xref: /linux/drivers/gpu/drm/scheduler/sched_main.c (revision 24168c5e6dfbdd5b414f048f47f75d64533296ca)
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 /**
52  * DOC: Flow Control
53  *
54  * The DRM GPU scheduler provides a flow control mechanism to regulate the rate
55  * in which the jobs fetched from scheduler entities are executed.
56  *
57  * In this context the &drm_gpu_scheduler keeps track of a driver specified
58  * credit limit representing the capacity of this scheduler and a credit count;
59  * every &drm_sched_job carries a driver specified number of credits.
60  *
61  * Once a job is executed (but not yet finished), the job's credits contribute
62  * to the scheduler's credit count until the job is finished. If by executing
63  * one more job the scheduler's credit count would exceed the scheduler's
64  * credit limit, the job won't be executed. Instead, the scheduler will wait
65  * until the credit count has decreased enough to not overflow its credit limit.
66  * This implies waiting for previously executed jobs.
67  *
68  * Optionally, drivers may register a callback (update_job_credits) provided by
69  * struct drm_sched_backend_ops to update the job's credits dynamically. The
70  * scheduler executes this callback every time the scheduler considers a job for
71  * execution and subsequently checks whether the job fits the scheduler's credit
72  * limit.
73  */
74 
75 #include <linux/wait.h>
76 #include <linux/sched.h>
77 #include <linux/completion.h>
78 #include <linux/dma-resv.h>
79 #include <uapi/linux/sched/types.h>
80 
81 #include <drm/drm_print.h>
82 #include <drm/drm_gem.h>
83 #include <drm/drm_syncobj.h>
84 #include <drm/gpu_scheduler.h>
85 #include <drm/spsc_queue.h>
86 
87 #define CREATE_TRACE_POINTS
88 #include "gpu_scheduler_trace.h"
89 
90 #define to_drm_sched_job(sched_job)		\
91 		container_of((sched_job), struct drm_sched_job, queue_node)
92 
93 int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
94 
95 /**
96  * DOC: sched_policy (int)
97  * Used to override default entities scheduling policy in a run queue.
98  */
99 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).");
100 module_param_named(sched_policy, drm_sched_policy, int, 0444);
101 
102 static u32 drm_sched_available_credits(struct drm_gpu_scheduler *sched)
103 {
104 	u32 credits;
105 
106 	drm_WARN_ON(sched, check_sub_overflow(sched->credit_limit,
107 					      atomic_read(&sched->credit_count),
108 					      &credits));
109 
110 	return credits;
111 }
112 
113 /**
114  * drm_sched_can_queue -- Can we queue more to the hardware?
115  * @sched: scheduler instance
116  * @entity: the scheduler entity
117  *
118  * Return true if we can push at least one more job from @entity, false
119  * otherwise.
120  */
121 static bool drm_sched_can_queue(struct drm_gpu_scheduler *sched,
122 				struct drm_sched_entity *entity)
123 {
124 	struct drm_sched_job *s_job;
125 
126 	s_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
127 	if (!s_job)
128 		return false;
129 
130 	if (sched->ops->update_job_credits) {
131 		s_job->credits = sched->ops->update_job_credits(s_job);
132 
133 		drm_WARN(sched, !s_job->credits,
134 			 "Jobs with zero credits bypass job-flow control.\n");
135 	}
136 
137 	/* If a job exceeds the credit limit, truncate it to the credit limit
138 	 * itself to guarantee forward progress.
139 	 */
140 	if (drm_WARN(sched, s_job->credits > sched->credit_limit,
141 		     "Jobs may not exceed the credit limit, truncate.\n"))
142 		s_job->credits = sched->credit_limit;
143 
144 	return drm_sched_available_credits(sched) >= s_job->credits;
145 }
146 
147 static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
148 							    const struct rb_node *b)
149 {
150 	struct drm_sched_entity *ent_a =  rb_entry((a), struct drm_sched_entity, rb_tree_node);
151 	struct drm_sched_entity *ent_b =  rb_entry((b), struct drm_sched_entity, rb_tree_node);
152 
153 	return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
154 }
155 
156 static inline void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity)
157 {
158 	struct drm_sched_rq *rq = entity->rq;
159 
160 	if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
161 		rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
162 		RB_CLEAR_NODE(&entity->rb_tree_node);
163 	}
164 }
165 
166 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts)
167 {
168 	/*
169 	 * Both locks need to be grabbed, one to protect from entity->rq change
170 	 * for entity from within concurrent drm_sched_entity_select_rq and the
171 	 * other to update the rb tree structure.
172 	 */
173 	spin_lock(&entity->rq_lock);
174 	spin_lock(&entity->rq->lock);
175 
176 	drm_sched_rq_remove_fifo_locked(entity);
177 
178 	entity->oldest_job_waiting = ts;
179 
180 	rb_add_cached(&entity->rb_tree_node, &entity->rq->rb_tree_root,
181 		      drm_sched_entity_compare_before);
182 
183 	spin_unlock(&entity->rq->lock);
184 	spin_unlock(&entity->rq_lock);
185 }
186 
187 /**
188  * drm_sched_rq_init - initialize a given run queue struct
189  *
190  * @sched: scheduler instance to associate with this run queue
191  * @rq: scheduler run queue
192  *
193  * Initializes a scheduler runqueue.
194  */
195 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
196 			      struct drm_sched_rq *rq)
197 {
198 	spin_lock_init(&rq->lock);
199 	INIT_LIST_HEAD(&rq->entities);
200 	rq->rb_tree_root = RB_ROOT_CACHED;
201 	rq->current_entity = NULL;
202 	rq->sched = sched;
203 }
204 
205 /**
206  * drm_sched_rq_add_entity - add an entity
207  *
208  * @rq: scheduler run queue
209  * @entity: scheduler entity
210  *
211  * Adds a scheduler entity to the run queue.
212  */
213 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
214 			     struct drm_sched_entity *entity)
215 {
216 	if (!list_empty(&entity->list))
217 		return;
218 
219 	spin_lock(&rq->lock);
220 
221 	atomic_inc(rq->sched->score);
222 	list_add_tail(&entity->list, &rq->entities);
223 
224 	spin_unlock(&rq->lock);
225 }
226 
227 /**
228  * drm_sched_rq_remove_entity - remove an entity
229  *
230  * @rq: scheduler run queue
231  * @entity: scheduler entity
232  *
233  * Removes a scheduler entity from the run queue.
234  */
235 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
236 				struct drm_sched_entity *entity)
237 {
238 	if (list_empty(&entity->list))
239 		return;
240 
241 	spin_lock(&rq->lock);
242 
243 	atomic_dec(rq->sched->score);
244 	list_del_init(&entity->list);
245 
246 	if (rq->current_entity == entity)
247 		rq->current_entity = NULL;
248 
249 	if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
250 		drm_sched_rq_remove_fifo_locked(entity);
251 
252 	spin_unlock(&rq->lock);
253 }
254 
255 /**
256  * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
257  *
258  * @sched: the gpu scheduler
259  * @rq: scheduler run queue to check.
260  *
261  * Try to find the next ready entity.
262  *
263  * Return an entity if one is found; return an error-pointer (!NULL) if an
264  * entity was ready, but the scheduler had insufficient credits to accommodate
265  * its job; return NULL, if no ready entity was found.
266  */
267 static struct drm_sched_entity *
268 drm_sched_rq_select_entity_rr(struct drm_gpu_scheduler *sched,
269 			      struct drm_sched_rq *rq)
270 {
271 	struct drm_sched_entity *entity;
272 
273 	spin_lock(&rq->lock);
274 
275 	entity = rq->current_entity;
276 	if (entity) {
277 		list_for_each_entry_continue(entity, &rq->entities, list) {
278 			if (drm_sched_entity_is_ready(entity)) {
279 				/* If we can't queue yet, preserve the current
280 				 * entity in terms of fairness.
281 				 */
282 				if (!drm_sched_can_queue(sched, entity)) {
283 					spin_unlock(&rq->lock);
284 					return ERR_PTR(-ENOSPC);
285 				}
286 
287 				rq->current_entity = entity;
288 				reinit_completion(&entity->entity_idle);
289 				spin_unlock(&rq->lock);
290 				return entity;
291 			}
292 		}
293 	}
294 
295 	list_for_each_entry(entity, &rq->entities, list) {
296 		if (drm_sched_entity_is_ready(entity)) {
297 			/* If we can't queue yet, preserve the current entity in
298 			 * terms of fairness.
299 			 */
300 			if (!drm_sched_can_queue(sched, entity)) {
301 				spin_unlock(&rq->lock);
302 				return ERR_PTR(-ENOSPC);
303 			}
304 
305 			rq->current_entity = entity;
306 			reinit_completion(&entity->entity_idle);
307 			spin_unlock(&rq->lock);
308 			return entity;
309 		}
310 
311 		if (entity == rq->current_entity)
312 			break;
313 	}
314 
315 	spin_unlock(&rq->lock);
316 
317 	return NULL;
318 }
319 
320 /**
321  * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
322  *
323  * @sched: the gpu scheduler
324  * @rq: scheduler run queue to check.
325  *
326  * Find oldest waiting ready entity.
327  *
328  * Return an entity if one is found; return an error-pointer (!NULL) if an
329  * entity was ready, but the scheduler had insufficient credits to accommodate
330  * its job; return NULL, if no ready entity was found.
331  */
332 static struct drm_sched_entity *
333 drm_sched_rq_select_entity_fifo(struct drm_gpu_scheduler *sched,
334 				struct drm_sched_rq *rq)
335 {
336 	struct rb_node *rb;
337 
338 	spin_lock(&rq->lock);
339 	for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
340 		struct drm_sched_entity *entity;
341 
342 		entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
343 		if (drm_sched_entity_is_ready(entity)) {
344 			/* If we can't queue yet, preserve the current entity in
345 			 * terms of fairness.
346 			 */
347 			if (!drm_sched_can_queue(sched, entity)) {
348 				spin_unlock(&rq->lock);
349 				return ERR_PTR(-ENOSPC);
350 			}
351 
352 			rq->current_entity = entity;
353 			reinit_completion(&entity->entity_idle);
354 			break;
355 		}
356 	}
357 	spin_unlock(&rq->lock);
358 
359 	return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
360 }
361 
362 /**
363  * drm_sched_run_job_queue - enqueue run-job work
364  * @sched: scheduler instance
365  */
366 static void drm_sched_run_job_queue(struct drm_gpu_scheduler *sched)
367 {
368 	if (!READ_ONCE(sched->pause_submit))
369 		queue_work(sched->submit_wq, &sched->work_run_job);
370 }
371 
372 /**
373  * __drm_sched_run_free_queue - enqueue free-job work
374  * @sched: scheduler instance
375  */
376 static void __drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
377 {
378 	if (!READ_ONCE(sched->pause_submit))
379 		queue_work(sched->submit_wq, &sched->work_free_job);
380 }
381 
382 /**
383  * drm_sched_run_free_queue - enqueue free-job work if ready
384  * @sched: scheduler instance
385  */
386 static void drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
387 {
388 	struct drm_sched_job *job;
389 
390 	spin_lock(&sched->job_list_lock);
391 	job = list_first_entry_or_null(&sched->pending_list,
392 				       struct drm_sched_job, list);
393 	if (job && dma_fence_is_signaled(&job->s_fence->finished))
394 		__drm_sched_run_free_queue(sched);
395 	spin_unlock(&sched->job_list_lock);
396 }
397 
398 /**
399  * drm_sched_job_done - complete a job
400  * @s_job: pointer to the job which is done
401  *
402  * Finish the job's fence and wake up the worker thread.
403  */
404 static void drm_sched_job_done(struct drm_sched_job *s_job, int result)
405 {
406 	struct drm_sched_fence *s_fence = s_job->s_fence;
407 	struct drm_gpu_scheduler *sched = s_fence->sched;
408 
409 	atomic_sub(s_job->credits, &sched->credit_count);
410 	atomic_dec(sched->score);
411 
412 	trace_drm_sched_process_job(s_fence);
413 
414 	dma_fence_get(&s_fence->finished);
415 	drm_sched_fence_finished(s_fence, result);
416 	dma_fence_put(&s_fence->finished);
417 	__drm_sched_run_free_queue(sched);
418 }
419 
420 /**
421  * drm_sched_job_done_cb - the callback for a done job
422  * @f: fence
423  * @cb: fence callbacks
424  */
425 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
426 {
427 	struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
428 
429 	drm_sched_job_done(s_job, f->error);
430 }
431 
432 /**
433  * drm_sched_start_timeout - start timeout for reset worker
434  *
435  * @sched: scheduler instance to start the worker for
436  *
437  * Start the timeout for the given scheduler.
438  */
439 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
440 {
441 	lockdep_assert_held(&sched->job_list_lock);
442 
443 	if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
444 	    !list_empty(&sched->pending_list))
445 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
446 }
447 
448 static void drm_sched_start_timeout_unlocked(struct drm_gpu_scheduler *sched)
449 {
450 	spin_lock(&sched->job_list_lock);
451 	drm_sched_start_timeout(sched);
452 	spin_unlock(&sched->job_list_lock);
453 }
454 
455 /**
456  * drm_sched_tdr_queue_imm: - immediately start job timeout handler
457  *
458  * @sched: scheduler for which the timeout handling should be started.
459  *
460  * Start timeout handling immediately for the named scheduler.
461  */
462 void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched)
463 {
464 	spin_lock(&sched->job_list_lock);
465 	sched->timeout = 0;
466 	drm_sched_start_timeout(sched);
467 	spin_unlock(&sched->job_list_lock);
468 }
469 EXPORT_SYMBOL(drm_sched_tdr_queue_imm);
470 
471 /**
472  * drm_sched_fault - immediately start timeout handler
473  *
474  * @sched: scheduler where the timeout handling should be started.
475  *
476  * Start timeout handling immediately when the driver detects a hardware fault.
477  */
478 void drm_sched_fault(struct drm_gpu_scheduler *sched)
479 {
480 	if (sched->timeout_wq)
481 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
482 }
483 EXPORT_SYMBOL(drm_sched_fault);
484 
485 /**
486  * drm_sched_suspend_timeout - Suspend scheduler job timeout
487  *
488  * @sched: scheduler instance for which to suspend the timeout
489  *
490  * Suspend the delayed work timeout for the scheduler. This is done by
491  * modifying the delayed work timeout to an arbitrary large value,
492  * MAX_SCHEDULE_TIMEOUT in this case.
493  *
494  * Returns the timeout remaining
495  *
496  */
497 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
498 {
499 	unsigned long sched_timeout, now = jiffies;
500 
501 	sched_timeout = sched->work_tdr.timer.expires;
502 
503 	/*
504 	 * Modify the timeout to an arbitrarily large value. This also prevents
505 	 * the timeout to be restarted when new submissions arrive
506 	 */
507 	if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
508 			&& time_after(sched_timeout, now))
509 		return sched_timeout - now;
510 	else
511 		return sched->timeout;
512 }
513 EXPORT_SYMBOL(drm_sched_suspend_timeout);
514 
515 /**
516  * drm_sched_resume_timeout - Resume scheduler job timeout
517  *
518  * @sched: scheduler instance for which to resume the timeout
519  * @remaining: remaining timeout
520  *
521  * Resume the delayed work timeout for the scheduler.
522  */
523 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
524 		unsigned long remaining)
525 {
526 	spin_lock(&sched->job_list_lock);
527 
528 	if (list_empty(&sched->pending_list))
529 		cancel_delayed_work(&sched->work_tdr);
530 	else
531 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
532 
533 	spin_unlock(&sched->job_list_lock);
534 }
535 EXPORT_SYMBOL(drm_sched_resume_timeout);
536 
537 static void drm_sched_job_begin(struct drm_sched_job *s_job)
538 {
539 	struct drm_gpu_scheduler *sched = s_job->sched;
540 
541 	spin_lock(&sched->job_list_lock);
542 	list_add_tail(&s_job->list, &sched->pending_list);
543 	drm_sched_start_timeout(sched);
544 	spin_unlock(&sched->job_list_lock);
545 }
546 
547 static void drm_sched_job_timedout(struct work_struct *work)
548 {
549 	struct drm_gpu_scheduler *sched;
550 	struct drm_sched_job *job;
551 	enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
552 
553 	sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
554 
555 	/* Protects against concurrent deletion in drm_sched_get_finished_job */
556 	spin_lock(&sched->job_list_lock);
557 	job = list_first_entry_or_null(&sched->pending_list,
558 				       struct drm_sched_job, list);
559 
560 	if (job) {
561 		/*
562 		 * Remove the bad job so it cannot be freed by concurrent
563 		 * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
564 		 * is parked at which point it's safe.
565 		 */
566 		list_del_init(&job->list);
567 		spin_unlock(&sched->job_list_lock);
568 
569 		status = job->sched->ops->timedout_job(job);
570 
571 		/*
572 		 * Guilty job did complete and hence needs to be manually removed
573 		 * See drm_sched_stop doc.
574 		 */
575 		if (sched->free_guilty) {
576 			job->sched->ops->free_job(job);
577 			sched->free_guilty = false;
578 		}
579 	} else {
580 		spin_unlock(&sched->job_list_lock);
581 	}
582 
583 	if (status != DRM_GPU_SCHED_STAT_ENODEV)
584 		drm_sched_start_timeout_unlocked(sched);
585 }
586 
587 /**
588  * drm_sched_stop - stop the scheduler
589  *
590  * @sched: scheduler instance
591  * @bad: job which caused the time out
592  *
593  * Stop the scheduler and also removes and frees all completed jobs.
594  * Note: bad job will not be freed as it might be used later and so it's
595  * callers responsibility to release it manually if it's not part of the
596  * pending list any more.
597  *
598  */
599 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
600 {
601 	struct drm_sched_job *s_job, *tmp;
602 
603 	drm_sched_wqueue_stop(sched);
604 
605 	/*
606 	 * Reinsert back the bad job here - now it's safe as
607 	 * drm_sched_get_finished_job cannot race against us and release the
608 	 * bad job at this point - we parked (waited for) any in progress
609 	 * (earlier) cleanups and drm_sched_get_finished_job will not be called
610 	 * now until the scheduler thread is unparked.
611 	 */
612 	if (bad && bad->sched == sched)
613 		/*
614 		 * Add at the head of the queue to reflect it was the earliest
615 		 * job extracted.
616 		 */
617 		list_add(&bad->list, &sched->pending_list);
618 
619 	/*
620 	 * Iterate the job list from later to  earlier one and either deactive
621 	 * their HW callbacks or remove them from pending list if they already
622 	 * signaled.
623 	 * This iteration is thread safe as sched thread is stopped.
624 	 */
625 	list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
626 					 list) {
627 		if (s_job->s_fence->parent &&
628 		    dma_fence_remove_callback(s_job->s_fence->parent,
629 					      &s_job->cb)) {
630 			dma_fence_put(s_job->s_fence->parent);
631 			s_job->s_fence->parent = NULL;
632 			atomic_sub(s_job->credits, &sched->credit_count);
633 		} else {
634 			/*
635 			 * remove job from pending_list.
636 			 * Locking here is for concurrent resume timeout
637 			 */
638 			spin_lock(&sched->job_list_lock);
639 			list_del_init(&s_job->list);
640 			spin_unlock(&sched->job_list_lock);
641 
642 			/*
643 			 * Wait for job's HW fence callback to finish using s_job
644 			 * before releasing it.
645 			 *
646 			 * Job is still alive so fence refcount at least 1
647 			 */
648 			dma_fence_wait(&s_job->s_fence->finished, false);
649 
650 			/*
651 			 * We must keep bad job alive for later use during
652 			 * recovery by some of the drivers but leave a hint
653 			 * that the guilty job must be released.
654 			 */
655 			if (bad != s_job)
656 				sched->ops->free_job(s_job);
657 			else
658 				sched->free_guilty = true;
659 		}
660 	}
661 
662 	/*
663 	 * Stop pending timer in flight as we rearm it in  drm_sched_start. This
664 	 * avoids the pending timeout work in progress to fire right away after
665 	 * this TDR finished and before the newly restarted jobs had a
666 	 * chance to complete.
667 	 */
668 	cancel_delayed_work(&sched->work_tdr);
669 }
670 
671 EXPORT_SYMBOL(drm_sched_stop);
672 
673 /**
674  * drm_sched_start - recover jobs after a reset
675  *
676  * @sched: scheduler instance
677  * @full_recovery: proceed with complete sched restart
678  *
679  */
680 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
681 {
682 	struct drm_sched_job *s_job, *tmp;
683 	int r;
684 
685 	/*
686 	 * Locking the list is not required here as the sched thread is parked
687 	 * so no new jobs are being inserted or removed. Also concurrent
688 	 * GPU recovers can't run in parallel.
689 	 */
690 	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
691 		struct dma_fence *fence = s_job->s_fence->parent;
692 
693 		atomic_add(s_job->credits, &sched->credit_count);
694 
695 		if (!full_recovery)
696 			continue;
697 
698 		if (fence) {
699 			r = dma_fence_add_callback(fence, &s_job->cb,
700 						   drm_sched_job_done_cb);
701 			if (r == -ENOENT)
702 				drm_sched_job_done(s_job, fence->error);
703 			else if (r)
704 				DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
705 					  r);
706 		} else
707 			drm_sched_job_done(s_job, -ECANCELED);
708 	}
709 
710 	if (full_recovery)
711 		drm_sched_start_timeout_unlocked(sched);
712 
713 	drm_sched_wqueue_start(sched);
714 }
715 EXPORT_SYMBOL(drm_sched_start);
716 
717 /**
718  * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
719  *
720  * @sched: scheduler instance
721  *
722  * Re-submitting jobs was a concept AMD came up as cheap way to implement
723  * recovery after a job timeout.
724  *
725  * This turned out to be not working very well. First of all there are many
726  * problem with the dma_fence implementation and requirements. Either the
727  * implementation is risking deadlocks with core memory management or violating
728  * documented implementation details of the dma_fence object.
729  *
730  * Drivers can still save and restore their state for recovery operations, but
731  * we shouldn't make this a general scheduler feature around the dma_fence
732  * interface.
733  */
734 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
735 {
736 	struct drm_sched_job *s_job, *tmp;
737 	uint64_t guilty_context;
738 	bool found_guilty = false;
739 	struct dma_fence *fence;
740 
741 	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
742 		struct drm_sched_fence *s_fence = s_job->s_fence;
743 
744 		if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
745 			found_guilty = true;
746 			guilty_context = s_job->s_fence->scheduled.context;
747 		}
748 
749 		if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
750 			dma_fence_set_error(&s_fence->finished, -ECANCELED);
751 
752 		fence = sched->ops->run_job(s_job);
753 
754 		if (IS_ERR_OR_NULL(fence)) {
755 			if (IS_ERR(fence))
756 				dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
757 
758 			s_job->s_fence->parent = NULL;
759 		} else {
760 
761 			s_job->s_fence->parent = dma_fence_get(fence);
762 
763 			/* Drop for orignal kref_init */
764 			dma_fence_put(fence);
765 		}
766 	}
767 }
768 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
769 
770 /**
771  * drm_sched_job_init - init a scheduler job
772  * @job: scheduler job to init
773  * @entity: scheduler entity to use
774  * @credits: the number of credits this job contributes to the schedulers
775  * credit limit
776  * @owner: job owner for debugging
777  *
778  * Refer to drm_sched_entity_push_job() documentation
779  * for locking considerations.
780  *
781  * Drivers must make sure drm_sched_job_cleanup() if this function returns
782  * successfully, even when @job is aborted before drm_sched_job_arm() is called.
783  *
784  * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
785  * has died, which can mean that there's no valid runqueue for a @entity.
786  * This function returns -ENOENT in this case (which probably should be -EIO as
787  * a more meanigful return value).
788  *
789  * Returns 0 for success, negative error code otherwise.
790  */
791 int drm_sched_job_init(struct drm_sched_job *job,
792 		       struct drm_sched_entity *entity,
793 		       u32 credits, void *owner)
794 {
795 	if (!entity->rq) {
796 		/* This will most likely be followed by missing frames
797 		 * or worse--a blank screen--leave a trail in the
798 		 * logs, so this can be debugged easier.
799 		 */
800 		drm_err(job->sched, "%s: entity has no rq!\n", __func__);
801 		return -ENOENT;
802 	}
803 
804 	if (unlikely(!credits)) {
805 		pr_err("*ERROR* %s: credits cannot be 0!\n", __func__);
806 		return -EINVAL;
807 	}
808 
809 	job->entity = entity;
810 	job->credits = credits;
811 	job->s_fence = drm_sched_fence_alloc(entity, owner);
812 	if (!job->s_fence)
813 		return -ENOMEM;
814 
815 	INIT_LIST_HEAD(&job->list);
816 
817 	xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
818 
819 	return 0;
820 }
821 EXPORT_SYMBOL(drm_sched_job_init);
822 
823 /**
824  * drm_sched_job_arm - arm a scheduler job for execution
825  * @job: scheduler job to arm
826  *
827  * This arms a scheduler job for execution. Specifically it initializes the
828  * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
829  * or other places that need to track the completion of this job.
830  *
831  * Refer to drm_sched_entity_push_job() documentation for locking
832  * considerations.
833  *
834  * This can only be called if drm_sched_job_init() succeeded.
835  */
836 void drm_sched_job_arm(struct drm_sched_job *job)
837 {
838 	struct drm_gpu_scheduler *sched;
839 	struct drm_sched_entity *entity = job->entity;
840 
841 	BUG_ON(!entity);
842 	drm_sched_entity_select_rq(entity);
843 	sched = entity->rq->sched;
844 
845 	job->sched = sched;
846 	job->s_priority = entity->priority;
847 	job->id = atomic64_inc_return(&sched->job_id_count);
848 
849 	drm_sched_fence_init(job->s_fence, job->entity);
850 }
851 EXPORT_SYMBOL(drm_sched_job_arm);
852 
853 /**
854  * drm_sched_job_add_dependency - adds the fence as a job dependency
855  * @job: scheduler job to add the dependencies to
856  * @fence: the dma_fence to add to the list of dependencies.
857  *
858  * Note that @fence is consumed in both the success and error cases.
859  *
860  * Returns:
861  * 0 on success, or an error on failing to expand the array.
862  */
863 int drm_sched_job_add_dependency(struct drm_sched_job *job,
864 				 struct dma_fence *fence)
865 {
866 	struct dma_fence *entry;
867 	unsigned long index;
868 	u32 id = 0;
869 	int ret;
870 
871 	if (!fence)
872 		return 0;
873 
874 	/* Deduplicate if we already depend on a fence from the same context.
875 	 * This lets the size of the array of deps scale with the number of
876 	 * engines involved, rather than the number of BOs.
877 	 */
878 	xa_for_each(&job->dependencies, index, entry) {
879 		if (entry->context != fence->context)
880 			continue;
881 
882 		if (dma_fence_is_later(fence, entry)) {
883 			dma_fence_put(entry);
884 			xa_store(&job->dependencies, index, fence, GFP_KERNEL);
885 		} else {
886 			dma_fence_put(fence);
887 		}
888 		return 0;
889 	}
890 
891 	ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
892 	if (ret != 0)
893 		dma_fence_put(fence);
894 
895 	return ret;
896 }
897 EXPORT_SYMBOL(drm_sched_job_add_dependency);
898 
899 /**
900  * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
901  * @job: scheduler job to add the dependencies to
902  * @file: drm file private pointer
903  * @handle: syncobj handle to lookup
904  * @point: timeline point
905  *
906  * This adds the fence matching the given syncobj to @job.
907  *
908  * Returns:
909  * 0 on success, or an error on failing to expand the array.
910  */
911 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
912 					 struct drm_file *file,
913 					 u32 handle,
914 					 u32 point)
915 {
916 	struct dma_fence *fence;
917 	int ret;
918 
919 	ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
920 	if (ret)
921 		return ret;
922 
923 	return drm_sched_job_add_dependency(job, fence);
924 }
925 EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
926 
927 /**
928  * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
929  * @job: scheduler job to add the dependencies to
930  * @resv: the dma_resv object to get the fences from
931  * @usage: the dma_resv_usage to use to filter the fences
932  *
933  * This adds all fences matching the given usage from @resv to @job.
934  * Must be called with the @resv lock held.
935  *
936  * Returns:
937  * 0 on success, or an error on failing to expand the array.
938  */
939 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
940 					struct dma_resv *resv,
941 					enum dma_resv_usage usage)
942 {
943 	struct dma_resv_iter cursor;
944 	struct dma_fence *fence;
945 	int ret;
946 
947 	dma_resv_assert_held(resv);
948 
949 	dma_resv_for_each_fence(&cursor, resv, usage, fence) {
950 		/* Make sure to grab an additional ref on the added fence */
951 		dma_fence_get(fence);
952 		ret = drm_sched_job_add_dependency(job, fence);
953 		if (ret) {
954 			dma_fence_put(fence);
955 			return ret;
956 		}
957 	}
958 	return 0;
959 }
960 EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
961 
962 /**
963  * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
964  *   dependencies
965  * @job: scheduler job to add the dependencies to
966  * @obj: the gem object to add new dependencies from.
967  * @write: whether the job might write the object (so we need to depend on
968  * shared fences in the reservation object).
969  *
970  * This should be called after drm_gem_lock_reservations() on your array of
971  * GEM objects used in the job but before updating the reservations with your
972  * own fences.
973  *
974  * Returns:
975  * 0 on success, or an error on failing to expand the array.
976  */
977 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
978 					    struct drm_gem_object *obj,
979 					    bool write)
980 {
981 	return drm_sched_job_add_resv_dependencies(job, obj->resv,
982 						   dma_resv_usage_rw(write));
983 }
984 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
985 
986 /**
987  * drm_sched_job_cleanup - clean up scheduler job resources
988  * @job: scheduler job to clean up
989  *
990  * Cleans up the resources allocated with drm_sched_job_init().
991  *
992  * Drivers should call this from their error unwind code if @job is aborted
993  * before drm_sched_job_arm() is called.
994  *
995  * After that point of no return @job is committed to be executed by the
996  * scheduler, and this function should be called from the
997  * &drm_sched_backend_ops.free_job callback.
998  */
999 void drm_sched_job_cleanup(struct drm_sched_job *job)
1000 {
1001 	struct dma_fence *fence;
1002 	unsigned long index;
1003 
1004 	if (kref_read(&job->s_fence->finished.refcount)) {
1005 		/* drm_sched_job_arm() has been called */
1006 		dma_fence_put(&job->s_fence->finished);
1007 	} else {
1008 		/* aborted job before committing to run it */
1009 		drm_sched_fence_free(job->s_fence);
1010 	}
1011 
1012 	job->s_fence = NULL;
1013 
1014 	xa_for_each(&job->dependencies, index, fence) {
1015 		dma_fence_put(fence);
1016 	}
1017 	xa_destroy(&job->dependencies);
1018 
1019 }
1020 EXPORT_SYMBOL(drm_sched_job_cleanup);
1021 
1022 /**
1023  * drm_sched_wakeup - Wake up the scheduler if it is ready to queue
1024  * @sched: scheduler instance
1025  * @entity: the scheduler entity
1026  *
1027  * Wake up the scheduler if we can queue jobs.
1028  */
1029 void drm_sched_wakeup(struct drm_gpu_scheduler *sched,
1030 		      struct drm_sched_entity *entity)
1031 {
1032 	if (drm_sched_can_queue(sched, entity))
1033 		drm_sched_run_job_queue(sched);
1034 }
1035 
1036 /**
1037  * drm_sched_select_entity - Select next entity to process
1038  *
1039  * @sched: scheduler instance
1040  *
1041  * Return an entity to process or NULL if none are found.
1042  *
1043  * Note, that we break out of the for-loop when "entity" is non-null, which can
1044  * also be an error-pointer--this assures we don't process lower priority
1045  * run-queues. See comments in the respectively called functions.
1046  */
1047 static struct drm_sched_entity *
1048 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
1049 {
1050 	struct drm_sched_entity *entity;
1051 	int i;
1052 
1053 	/* Start with the highest priority.
1054 	 */
1055 	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1056 		entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
1057 			drm_sched_rq_select_entity_fifo(sched, sched->sched_rq[i]) :
1058 			drm_sched_rq_select_entity_rr(sched, sched->sched_rq[i]);
1059 		if (entity)
1060 			break;
1061 	}
1062 
1063 	return IS_ERR(entity) ? NULL : entity;
1064 }
1065 
1066 /**
1067  * drm_sched_get_finished_job - fetch the next finished job to be destroyed
1068  *
1069  * @sched: scheduler instance
1070  *
1071  * Returns the next finished job from the pending list (if there is one)
1072  * ready for it to be destroyed.
1073  */
1074 static struct drm_sched_job *
1075 drm_sched_get_finished_job(struct drm_gpu_scheduler *sched)
1076 {
1077 	struct drm_sched_job *job, *next;
1078 
1079 	spin_lock(&sched->job_list_lock);
1080 
1081 	job = list_first_entry_or_null(&sched->pending_list,
1082 				       struct drm_sched_job, list);
1083 
1084 	if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
1085 		/* remove job from pending_list */
1086 		list_del_init(&job->list);
1087 
1088 		/* cancel this job's TO timer */
1089 		cancel_delayed_work(&sched->work_tdr);
1090 		/* make the scheduled timestamp more accurate */
1091 		next = list_first_entry_or_null(&sched->pending_list,
1092 						typeof(*next), list);
1093 
1094 		if (next) {
1095 			if (test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT,
1096 				     &next->s_fence->scheduled.flags))
1097 				next->s_fence->scheduled.timestamp =
1098 					dma_fence_timestamp(&job->s_fence->finished);
1099 			/* start TO timer for next job */
1100 			drm_sched_start_timeout(sched);
1101 		}
1102 	} else {
1103 		job = NULL;
1104 	}
1105 
1106 	spin_unlock(&sched->job_list_lock);
1107 
1108 	return job;
1109 }
1110 
1111 /**
1112  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
1113  * @sched_list: list of drm_gpu_schedulers
1114  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
1115  *
1116  * Returns pointer of the sched with the least load or NULL if none of the
1117  * drm_gpu_schedulers are ready
1118  */
1119 struct drm_gpu_scheduler *
1120 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
1121 		     unsigned int num_sched_list)
1122 {
1123 	struct drm_gpu_scheduler *sched, *picked_sched = NULL;
1124 	int i;
1125 	unsigned int min_score = UINT_MAX, num_score;
1126 
1127 	for (i = 0; i < num_sched_list; ++i) {
1128 		sched = sched_list[i];
1129 
1130 		if (!sched->ready) {
1131 			DRM_WARN("scheduler %s is not ready, skipping",
1132 				 sched->name);
1133 			continue;
1134 		}
1135 
1136 		num_score = atomic_read(sched->score);
1137 		if (num_score < min_score) {
1138 			min_score = num_score;
1139 			picked_sched = sched;
1140 		}
1141 	}
1142 
1143 	return picked_sched;
1144 }
1145 EXPORT_SYMBOL(drm_sched_pick_best);
1146 
1147 /**
1148  * drm_sched_free_job_work - worker to call free_job
1149  *
1150  * @w: free job work
1151  */
1152 static void drm_sched_free_job_work(struct work_struct *w)
1153 {
1154 	struct drm_gpu_scheduler *sched =
1155 		container_of(w, struct drm_gpu_scheduler, work_free_job);
1156 	struct drm_sched_job *job;
1157 
1158 	if (READ_ONCE(sched->pause_submit))
1159 		return;
1160 
1161 	job = drm_sched_get_finished_job(sched);
1162 	if (job)
1163 		sched->ops->free_job(job);
1164 
1165 	drm_sched_run_free_queue(sched);
1166 	drm_sched_run_job_queue(sched);
1167 }
1168 
1169 /**
1170  * drm_sched_run_job_work - worker to call run_job
1171  *
1172  * @w: run job work
1173  */
1174 static void drm_sched_run_job_work(struct work_struct *w)
1175 {
1176 	struct drm_gpu_scheduler *sched =
1177 		container_of(w, struct drm_gpu_scheduler, work_run_job);
1178 	struct drm_sched_entity *entity;
1179 	struct dma_fence *fence;
1180 	struct drm_sched_fence *s_fence;
1181 	struct drm_sched_job *sched_job;
1182 	int r;
1183 
1184 	if (READ_ONCE(sched->pause_submit))
1185 		return;
1186 
1187 	/* Find entity with a ready job */
1188 	entity = drm_sched_select_entity(sched);
1189 	if (!entity)
1190 		return;	/* No more work */
1191 
1192 	sched_job = drm_sched_entity_pop_job(entity);
1193 	if (!sched_job) {
1194 		complete_all(&entity->entity_idle);
1195 		drm_sched_run_job_queue(sched);
1196 		return;
1197 	}
1198 
1199 	s_fence = sched_job->s_fence;
1200 
1201 	atomic_add(sched_job->credits, &sched->credit_count);
1202 	drm_sched_job_begin(sched_job);
1203 
1204 	trace_drm_run_job(sched_job, entity);
1205 	fence = sched->ops->run_job(sched_job);
1206 	complete_all(&entity->entity_idle);
1207 	drm_sched_fence_scheduled(s_fence, fence);
1208 
1209 	if (!IS_ERR_OR_NULL(fence)) {
1210 		/* Drop for original kref_init of the fence */
1211 		dma_fence_put(fence);
1212 
1213 		r = dma_fence_add_callback(fence, &sched_job->cb,
1214 					   drm_sched_job_done_cb);
1215 		if (r == -ENOENT)
1216 			drm_sched_job_done(sched_job, fence->error);
1217 		else if (r)
1218 			DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
1219 	} else {
1220 		drm_sched_job_done(sched_job, IS_ERR(fence) ?
1221 				   PTR_ERR(fence) : 0);
1222 	}
1223 
1224 	wake_up(&sched->job_scheduled);
1225 	drm_sched_run_job_queue(sched);
1226 }
1227 
1228 /**
1229  * drm_sched_init - Init a gpu scheduler instance
1230  *
1231  * @sched: scheduler instance
1232  * @ops: backend operations for this scheduler
1233  * @submit_wq: workqueue to use for submission. If NULL, an ordered wq is
1234  *	       allocated and used
1235  * @num_rqs: number of runqueues, one for each priority, up to DRM_SCHED_PRIORITY_COUNT
1236  * @credit_limit: the number of credits this scheduler can hold from all jobs
1237  * @hang_limit: number of times to allow a job to hang before dropping it
1238  * @timeout: timeout value in jiffies for the scheduler
1239  * @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
1240  *		used
1241  * @score: optional score atomic shared with other schedulers
1242  * @name: name used for debugging
1243  * @dev: target &struct device
1244  *
1245  * Return 0 on success, otherwise error code.
1246  */
1247 int drm_sched_init(struct drm_gpu_scheduler *sched,
1248 		   const struct drm_sched_backend_ops *ops,
1249 		   struct workqueue_struct *submit_wq,
1250 		   u32 num_rqs, u32 credit_limit, unsigned int hang_limit,
1251 		   long timeout, struct workqueue_struct *timeout_wq,
1252 		   atomic_t *score, const char *name, struct device *dev)
1253 {
1254 	int i;
1255 
1256 	sched->ops = ops;
1257 	sched->credit_limit = credit_limit;
1258 	sched->name = name;
1259 	sched->timeout = timeout;
1260 	sched->timeout_wq = timeout_wq ? : system_wq;
1261 	sched->hang_limit = hang_limit;
1262 	sched->score = score ? score : &sched->_score;
1263 	sched->dev = dev;
1264 
1265 	if (num_rqs > DRM_SCHED_PRIORITY_COUNT) {
1266 		/* This is a gross violation--tell drivers what the  problem is.
1267 		 */
1268 		drm_err(sched, "%s: num_rqs cannot be greater than DRM_SCHED_PRIORITY_COUNT\n",
1269 			__func__);
1270 		return -EINVAL;
1271 	} else if (sched->sched_rq) {
1272 		/* Not an error, but warn anyway so drivers can
1273 		 * fine-tune their DRM calling order, and return all
1274 		 * is good.
1275 		 */
1276 		drm_warn(sched, "%s: scheduler already initialized!\n", __func__);
1277 		return 0;
1278 	}
1279 
1280 	if (submit_wq) {
1281 		sched->submit_wq = submit_wq;
1282 		sched->own_submit_wq = false;
1283 	} else {
1284 		sched->submit_wq = alloc_ordered_workqueue(name, 0);
1285 		if (!sched->submit_wq)
1286 			return -ENOMEM;
1287 
1288 		sched->own_submit_wq = true;
1289 	}
1290 
1291 	sched->sched_rq = kmalloc_array(num_rqs, sizeof(*sched->sched_rq),
1292 					GFP_KERNEL | __GFP_ZERO);
1293 	if (!sched->sched_rq)
1294 		goto Out_check_own;
1295 	sched->num_rqs = num_rqs;
1296 	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1297 		sched->sched_rq[i] = kzalloc(sizeof(*sched->sched_rq[i]), GFP_KERNEL);
1298 		if (!sched->sched_rq[i])
1299 			goto Out_unroll;
1300 		drm_sched_rq_init(sched, sched->sched_rq[i]);
1301 	}
1302 
1303 	init_waitqueue_head(&sched->job_scheduled);
1304 	INIT_LIST_HEAD(&sched->pending_list);
1305 	spin_lock_init(&sched->job_list_lock);
1306 	atomic_set(&sched->credit_count, 0);
1307 	INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1308 	INIT_WORK(&sched->work_run_job, drm_sched_run_job_work);
1309 	INIT_WORK(&sched->work_free_job, drm_sched_free_job_work);
1310 	atomic_set(&sched->_score, 0);
1311 	atomic64_set(&sched->job_id_count, 0);
1312 	sched->pause_submit = false;
1313 
1314 	sched->ready = true;
1315 	return 0;
1316 Out_unroll:
1317 	for (--i ; i >= DRM_SCHED_PRIORITY_KERNEL; i--)
1318 		kfree(sched->sched_rq[i]);
1319 
1320 	kfree(sched->sched_rq);
1321 	sched->sched_rq = NULL;
1322 Out_check_own:
1323 	if (sched->own_submit_wq)
1324 		destroy_workqueue(sched->submit_wq);
1325 	drm_err(sched, "%s: Failed to setup GPU scheduler--out of memory\n", __func__);
1326 	return -ENOMEM;
1327 }
1328 EXPORT_SYMBOL(drm_sched_init);
1329 
1330 /**
1331  * drm_sched_fini - Destroy a gpu scheduler
1332  *
1333  * @sched: scheduler instance
1334  *
1335  * Tears down and cleans up the scheduler.
1336  */
1337 void drm_sched_fini(struct drm_gpu_scheduler *sched)
1338 {
1339 	struct drm_sched_entity *s_entity;
1340 	int i;
1341 
1342 	drm_sched_wqueue_stop(sched);
1343 
1344 	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1345 		struct drm_sched_rq *rq = sched->sched_rq[i];
1346 
1347 		spin_lock(&rq->lock);
1348 		list_for_each_entry(s_entity, &rq->entities, list)
1349 			/*
1350 			 * Prevents reinsertion and marks job_queue as idle,
1351 			 * it will removed from rq in drm_sched_entity_fini
1352 			 * eventually
1353 			 */
1354 			s_entity->stopped = true;
1355 		spin_unlock(&rq->lock);
1356 		kfree(sched->sched_rq[i]);
1357 	}
1358 
1359 	/* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1360 	wake_up_all(&sched->job_scheduled);
1361 
1362 	/* Confirm no work left behind accessing device structures */
1363 	cancel_delayed_work_sync(&sched->work_tdr);
1364 
1365 	if (sched->own_submit_wq)
1366 		destroy_workqueue(sched->submit_wq);
1367 	sched->ready = false;
1368 	kfree(sched->sched_rq);
1369 	sched->sched_rq = NULL;
1370 }
1371 EXPORT_SYMBOL(drm_sched_fini);
1372 
1373 /**
1374  * drm_sched_increase_karma - Update sched_entity guilty flag
1375  *
1376  * @bad: The job guilty of time out
1377  *
1378  * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1379  * limit of the scheduler then the respective sched entity is marked guilty and
1380  * jobs from it will not be scheduled further
1381  */
1382 void drm_sched_increase_karma(struct drm_sched_job *bad)
1383 {
1384 	int i;
1385 	struct drm_sched_entity *tmp;
1386 	struct drm_sched_entity *entity;
1387 	struct drm_gpu_scheduler *sched = bad->sched;
1388 
1389 	/* don't change @bad's karma if it's from KERNEL RQ,
1390 	 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1391 	 * corrupt but keep in mind that kernel jobs always considered good.
1392 	 */
1393 	if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1394 		atomic_inc(&bad->karma);
1395 
1396 		for (i = DRM_SCHED_PRIORITY_HIGH; i < sched->num_rqs; i++) {
1397 			struct drm_sched_rq *rq = sched->sched_rq[i];
1398 
1399 			spin_lock(&rq->lock);
1400 			list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1401 				if (bad->s_fence->scheduled.context ==
1402 				    entity->fence_context) {
1403 					if (entity->guilty)
1404 						atomic_set(entity->guilty, 1);
1405 					break;
1406 				}
1407 			}
1408 			spin_unlock(&rq->lock);
1409 			if (&entity->list != &rq->entities)
1410 				break;
1411 		}
1412 	}
1413 }
1414 EXPORT_SYMBOL(drm_sched_increase_karma);
1415 
1416 /**
1417  * drm_sched_wqueue_ready - Is the scheduler ready for submission
1418  *
1419  * @sched: scheduler instance
1420  *
1421  * Returns true if submission is ready
1422  */
1423 bool drm_sched_wqueue_ready(struct drm_gpu_scheduler *sched)
1424 {
1425 	return sched->ready;
1426 }
1427 EXPORT_SYMBOL(drm_sched_wqueue_ready);
1428 
1429 /**
1430  * drm_sched_wqueue_stop - stop scheduler submission
1431  *
1432  * @sched: scheduler instance
1433  */
1434 void drm_sched_wqueue_stop(struct drm_gpu_scheduler *sched)
1435 {
1436 	WRITE_ONCE(sched->pause_submit, true);
1437 	cancel_work_sync(&sched->work_run_job);
1438 	cancel_work_sync(&sched->work_free_job);
1439 }
1440 EXPORT_SYMBOL(drm_sched_wqueue_stop);
1441 
1442 /**
1443  * drm_sched_wqueue_start - start scheduler submission
1444  *
1445  * @sched: scheduler instance
1446  */
1447 void drm_sched_wqueue_start(struct drm_gpu_scheduler *sched)
1448 {
1449 	WRITE_ONCE(sched->pause_submit, false);
1450 	queue_work(sched->submit_wq, &sched->work_run_job);
1451 	queue_work(sched->submit_wq, &sched->work_free_job);
1452 }
1453 EXPORT_SYMBOL(drm_sched_wqueue_start);
1454