (!job->s_fence)
594		return -ENOMEM;
595	job->id = atomic64_inc_return(&sched->job_id_count);
596
597	INIT_LIST_HEAD(&job->list);
598
599	return 0;
600}
601EXPORT_SYMBOL(drm_sched_job_init);
602
603/**
604 * drm_sched_job_cleanup - clean up scheduler job resources
605 *
606 * @job: scheduler job to clean up
607 */
608void drm_sched_job_cleanup(struct drm_sched_job *job)
609{
610	dma_fence_put(&job->s_fence->finished);
611	job->s_fence = NULL;
612}
613EXPORT_SYMBOL(drm_sched_job_cleanup);
614
615/**
616 * drm_sched_ready - is the scheduler ready
617 *
618 * @sched: scheduler instance
619 *
620 * Return true if we can push more jobs to the hw, otherwise false.
621 */
622static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
623{
624	return atomic_read(&sched->hw_rq_count) <
625		sched->hw_submission_limit;
626}
627
628/**
629 * drm_sched_wakeup - Wake up the scheduler when it is ready
630 *
631 * @sched: scheduler instance
632 *
633 */
634void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
635{
636	if (drm_sched_ready(sched))
637		wake_up_interruptible(&sched->wake_up_worker);
638}
639
640/**
641 * drm_sched_select_entity - Select next entity to process
642 *
643 * @sched: scheduler instance
644 *
645 * Returns the entity to process or NULL if none are found.
646 */
647static struct drm_sched_entity *
648drm_sched_select_entity(struct drm_gpu_scheduler *sched)
649{
650	struct drm_sched_entity *entity;
651	int i;
652
653	if (!drm_sched_ready(sched))
654		return NULL;
655
656	/* Kernel run queue has higher priority than normal run queue*/
657	for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
658		entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
659		if (entity)
660			break;
661	}
662
663	return entity;
664}
665
666/**
667 * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
668 *
669 * @sched: scheduler instance
670 *
671 * Returns the next finished job from the pending list (if there is one)
672 * ready for it to be destroyed.
673 */
674static struct drm_sched_job *
675drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
676{
677	struct drm_sched_job *job, *next;
678
679	/*
680	 * Don't destroy jobs while the timeout worker is running  OR thread
681	 * is being parked and hence assumed to not touch pending_list
682	 */
683	if ((sched->timeout != MAX_SCHEDULE_TIMEOUT &&
684	    !cancel_delayed_work(&sched->work_tdr)) ||
685	    kthread_should_park())
686		return NULL;
687
688	spin_lock(&sched->job_list_lock);
689
690	job = list_first_entry_or_null(&sched->pending_list,
691				       struct drm_sched_job, list);
692
693	if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
694		/* remove job from pending_list */
695		list_del_init(&job->list);
696		/* make the scheduled timestamp more accurate */
697		next = list_first_entry_or_null(&sched->pending_list,
698						typeof(*next), list);
699		if (next)
700			next->s_fence->scheduled.timestamp =
701				job->s_fence->finished.timestamp;
702
703	} else {
704		job = NULL;
705		/* queue timeout for next job */
706		drm_sched_start_timeout(sched);
707	}
708
709	spin_unlock(&sched->job_list_lock);
710
711	return job;
712}
713
714/**
715 * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
716 * @sched_list: list of drm_gpu_schedulers
717 * @num_sched_list: number of drm_gpu_schedulers in the sched_list
718 *
719 * Returns pointer of the sched with the least load or NULL if none of the
720 * drm_gpu_schedulers are ready
721 */
722struct drm_gpu_scheduler *
723drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
724		     unsigned int num_sched_list)
725{
726	struct drm_gpu_scheduler *sched, *picked_sched = NULL;
727	int i;
728	unsigned int min_score = UINT_MAX, num_score;
729
730	for (i = 0; i < num_sched_list; ++i) {
731		sched = sched_list[i];
732
733		if (!sched->ready) {
734			DRM_WARN("scheduler %s is not ready, skipping",
735				 sched->name);
736			continue;
737		}
738
739		num_score = atomic_read(sched->score);
740		if (num_score < min_score) {
741			min_score = num_score;
742			picked_sched = sched;
743		}
744	}
745
746	return picked_sched;
747}
748EXPORT_SYMBOL(drm_sched_pick_best);
749
750/**
751 * drm_sched_blocked - check if the scheduler is blocked
752 *
753 * @sched: scheduler instance
754 *
755 * Returns true if blocked, otherwise false.
756 */
757static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
758{
759	if (kthread_should_park()) {
760		kthread_parkme();
761		return true;
762	}
763
764	return false;
765}
766
767/**
768 * drm_sched_main - main scheduler thread
769 *
770 * @param: scheduler instance
771 *
772 * Returns 0.
773 */
774static int drm_sched_main(void *param)
775{
776	struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
777	int r;
778
779	sched_set_fifo_low(current);
780
781	while (!kthread_should_stop()) {
782		struct drm_sched_entity *entity = NULL;
783		struct drm_sched_fence *s_fence;
784		struct drm_sched_job *sched_job;
785		struct dma_fence *fence;
786		struct drm_sched_job *cleanup_job = NULL;
787
788		wait_event_interruptible(sched->wake_up_worker,
789					 (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
790					 (!drm_sched_blocked(sched) &&
791					  (entity = drm_sched_select_entity(sched))) ||
792					 kthread_should_stop());
793
794		if (cleanup_job) {
795			sched->ops->free_job(cleanup_job);
796			/* queue timeout for next job */
797			drm_sched_start_timeout(sched);
798		}
799
800		if (!entity)
801			continue;
802
803		sched_job = drm_sched_entity_pop_job(entity);
804
805		complete(&entity->entity_idle);
806
807		if (!sched_job)
808			continue;
809
810		s_fence = sched_job->s_fence;
811
812		atomic_inc(&sched->hw_rq_count);
813		drm_sched_job_begin(sched_job);
814
815		trace_drm_run_job(sched_job, entity);
816		fence = sched->ops->run_job(sched_job);
817		drm_sched_fence_scheduled(s_fence);
818
819		if (!IS_ERR_OR_NULL(fence)) {
820			s_fence->parent = dma_fence_get(fence);
821			r = dma_fence_add_callback(fence, &sched_job->cb,
822						   drm_sched_job_done_cb);
823			if (r == -ENOENT)
824				drm_sched_job_done(sched_job);
825			else if (r)
826				DRM_ERROR("fence add callback failed (%d)\n",
827					  r);
828			dma_fence_put(fence);
829		} else {
830			if (IS_ERR(fence))
831				dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
832
833			drm_sched_job_done(sched_job);
834		}
835
836		wake_up(&sched->job_scheduled);
837	}
838	return 0;
839}
840
841/**
842 * drm_sched_init - Init a gpu scheduler instance
843 *
844 * @sched: scheduler instance
845 * @ops: backend operations for this scheduler
846 * @hw_submission: number of hw submissions that can be in flight
847 * @hang_limit: number of times to allow a job to hang before dropping it
848 * @timeout: timeout value in jiffies for the scheduler
849 * @score: optional score atomic shared with other schedulers
850 * @name: name used for debugging
851 *
852 * Return 0 on success, otherwise error code.
853 */
854int drm_sched_init(struct drm_gpu_scheduler *sched,
855		   const struct drm_sched_backend_ops *ops,
856		   unsigned hw_submission, unsigned hang_limit, long timeout,
857		   atomic_t *score, const char *name)
858{
859	int i, ret;
860	sched->ops = ops;
861	sched->hw_submission_limit = hw_submission;
862	sched->name = name;
863	sched->timeout = timeout;
864	sched->hang_limit = hang_limit;
865	sched->score = score ? score : &sched->_score;
866	for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
867		drm_sched_rq_init(sched, &sched->sched_rq[i]);
868
869	init_waitqueue_head(&sched->wake_up_worker);
870	init_waitqueue_head(&sched->job_scheduled);
871	INIT_LIST_HEAD(&sched->pending_list);
872	spin_lock_init(&sched->job_list_lock);
873	atomic_set(&sched->hw_rq_count, 0);
874	INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
875	atomic_set(&sched->_score, 0);
876	atomic64_set(&sched->job_id_count, 0);
877
878	/* Each scheduler will run on a seperate kernel thread */
879	sched->thread = kthread_run(drm_sched_main, sched, sched->name);
880	if (IS_ERR(sched->thread)) {
881		ret = PTR_ERR(sched->thread);
882		sched->thread = NULL;
883		DRM_ERROR("Failed to create scheduler for %s.\n", name);
884		return ret;
885	}
886
887	sched->ready = true;
888	return 0;
889}
890EXPORT_SYMBOL(drm_sched_init);
891
892/**
893 * drm_sched_fini - Destroy a gpu scheduler
894 *
895 * @sched: scheduler instance
896 *
897 * Tears down and cleans up the scheduler.
898 */
899void drm_sched_fini(struct drm_gpu_scheduler *sched)
900{
901	struct drm_sched_entity *s_entity;
902	int i;
903
904	if (sched->thread)
905		kthread_stop(sched->thread);
906
907	for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
908		struct drm_sched_rq *rq = &sched->sched_rq[i];
909
910		if (!rq)
911			continue;
912
913		spin_lock(&rq->lock);
914		list_for_each_entry(s_entity, &rq->entities, list)
915			/*
916			 * Prevents reinsertion and marks job_queue as idle,
917			 * it will removed from rq in drm_sched_entity_fini
918			 * eventually
919			 */
920			s_entity->stopped = true;
921		spin_unlock(&rq->lock);
922
923	}
924
925	/* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
926	wake_up_all(&sched->job_scheduled);
927
928	/* Confirm no work left behind accessing device structures */
929	cancel_delayed_work_sync(&sched->work_tdr);
930
931	sched->ready = false;
932}
933EXPORT_SYMBOL(drm_sched_fini);
934
935/**
936 * drm_sched_increase_karma_ext - Update sched_entity guilty flag
937 *
938 * @bad: The job guilty of time out
939 * @type: type for increase/reset karma
940 *
941 */
942void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type)
943{
944	int i;
945	struct drm_sched_entity *tmp;
946	struct drm_sched_entity *entity;
947	struct drm_gpu_scheduler *sched = bad->sched;
948
949	/* don't change @bad's karma if it's from KERNEL RQ,
950	 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
951	 * corrupt but keep in mind that kernel jobs always considered good.
952	 */
953	if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
954		if (type == 0)
955			atomic_set(&bad->karma, 0);
956		else if (type == 1)
957			atomic_inc(&bad->karma);
958
959		for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
960		     i++) {
961			struct drm_sched_rq *rq = &sched->sched_rq[i];
962
963			spin_lock(&rq->lock);
964			list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
965				if (bad->s_fence->scheduled.context ==
966				    entity->fence_context) {
967					if (entity->guilty)
968						atomic_set(entity->guilty, type);
969					break;
970				}
971			}
972			spin_unlock(&rq->lock);
973			if (&entity->list != &rq->entities)
974				break;
975		}
976	}
977}
978EXPORT_SYMBOL(drm_sched_increase_karma_ext);