xref: /linux/drivers/gpu/drm/scheduler/sched_main.c (revision 504f9bdd3a1588604b0452bfe927ff86e5f6e6df)
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 an entity are always scheduled in the order in which 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 
69 #include <linux/export.h>
70 #include <linux/wait.h>
71 #include <linux/sched.h>
72 #include <linux/completion.h>
73 #include <linux/dma-resv.h>
74 #include <uapi/linux/sched/types.h>
75 
76 #include <drm/drm_print.h>
77 #include <drm/drm_gem.h>
78 #include <drm/drm_syncobj.h>
79 #include <drm/gpu_scheduler.h>
80 #include <drm/spsc_queue.h>
81 
82 #include "sched_internal.h"
83 
84 #define CREATE_TRACE_POINTS
85 #include "gpu_scheduler_trace.h"
86 
87 int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
88 
89 /**
90  * DOC: sched_policy (int)
91  * Used to override default entities scheduling policy in a run queue.
92  */
93 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).");
94 module_param_named(sched_policy, drm_sched_policy, int, 0444);
95 
96 static u32 drm_sched_available_credits(struct drm_gpu_scheduler *sched)
97 {
98 	u32 credits;
99 
100 	WARN_ON(check_sub_overflow(sched->credit_limit,
101 				   atomic_read(&sched->credit_count),
102 				   &credits));
103 
104 	return credits;
105 }
106 
107 /**
108  * drm_sched_can_queue -- Can we queue more to the hardware?
109  * @sched: scheduler instance
110  * @entity: the scheduler entity
111  *
112  * Return true if we can push at least one more job from @entity, false
113  * otherwise.
114  */
115 static bool drm_sched_can_queue(struct drm_gpu_scheduler *sched,
116 				struct drm_sched_entity *entity)
117 {
118 	struct drm_sched_job *s_job;
119 
120 	s_job = drm_sched_entity_queue_peek(entity);
121 	if (!s_job)
122 		return false;
123 
124 	/* If a job exceeds the credit limit, truncate it to the credit limit
125 	 * itself to guarantee forward progress.
126 	 */
127 	if (s_job->credits > sched->credit_limit) {
128 		dev_WARN(sched->dev,
129 			 "Jobs may not exceed the credit limit, truncate.\n");
130 		s_job->credits = sched->credit_limit;
131 	}
132 
133 	return drm_sched_available_credits(sched) >= s_job->credits;
134 }
135 
136 static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
137 							    const struct rb_node *b)
138 {
139 	struct drm_sched_entity *ent_a =  rb_entry((a), struct drm_sched_entity, rb_tree_node);
140 	struct drm_sched_entity *ent_b =  rb_entry((b), struct drm_sched_entity, rb_tree_node);
141 
142 	return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
143 }
144 
145 static void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity,
146 					    struct drm_sched_rq *rq)
147 {
148 	if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
149 		rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
150 		RB_CLEAR_NODE(&entity->rb_tree_node);
151 	}
152 }
153 
154 void drm_sched_rq_update_fifo_locked(struct drm_sched_entity *entity,
155 				     struct drm_sched_rq *rq,
156 				     ktime_t ts)
157 {
158 	/*
159 	 * Both locks need to be grabbed, one to protect from entity->rq change
160 	 * for entity from within concurrent drm_sched_entity_select_rq and the
161 	 * other to update the rb tree structure.
162 	 */
163 	lockdep_assert_held(&entity->lock);
164 	lockdep_assert_held(&rq->lock);
165 
166 	drm_sched_rq_remove_fifo_locked(entity, rq);
167 
168 	entity->oldest_job_waiting = ts;
169 
170 	rb_add_cached(&entity->rb_tree_node, &rq->rb_tree_root,
171 		      drm_sched_entity_compare_before);
172 }
173 
174 /**
175  * drm_sched_rq_init - initialize a given run queue struct
176  *
177  * @sched: scheduler instance to associate with this run queue
178  * @rq: scheduler run queue
179  *
180  * Initializes a scheduler runqueue.
181  */
182 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
183 			      struct drm_sched_rq *rq)
184 {
185 	spin_lock_init(&rq->lock);
186 	INIT_LIST_HEAD(&rq->entities);
187 	rq->rb_tree_root = RB_ROOT_CACHED;
188 	rq->current_entity = NULL;
189 	rq->sched = sched;
190 }
191 
192 /**
193  * drm_sched_rq_add_entity - add an entity
194  *
195  * @rq: scheduler run queue
196  * @entity: scheduler entity
197  *
198  * Adds a scheduler entity to the run queue.
199  */
200 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
201 			     struct drm_sched_entity *entity)
202 {
203 	lockdep_assert_held(&entity->lock);
204 	lockdep_assert_held(&rq->lock);
205 
206 	if (!list_empty(&entity->list))
207 		return;
208 
209 	atomic_inc(rq->sched->score);
210 	list_add_tail(&entity->list, &rq->entities);
211 }
212 
213 /**
214  * drm_sched_rq_remove_entity - remove an entity
215  *
216  * @rq: scheduler run queue
217  * @entity: scheduler entity
218  *
219  * Removes a scheduler entity from the run queue.
220  */
221 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
222 				struct drm_sched_entity *entity)
223 {
224 	lockdep_assert_held(&entity->lock);
225 
226 	if (list_empty(&entity->list))
227 		return;
228 
229 	spin_lock(&rq->lock);
230 
231 	atomic_dec(rq->sched->score);
232 	list_del_init(&entity->list);
233 
234 	if (rq->current_entity == entity)
235 		rq->current_entity = NULL;
236 
237 	if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
238 		drm_sched_rq_remove_fifo_locked(entity, rq);
239 
240 	spin_unlock(&rq->lock);
241 }
242 
243 /**
244  * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
245  *
246  * @sched: the gpu scheduler
247  * @rq: scheduler run queue to check.
248  *
249  * Try to find the next ready entity.
250  *
251  * Return an entity if one is found; return an error-pointer (!NULL) if an
252  * entity was ready, but the scheduler had insufficient credits to accommodate
253  * its job; return NULL, if no ready entity was found.
254  */
255 static struct drm_sched_entity *
256 drm_sched_rq_select_entity_rr(struct drm_gpu_scheduler *sched,
257 			      struct drm_sched_rq *rq)
258 {
259 	struct drm_sched_entity *entity;
260 
261 	spin_lock(&rq->lock);
262 
263 	entity = rq->current_entity;
264 	if (entity) {
265 		list_for_each_entry_continue(entity, &rq->entities, list) {
266 			if (drm_sched_entity_is_ready(entity))
267 				goto found;
268 		}
269 	}
270 
271 	list_for_each_entry(entity, &rq->entities, list) {
272 		if (drm_sched_entity_is_ready(entity))
273 			goto found;
274 
275 		if (entity == rq->current_entity)
276 			break;
277 	}
278 
279 	spin_unlock(&rq->lock);
280 
281 	return NULL;
282 
283 found:
284 	if (!drm_sched_can_queue(sched, entity)) {
285 		/*
286 		 * If scheduler cannot take more jobs signal the caller to not
287 		 * consider lower priority queues.
288 		 */
289 		entity = ERR_PTR(-ENOSPC);
290 	} else {
291 		rq->current_entity = entity;
292 		reinit_completion(&entity->entity_idle);
293 	}
294 
295 	spin_unlock(&rq->lock);
296 
297 	return entity;
298 }
299 
300 /**
301  * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
302  *
303  * @sched: the gpu scheduler
304  * @rq: scheduler run queue to check.
305  *
306  * Find oldest waiting ready entity.
307  *
308  * Return an entity if one is found; return an error-pointer (!NULL) if an
309  * entity was ready, but the scheduler had insufficient credits to accommodate
310  * its job; return NULL, if no ready entity was found.
311  */
312 static struct drm_sched_entity *
313 drm_sched_rq_select_entity_fifo(struct drm_gpu_scheduler *sched,
314 				struct drm_sched_rq *rq)
315 {
316 	struct rb_node *rb;
317 
318 	spin_lock(&rq->lock);
319 	for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
320 		struct drm_sched_entity *entity;
321 
322 		entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
323 		if (drm_sched_entity_is_ready(entity)) {
324 			/* If we can't queue yet, preserve the current entity in
325 			 * terms of fairness.
326 			 */
327 			if (!drm_sched_can_queue(sched, entity)) {
328 				spin_unlock(&rq->lock);
329 				return ERR_PTR(-ENOSPC);
330 			}
331 
332 			reinit_completion(&entity->entity_idle);
333 			break;
334 		}
335 	}
336 	spin_unlock(&rq->lock);
337 
338 	return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
339 }
340 
341 /**
342  * drm_sched_run_job_queue - enqueue run-job work
343  * @sched: scheduler instance
344  */
345 static void drm_sched_run_job_queue(struct drm_gpu_scheduler *sched)
346 {
347 	if (!READ_ONCE(sched->pause_submit))
348 		queue_work(sched->submit_wq, &sched->work_run_job);
349 }
350 
351 /**
352  * __drm_sched_run_free_queue - enqueue free-job work
353  * @sched: scheduler instance
354  */
355 static void __drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
356 {
357 	if (!READ_ONCE(sched->pause_submit))
358 		queue_work(sched->submit_wq, &sched->work_free_job);
359 }
360 
361 /**
362  * drm_sched_run_free_queue - enqueue free-job work if ready
363  * @sched: scheduler instance
364  */
365 static void drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
366 {
367 	struct drm_sched_job *job;
368 
369 	spin_lock(&sched->job_list_lock);
370 	job = list_first_entry_or_null(&sched->pending_list,
371 				       struct drm_sched_job, list);
372 	if (job && dma_fence_is_signaled(&job->s_fence->finished))
373 		__drm_sched_run_free_queue(sched);
374 	spin_unlock(&sched->job_list_lock);
375 }
376 
377 /**
378  * drm_sched_job_done - complete a job
379  * @s_job: pointer to the job which is done
380  *
381  * Finish the job's fence and resubmit the work items.
382  */
383 static void drm_sched_job_done(struct drm_sched_job *s_job, int result)
384 {
385 	struct drm_sched_fence *s_fence = s_job->s_fence;
386 	struct drm_gpu_scheduler *sched = s_fence->sched;
387 
388 	atomic_sub(s_job->credits, &sched->credit_count);
389 	atomic_dec(sched->score);
390 
391 	trace_drm_sched_job_done(s_fence);
392 
393 	dma_fence_get(&s_fence->finished);
394 	drm_sched_fence_finished(s_fence, result);
395 	dma_fence_put(&s_fence->finished);
396 	__drm_sched_run_free_queue(sched);
397 }
398 
399 /**
400  * drm_sched_job_done_cb - the callback for a done job
401  * @f: fence
402  * @cb: fence callbacks
403  */
404 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
405 {
406 	struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
407 
408 	drm_sched_job_done(s_job, f->error);
409 }
410 
411 /**
412  * drm_sched_start_timeout - start timeout for reset worker
413  *
414  * @sched: scheduler instance to start the worker for
415  *
416  * Start the timeout for the given scheduler.
417  */
418 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
419 {
420 	lockdep_assert_held(&sched->job_list_lock);
421 
422 	if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
423 	    !list_empty(&sched->pending_list))
424 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
425 }
426 
427 static void drm_sched_start_timeout_unlocked(struct drm_gpu_scheduler *sched)
428 {
429 	spin_lock(&sched->job_list_lock);
430 	drm_sched_start_timeout(sched);
431 	spin_unlock(&sched->job_list_lock);
432 }
433 
434 /**
435  * drm_sched_tdr_queue_imm: - immediately start job timeout handler
436  *
437  * @sched: scheduler for which the timeout handling should be started.
438  *
439  * Start timeout handling immediately for the named scheduler.
440  */
441 void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched)
442 {
443 	spin_lock(&sched->job_list_lock);
444 	sched->timeout = 0;
445 	drm_sched_start_timeout(sched);
446 	spin_unlock(&sched->job_list_lock);
447 }
448 EXPORT_SYMBOL(drm_sched_tdr_queue_imm);
449 
450 /**
451  * drm_sched_fault - immediately start timeout handler
452  *
453  * @sched: scheduler where the timeout handling should be started.
454  *
455  * Start timeout handling immediately when the driver detects a hardware fault.
456  */
457 void drm_sched_fault(struct drm_gpu_scheduler *sched)
458 {
459 	if (sched->timeout_wq)
460 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
461 }
462 EXPORT_SYMBOL(drm_sched_fault);
463 
464 /**
465  * drm_sched_suspend_timeout - Suspend scheduler job timeout
466  *
467  * @sched: scheduler instance for which to suspend the timeout
468  *
469  * Suspend the delayed work timeout for the scheduler. This is done by
470  * modifying the delayed work timeout to an arbitrary large value,
471  * MAX_SCHEDULE_TIMEOUT in this case.
472  *
473  * Returns the timeout remaining
474  *
475  */
476 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
477 {
478 	unsigned long sched_timeout, now = jiffies;
479 
480 	sched_timeout = sched->work_tdr.timer.expires;
481 
482 	/*
483 	 * Modify the timeout to an arbitrarily large value. This also prevents
484 	 * the timeout to be restarted when new submissions arrive
485 	 */
486 	if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
487 			&& time_after(sched_timeout, now))
488 		return sched_timeout - now;
489 	else
490 		return sched->timeout;
491 }
492 EXPORT_SYMBOL(drm_sched_suspend_timeout);
493 
494 /**
495  * drm_sched_resume_timeout - Resume scheduler job timeout
496  *
497  * @sched: scheduler instance for which to resume the timeout
498  * @remaining: remaining timeout
499  *
500  * Resume the delayed work timeout for the scheduler.
501  */
502 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
503 		unsigned long remaining)
504 {
505 	spin_lock(&sched->job_list_lock);
506 
507 	if (list_empty(&sched->pending_list))
508 		cancel_delayed_work(&sched->work_tdr);
509 	else
510 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
511 
512 	spin_unlock(&sched->job_list_lock);
513 }
514 EXPORT_SYMBOL(drm_sched_resume_timeout);
515 
516 static void drm_sched_job_begin(struct drm_sched_job *s_job)
517 {
518 	struct drm_gpu_scheduler *sched = s_job->sched;
519 
520 	spin_lock(&sched->job_list_lock);
521 	list_add_tail(&s_job->list, &sched->pending_list);
522 	drm_sched_start_timeout(sched);
523 	spin_unlock(&sched->job_list_lock);
524 }
525 
526 static void drm_sched_job_timedout(struct work_struct *work)
527 {
528 	struct drm_gpu_scheduler *sched;
529 	struct drm_sched_job *job;
530 	enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
531 
532 	sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
533 
534 	/* Protects against concurrent deletion in drm_sched_get_finished_job */
535 	spin_lock(&sched->job_list_lock);
536 	job = list_first_entry_or_null(&sched->pending_list,
537 				       struct drm_sched_job, list);
538 
539 	if (job) {
540 		/*
541 		 * Remove the bad job so it cannot be freed by a concurrent
542 		 * &struct drm_sched_backend_ops.free_job. It will be
543 		 * reinserted after the scheduler's work items have been
544 		 * cancelled, at which point it's safe.
545 		 */
546 		list_del_init(&job->list);
547 		spin_unlock(&sched->job_list_lock);
548 
549 		status = job->sched->ops->timedout_job(job);
550 
551 		/*
552 		 * Guilty job did complete and hence needs to be manually removed
553 		 * See drm_sched_stop doc.
554 		 */
555 		if (sched->free_guilty) {
556 			job->sched->ops->free_job(job);
557 			sched->free_guilty = false;
558 		}
559 	} else {
560 		spin_unlock(&sched->job_list_lock);
561 	}
562 
563 	if (status != DRM_GPU_SCHED_STAT_ENODEV)
564 		drm_sched_start_timeout_unlocked(sched);
565 }
566 
567 /**
568  * drm_sched_stop - stop the scheduler
569  *
570  * @sched: scheduler instance
571  * @bad: job which caused the time out
572  *
573  * Stop the scheduler and also removes and frees all completed jobs.
574  * Note: bad job will not be freed as it might be used later and so it's
575  * callers responsibility to release it manually if it's not part of the
576  * pending list any more.
577  *
578  * This function is typically used for reset recovery (see the docu of
579  * drm_sched_backend_ops.timedout_job() for details). Do not call it for
580  * scheduler teardown, i.e., before calling drm_sched_fini().
581  */
582 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
583 {
584 	struct drm_sched_job *s_job, *tmp;
585 
586 	drm_sched_wqueue_stop(sched);
587 
588 	/*
589 	 * Reinsert back the bad job here - now it's safe as
590 	 * drm_sched_get_finished_job() cannot race against us and release the
591 	 * bad job at this point - we parked (waited for) any in progress
592 	 * (earlier) cleanups and drm_sched_get_finished_job() will not be
593 	 * called now until the scheduler's work items are submitted again.
594 	 */
595 	if (bad && bad->sched == sched)
596 		/*
597 		 * Add at the head of the queue to reflect it was the earliest
598 		 * job extracted.
599 		 */
600 		list_add(&bad->list, &sched->pending_list);
601 
602 	/*
603 	 * Iterate the job list from later to  earlier one and either deactive
604 	 * their HW callbacks or remove them from pending list if they already
605 	 * signaled.
606 	 * This iteration is thread safe as the scheduler's work items have been
607 	 * cancelled.
608 	 */
609 	list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
610 					 list) {
611 		if (s_job->s_fence->parent &&
612 		    dma_fence_remove_callback(s_job->s_fence->parent,
613 					      &s_job->cb)) {
614 			dma_fence_put(s_job->s_fence->parent);
615 			s_job->s_fence->parent = NULL;
616 			atomic_sub(s_job->credits, &sched->credit_count);
617 		} else {
618 			/*
619 			 * remove job from pending_list.
620 			 * Locking here is for concurrent resume timeout
621 			 */
622 			spin_lock(&sched->job_list_lock);
623 			list_del_init(&s_job->list);
624 			spin_unlock(&sched->job_list_lock);
625 
626 			/*
627 			 * Wait for job's HW fence callback to finish using s_job
628 			 * before releasing it.
629 			 *
630 			 * Job is still alive so fence refcount at least 1
631 			 */
632 			dma_fence_wait(&s_job->s_fence->finished, false);
633 
634 			/*
635 			 * We must keep bad job alive for later use during
636 			 * recovery by some of the drivers but leave a hint
637 			 * that the guilty job must be released.
638 			 */
639 			if (bad != s_job)
640 				sched->ops->free_job(s_job);
641 			else
642 				sched->free_guilty = true;
643 		}
644 	}
645 
646 	/*
647 	 * Stop pending timer in flight as we rearm it in  drm_sched_start. This
648 	 * avoids the pending timeout work in progress to fire right away after
649 	 * this TDR finished and before the newly restarted jobs had a
650 	 * chance to complete.
651 	 */
652 	cancel_delayed_work(&sched->work_tdr);
653 }
654 EXPORT_SYMBOL(drm_sched_stop);
655 
656 /**
657  * drm_sched_start - recover jobs after a reset
658  *
659  * @sched: scheduler instance
660  * @errno: error to set on the pending fences
661  *
662  * This function is typically used for reset recovery (see the docu of
663  * drm_sched_backend_ops.timedout_job() for details). Do not call it for
664  * scheduler startup. The scheduler itself is fully operational after
665  * drm_sched_init() succeeded.
666  */
667 void drm_sched_start(struct drm_gpu_scheduler *sched, int errno)
668 {
669 	struct drm_sched_job *s_job, *tmp;
670 
671 	/*
672 	 * Locking the list is not required here as the scheduler's work items
673 	 * are currently not running, so no new jobs are being inserted or
674 	 * removed. Also concurrent GPU recovers can't run in parallel.
675 	 */
676 	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
677 		struct dma_fence *fence = s_job->s_fence->parent;
678 
679 		atomic_add(s_job->credits, &sched->credit_count);
680 
681 		if (!fence) {
682 			drm_sched_job_done(s_job, errno ?: -ECANCELED);
683 			continue;
684 		}
685 
686 		if (dma_fence_add_callback(fence, &s_job->cb,
687 					   drm_sched_job_done_cb))
688 			drm_sched_job_done(s_job, fence->error ?: errno);
689 	}
690 
691 	drm_sched_start_timeout_unlocked(sched);
692 	drm_sched_wqueue_start(sched);
693 }
694 EXPORT_SYMBOL(drm_sched_start);
695 
696 /**
697  * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
698  *
699  * @sched: scheduler instance
700  *
701  * Re-submitting jobs was a concept AMD came up as cheap way to implement
702  * recovery after a job timeout.
703  *
704  * This turned out to be not working very well. First of all there are many
705  * problem with the dma_fence implementation and requirements. Either the
706  * implementation is risking deadlocks with core memory management or violating
707  * documented implementation details of the dma_fence object.
708  *
709  * Drivers can still save and restore their state for recovery operations, but
710  * we shouldn't make this a general scheduler feature around the dma_fence
711  * interface.
712  */
713 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
714 {
715 	struct drm_sched_job *s_job, *tmp;
716 	uint64_t guilty_context;
717 	bool found_guilty = false;
718 	struct dma_fence *fence;
719 
720 	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
721 		struct drm_sched_fence *s_fence = s_job->s_fence;
722 
723 		if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
724 			found_guilty = true;
725 			guilty_context = s_job->s_fence->scheduled.context;
726 		}
727 
728 		if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
729 			dma_fence_set_error(&s_fence->finished, -ECANCELED);
730 
731 		fence = sched->ops->run_job(s_job);
732 
733 		if (IS_ERR_OR_NULL(fence)) {
734 			if (IS_ERR(fence))
735 				dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
736 
737 			s_job->s_fence->parent = NULL;
738 		} else {
739 
740 			s_job->s_fence->parent = dma_fence_get(fence);
741 
742 			/* Drop for orignal kref_init */
743 			dma_fence_put(fence);
744 		}
745 	}
746 }
747 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
748 
749 /**
750  * drm_sched_job_init - init a scheduler job
751  * @job: scheduler job to init
752  * @entity: scheduler entity to use
753  * @credits: the number of credits this job contributes to the schedulers
754  * credit limit
755  * @owner: job owner for debugging
756  * @drm_client_id: &struct drm_file.client_id of the owner (used by trace
757  * events)
758  *
759  * Refer to drm_sched_entity_push_job() documentation
760  * for locking considerations.
761  *
762  * Drivers must make sure drm_sched_job_cleanup() if this function returns
763  * successfully, even when @job is aborted before drm_sched_job_arm() is called.
764  *
765  * Note that this function does not assign a valid value to each struct member
766  * of struct drm_sched_job. Take a look at that struct's documentation to see
767  * who sets which struct member with what lifetime.
768  *
769  * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
770  * has died, which can mean that there's no valid runqueue for a @entity.
771  * This function returns -ENOENT in this case (which probably should be -EIO as
772  * a more meanigful return value).
773  *
774  * Returns 0 for success, negative error code otherwise.
775  */
776 int drm_sched_job_init(struct drm_sched_job *job,
777 		       struct drm_sched_entity *entity,
778 		       u32 credits, void *owner,
779 		       uint64_t drm_client_id)
780 {
781 	if (!entity->rq) {
782 		/* This will most likely be followed by missing frames
783 		 * or worse--a blank screen--leave a trail in the
784 		 * logs, so this can be debugged easier.
785 		 */
786 		dev_err(job->sched->dev, "%s: entity has no rq!\n", __func__);
787 		return -ENOENT;
788 	}
789 
790 	if (unlikely(!credits)) {
791 		pr_err("*ERROR* %s: credits cannot be 0!\n", __func__);
792 		return -EINVAL;
793 	}
794 
795 	/*
796 	 * We don't know for sure how the user has allocated. Thus, zero the
797 	 * struct so that unallowed (i.e., too early) usage of pointers that
798 	 * this function does not set is guaranteed to lead to a NULL pointer
799 	 * exception instead of UB.
800 	 */
801 	memset(job, 0, sizeof(*job));
802 
803 	job->entity = entity;
804 	job->credits = credits;
805 	job->s_fence = drm_sched_fence_alloc(entity, owner, drm_client_id);
806 	if (!job->s_fence)
807 		return -ENOMEM;
808 
809 	INIT_LIST_HEAD(&job->list);
810 
811 	xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
812 
813 	return 0;
814 }
815 EXPORT_SYMBOL(drm_sched_job_init);
816 
817 /**
818  * drm_sched_job_arm - arm a scheduler job for execution
819  * @job: scheduler job to arm
820  *
821  * This arms a scheduler job for execution. Specifically it initializes the
822  * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
823  * or other places that need to track the completion of this job. It also
824  * initializes sequence numbers, which are fundamental for fence ordering.
825  *
826  * Refer to drm_sched_entity_push_job() documentation for locking
827  * considerations.
828  *
829  * Once this function was called, you *must* submit @job with
830  * drm_sched_entity_push_job().
831  *
832  * This can only be called if drm_sched_job_init() succeeded.
833  */
834 void drm_sched_job_arm(struct drm_sched_job *job)
835 {
836 	struct drm_gpu_scheduler *sched;
837 	struct drm_sched_entity *entity = job->entity;
838 
839 	BUG_ON(!entity);
840 	drm_sched_entity_select_rq(entity);
841 	sched = entity->rq->sched;
842 
843 	job->sched = sched;
844 	job->s_priority = entity->priority;
845 
846 	drm_sched_fence_init(job->s_fence, job->entity);
847 }
848 EXPORT_SYMBOL(drm_sched_job_arm);
849 
850 /**
851  * drm_sched_job_add_dependency - adds the fence as a job dependency
852  * @job: scheduler job to add the dependencies to
853  * @fence: the dma_fence to add to the list of dependencies.
854  *
855  * Note that @fence is consumed in both the success and error cases.
856  *
857  * Returns:
858  * 0 on success, or an error on failing to expand the array.
859  */
860 int drm_sched_job_add_dependency(struct drm_sched_job *job,
861 				 struct dma_fence *fence)
862 {
863 	struct dma_fence *entry;
864 	unsigned long index;
865 	u32 id = 0;
866 	int ret;
867 
868 	if (!fence)
869 		return 0;
870 
871 	/* Deduplicate if we already depend on a fence from the same context.
872 	 * This lets the size of the array of deps scale with the number of
873 	 * engines involved, rather than the number of BOs.
874 	 */
875 	xa_for_each(&job->dependencies, index, entry) {
876 		if (entry->context != fence->context)
877 			continue;
878 
879 		if (dma_fence_is_later(fence, entry)) {
880 			dma_fence_put(entry);
881 			xa_store(&job->dependencies, index, fence, GFP_KERNEL);
882 		} else {
883 			dma_fence_put(fence);
884 		}
885 		return 0;
886 	}
887 
888 	ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
889 	if (ret != 0)
890 		dma_fence_put(fence);
891 
892 	return ret;
893 }
894 EXPORT_SYMBOL(drm_sched_job_add_dependency);
895 
896 /**
897  * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
898  * @job: scheduler job to add the dependencies to
899  * @file: drm file private pointer
900  * @handle: syncobj handle to lookup
901  * @point: timeline point
902  *
903  * This adds the fence matching the given syncobj to @job.
904  *
905  * Returns:
906  * 0 on success, or an error on failing to expand the array.
907  */
908 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
909 					 struct drm_file *file,
910 					 u32 handle,
911 					 u32 point)
912 {
913 	struct dma_fence *fence;
914 	int ret;
915 
916 	ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
917 	if (ret)
918 		return ret;
919 
920 	return drm_sched_job_add_dependency(job, fence);
921 }
922 EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
923 
924 /**
925  * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
926  * @job: scheduler job to add the dependencies to
927  * @resv: the dma_resv object to get the fences from
928  * @usage: the dma_resv_usage to use to filter the fences
929  *
930  * This adds all fences matching the given usage from @resv to @job.
931  * Must be called with the @resv lock held.
932  *
933  * Returns:
934  * 0 on success, or an error on failing to expand the array.
935  */
936 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
937 					struct dma_resv *resv,
938 					enum dma_resv_usage usage)
939 {
940 	struct dma_resv_iter cursor;
941 	struct dma_fence *fence;
942 	int ret;
943 
944 	dma_resv_assert_held(resv);
945 
946 	dma_resv_for_each_fence(&cursor, resv, usage, fence) {
947 		/* Make sure to grab an additional ref on the added fence */
948 		dma_fence_get(fence);
949 		ret = drm_sched_job_add_dependency(job, fence);
950 		if (ret) {
951 			dma_fence_put(fence);
952 			return ret;
953 		}
954 	}
955 	return 0;
956 }
957 EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
958 
959 /**
960  * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
961  *   dependencies
962  * @job: scheduler job to add the dependencies to
963  * @obj: the gem object to add new dependencies from.
964  * @write: whether the job might write the object (so we need to depend on
965  * shared fences in the reservation object).
966  *
967  * This should be called after drm_gem_lock_reservations() on your array of
968  * GEM objects used in the job but before updating the reservations with your
969  * own fences.
970  *
971  * Returns:
972  * 0 on success, or an error on failing to expand the array.
973  */
974 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
975 					    struct drm_gem_object *obj,
976 					    bool write)
977 {
978 	return drm_sched_job_add_resv_dependencies(job, obj->resv,
979 						   dma_resv_usage_rw(write));
980 }
981 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
982 
983 /**
984  * drm_sched_job_has_dependency - check whether fence is the job's dependency
985  * @job: scheduler job to check
986  * @fence: fence to look for
987  *
988  * Returns:
989  * True if @fence is found within the job's dependencies, or otherwise false.
990  */
991 bool drm_sched_job_has_dependency(struct drm_sched_job *job,
992 				  struct dma_fence *fence)
993 {
994 	struct dma_fence *f;
995 	unsigned long index;
996 
997 	xa_for_each(&job->dependencies, index, f) {
998 		if (f == fence)
999 			return true;
1000 	}
1001 
1002 	return false;
1003 }
1004 EXPORT_SYMBOL(drm_sched_job_has_dependency);
1005 
1006 /**
1007  * drm_sched_job_cleanup - clean up scheduler job resources
1008  * @job: scheduler job to clean up
1009  *
1010  * Cleans up the resources allocated with drm_sched_job_init().
1011  *
1012  * Drivers should call this from their error unwind code if @job is aborted
1013  * before drm_sched_job_arm() is called.
1014  *
1015  * drm_sched_job_arm() is a point of no return since it initializes the fences
1016  * and their sequence number etc. Once that function has been called, you *must*
1017  * submit it with drm_sched_entity_push_job() and cannot simply abort it by
1018  * calling drm_sched_job_cleanup().
1019  *
1020  * This function should be called in the &drm_sched_backend_ops.free_job callback.
1021  */
1022 void drm_sched_job_cleanup(struct drm_sched_job *job)
1023 {
1024 	struct dma_fence *fence;
1025 	unsigned long index;
1026 
1027 	if (kref_read(&job->s_fence->finished.refcount)) {
1028 		/* The job has been processed by the scheduler, i.e.,
1029 		 * drm_sched_job_arm() and drm_sched_entity_push_job() have
1030 		 * been called.
1031 		 */
1032 		dma_fence_put(&job->s_fence->finished);
1033 	} else {
1034 		/* The job was aborted before it has been committed to be run;
1035 		 * notably, drm_sched_job_arm() has not been called.
1036 		 */
1037 		drm_sched_fence_free(job->s_fence);
1038 	}
1039 
1040 	job->s_fence = NULL;
1041 
1042 	xa_for_each(&job->dependencies, index, fence) {
1043 		dma_fence_put(fence);
1044 	}
1045 	xa_destroy(&job->dependencies);
1046 
1047 }
1048 EXPORT_SYMBOL(drm_sched_job_cleanup);
1049 
1050 /**
1051  * drm_sched_wakeup - Wake up the scheduler if it is ready to queue
1052  * @sched: scheduler instance
1053  *
1054  * Wake up the scheduler if we can queue jobs.
1055  */
1056 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
1057 {
1058 	drm_sched_run_job_queue(sched);
1059 }
1060 
1061 /**
1062  * drm_sched_select_entity - Select next entity to process
1063  *
1064  * @sched: scheduler instance
1065  *
1066  * Return an entity to process or NULL if none are found.
1067  *
1068  * Note, that we break out of the for-loop when "entity" is non-null, which can
1069  * also be an error-pointer--this assures we don't process lower priority
1070  * run-queues. See comments in the respectively called functions.
1071  */
1072 static struct drm_sched_entity *
1073 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
1074 {
1075 	struct drm_sched_entity *entity;
1076 	int i;
1077 
1078 	/* Start with the highest priority.
1079 	 */
1080 	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1081 		entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
1082 			drm_sched_rq_select_entity_fifo(sched, sched->sched_rq[i]) :
1083 			drm_sched_rq_select_entity_rr(sched, sched->sched_rq[i]);
1084 		if (entity)
1085 			break;
1086 	}
1087 
1088 	return IS_ERR(entity) ? NULL : entity;
1089 }
1090 
1091 /**
1092  * drm_sched_get_finished_job - fetch the next finished job to be destroyed
1093  *
1094  * @sched: scheduler instance
1095  *
1096  * Returns the next finished job from the pending list (if there is one)
1097  * ready for it to be destroyed.
1098  */
1099 static struct drm_sched_job *
1100 drm_sched_get_finished_job(struct drm_gpu_scheduler *sched)
1101 {
1102 	struct drm_sched_job *job, *next;
1103 
1104 	spin_lock(&sched->job_list_lock);
1105 
1106 	job = list_first_entry_or_null(&sched->pending_list,
1107 				       struct drm_sched_job, list);
1108 
1109 	if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
1110 		/* remove job from pending_list */
1111 		list_del_init(&job->list);
1112 
1113 		/* cancel this job's TO timer */
1114 		cancel_delayed_work(&sched->work_tdr);
1115 		/* make the scheduled timestamp more accurate */
1116 		next = list_first_entry_or_null(&sched->pending_list,
1117 						typeof(*next), list);
1118 
1119 		if (next) {
1120 			if (test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT,
1121 				     &next->s_fence->scheduled.flags))
1122 				next->s_fence->scheduled.timestamp =
1123 					dma_fence_timestamp(&job->s_fence->finished);
1124 			/* start TO timer for next job */
1125 			drm_sched_start_timeout(sched);
1126 		}
1127 	} else {
1128 		job = NULL;
1129 	}
1130 
1131 	spin_unlock(&sched->job_list_lock);
1132 
1133 	return job;
1134 }
1135 
1136 /**
1137  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
1138  * @sched_list: list of drm_gpu_schedulers
1139  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
1140  *
1141  * Returns pointer of the sched with the least load or NULL if none of the
1142  * drm_gpu_schedulers are ready
1143  */
1144 struct drm_gpu_scheduler *
1145 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
1146 		     unsigned int num_sched_list)
1147 {
1148 	struct drm_gpu_scheduler *sched, *picked_sched = NULL;
1149 	int i;
1150 	unsigned int min_score = UINT_MAX, num_score;
1151 
1152 	for (i = 0; i < num_sched_list; ++i) {
1153 		sched = sched_list[i];
1154 
1155 		if (!sched->ready) {
1156 			DRM_WARN("scheduler %s is not ready, skipping",
1157 				 sched->name);
1158 			continue;
1159 		}
1160 
1161 		num_score = atomic_read(sched->score);
1162 		if (num_score < min_score) {
1163 			min_score = num_score;
1164 			picked_sched = sched;
1165 		}
1166 	}
1167 
1168 	return picked_sched;
1169 }
1170 EXPORT_SYMBOL(drm_sched_pick_best);
1171 
1172 /**
1173  * drm_sched_free_job_work - worker to call free_job
1174  *
1175  * @w: free job work
1176  */
1177 static void drm_sched_free_job_work(struct work_struct *w)
1178 {
1179 	struct drm_gpu_scheduler *sched =
1180 		container_of(w, struct drm_gpu_scheduler, work_free_job);
1181 	struct drm_sched_job *job;
1182 
1183 	job = drm_sched_get_finished_job(sched);
1184 	if (job)
1185 		sched->ops->free_job(job);
1186 
1187 	drm_sched_run_free_queue(sched);
1188 	drm_sched_run_job_queue(sched);
1189 }
1190 
1191 /**
1192  * drm_sched_run_job_work - worker to call run_job
1193  *
1194  * @w: run job work
1195  */
1196 static void drm_sched_run_job_work(struct work_struct *w)
1197 {
1198 	struct drm_gpu_scheduler *sched =
1199 		container_of(w, struct drm_gpu_scheduler, work_run_job);
1200 	struct drm_sched_entity *entity;
1201 	struct dma_fence *fence;
1202 	struct drm_sched_fence *s_fence;
1203 	struct drm_sched_job *sched_job;
1204 	int r;
1205 
1206 	/* Find entity with a ready job */
1207 	entity = drm_sched_select_entity(sched);
1208 	if (!entity)
1209 		return;	/* No more work */
1210 
1211 	sched_job = drm_sched_entity_pop_job(entity);
1212 	if (!sched_job) {
1213 		complete_all(&entity->entity_idle);
1214 		drm_sched_run_job_queue(sched);
1215 		return;
1216 	}
1217 
1218 	s_fence = sched_job->s_fence;
1219 
1220 	atomic_add(sched_job->credits, &sched->credit_count);
1221 	drm_sched_job_begin(sched_job);
1222 
1223 	trace_drm_sched_job_run(sched_job, entity);
1224 	/*
1225 	 * The run_job() callback must by definition return a fence whose
1226 	 * refcount has been incremented for the scheduler already.
1227 	 */
1228 	fence = sched->ops->run_job(sched_job);
1229 	complete_all(&entity->entity_idle);
1230 	drm_sched_fence_scheduled(s_fence, fence);
1231 
1232 	if (!IS_ERR_OR_NULL(fence)) {
1233 		r = dma_fence_add_callback(fence, &sched_job->cb,
1234 					   drm_sched_job_done_cb);
1235 		if (r == -ENOENT)
1236 			drm_sched_job_done(sched_job, fence->error);
1237 		else if (r)
1238 			DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
1239 
1240 		dma_fence_put(fence);
1241 	} else {
1242 		drm_sched_job_done(sched_job, IS_ERR(fence) ?
1243 				   PTR_ERR(fence) : 0);
1244 	}
1245 
1246 	wake_up(&sched->job_scheduled);
1247 	drm_sched_run_job_queue(sched);
1248 }
1249 
1250 static struct workqueue_struct *drm_sched_alloc_wq(const char *name)
1251 {
1252 #if (IS_ENABLED(CONFIG_LOCKDEP))
1253 	static struct lockdep_map map = {
1254 		.name = "drm_sched_lockdep_map"
1255 	};
1256 
1257 	/*
1258 	 * Avoid leaking a lockdep map on each drm sched creation and
1259 	 * destruction by using a single lockdep map for all drm sched
1260 	 * allocated submit_wq.
1261 	 */
1262 
1263 	return alloc_ordered_workqueue_lockdep_map(name, WQ_MEM_RECLAIM, &map);
1264 #else
1265 	return alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
1266 #endif
1267 }
1268 
1269 /**
1270  * drm_sched_init - Init a gpu scheduler instance
1271  *
1272  * @sched: scheduler instance
1273  * @args: scheduler initialization arguments
1274  *
1275  * Return 0 on success, otherwise error code.
1276  */
1277 int drm_sched_init(struct drm_gpu_scheduler *sched, const struct drm_sched_init_args *args)
1278 {
1279 	int i;
1280 
1281 	sched->ops = args->ops;
1282 	sched->credit_limit = args->credit_limit;
1283 	sched->name = args->name;
1284 	sched->timeout = args->timeout;
1285 	sched->hang_limit = args->hang_limit;
1286 	sched->timeout_wq = args->timeout_wq ? args->timeout_wq : system_wq;
1287 	sched->score = args->score ? args->score : &sched->_score;
1288 	sched->dev = args->dev;
1289 
1290 	if (args->num_rqs > DRM_SCHED_PRIORITY_COUNT) {
1291 		/* This is a gross violation--tell drivers what the  problem is.
1292 		 */
1293 		dev_err(sched->dev, "%s: num_rqs cannot be greater than DRM_SCHED_PRIORITY_COUNT\n",
1294 			__func__);
1295 		return -EINVAL;
1296 	} else if (sched->sched_rq) {
1297 		/* Not an error, but warn anyway so drivers can
1298 		 * fine-tune their DRM calling order, and return all
1299 		 * is good.
1300 		 */
1301 		dev_warn(sched->dev, "%s: scheduler already initialized!\n", __func__);
1302 		return 0;
1303 	}
1304 
1305 	if (args->submit_wq) {
1306 		sched->submit_wq = args->submit_wq;
1307 		sched->own_submit_wq = false;
1308 	} else {
1309 		sched->submit_wq = drm_sched_alloc_wq(args->name);
1310 		if (!sched->submit_wq)
1311 			return -ENOMEM;
1312 
1313 		sched->own_submit_wq = true;
1314 	}
1315 
1316 	sched->sched_rq = kmalloc_array(args->num_rqs, sizeof(*sched->sched_rq),
1317 					GFP_KERNEL | __GFP_ZERO);
1318 	if (!sched->sched_rq)
1319 		goto Out_check_own;
1320 	sched->num_rqs = args->num_rqs;
1321 	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1322 		sched->sched_rq[i] = kzalloc(sizeof(*sched->sched_rq[i]), GFP_KERNEL);
1323 		if (!sched->sched_rq[i])
1324 			goto Out_unroll;
1325 		drm_sched_rq_init(sched, sched->sched_rq[i]);
1326 	}
1327 
1328 	init_waitqueue_head(&sched->job_scheduled);
1329 	INIT_LIST_HEAD(&sched->pending_list);
1330 	spin_lock_init(&sched->job_list_lock);
1331 	atomic_set(&sched->credit_count, 0);
1332 	INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1333 	INIT_WORK(&sched->work_run_job, drm_sched_run_job_work);
1334 	INIT_WORK(&sched->work_free_job, drm_sched_free_job_work);
1335 	atomic_set(&sched->_score, 0);
1336 	atomic64_set(&sched->job_id_count, 0);
1337 	sched->pause_submit = false;
1338 
1339 	sched->ready = true;
1340 	return 0;
1341 Out_unroll:
1342 	for (--i ; i >= DRM_SCHED_PRIORITY_KERNEL; i--)
1343 		kfree(sched->sched_rq[i]);
1344 
1345 	kfree(sched->sched_rq);
1346 	sched->sched_rq = NULL;
1347 Out_check_own:
1348 	if (sched->own_submit_wq)
1349 		destroy_workqueue(sched->submit_wq);
1350 	dev_err(sched->dev, "%s: Failed to setup GPU scheduler--out of memory\n", __func__);
1351 	return -ENOMEM;
1352 }
1353 EXPORT_SYMBOL(drm_sched_init);
1354 
1355 /**
1356  * drm_sched_fini - Destroy a gpu scheduler
1357  *
1358  * @sched: scheduler instance
1359  *
1360  * Tears down and cleans up the scheduler.
1361  *
1362  * This stops submission of new jobs to the hardware through
1363  * drm_sched_backend_ops.run_job(). Consequently, drm_sched_backend_ops.free_job()
1364  * will not be called for all jobs still in drm_gpu_scheduler.pending_list.
1365  * There is no solution for this currently. Thus, it is up to the driver to make
1366  * sure that:
1367  *
1368  *  a) drm_sched_fini() is only called after for all submitted jobs
1369  *     drm_sched_backend_ops.free_job() has been called or that
1370  *  b) the jobs for which drm_sched_backend_ops.free_job() has not been called
1371  *     after drm_sched_fini() ran are freed manually.
1372  *
1373  * FIXME: Take care of the above problem and prevent this function from leaking
1374  * the jobs in drm_gpu_scheduler.pending_list under any circumstances.
1375  */
1376 void drm_sched_fini(struct drm_gpu_scheduler *sched)
1377 {
1378 	struct drm_sched_entity *s_entity;
1379 	int i;
1380 
1381 	drm_sched_wqueue_stop(sched);
1382 
1383 	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1384 		struct drm_sched_rq *rq = sched->sched_rq[i];
1385 
1386 		spin_lock(&rq->lock);
1387 		list_for_each_entry(s_entity, &rq->entities, list)
1388 			/*
1389 			 * Prevents reinsertion and marks job_queue as idle,
1390 			 * it will be removed from the rq in drm_sched_entity_fini()
1391 			 * eventually
1392 			 */
1393 			s_entity->stopped = true;
1394 		spin_unlock(&rq->lock);
1395 		kfree(sched->sched_rq[i]);
1396 	}
1397 
1398 	/* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1399 	wake_up_all(&sched->job_scheduled);
1400 
1401 	/* Confirm no work left behind accessing device structures */
1402 	cancel_delayed_work_sync(&sched->work_tdr);
1403 
1404 	if (sched->own_submit_wq)
1405 		destroy_workqueue(sched->submit_wq);
1406 	sched->ready = false;
1407 	kfree(sched->sched_rq);
1408 	sched->sched_rq = NULL;
1409 }
1410 EXPORT_SYMBOL(drm_sched_fini);
1411 
1412 /**
1413  * drm_sched_increase_karma - Update sched_entity guilty flag
1414  *
1415  * @bad: The job guilty of time out
1416  *
1417  * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1418  * limit of the scheduler then the respective sched entity is marked guilty and
1419  * jobs from it will not be scheduled further
1420  */
1421 void drm_sched_increase_karma(struct drm_sched_job *bad)
1422 {
1423 	int i;
1424 	struct drm_sched_entity *tmp;
1425 	struct drm_sched_entity *entity;
1426 	struct drm_gpu_scheduler *sched = bad->sched;
1427 
1428 	/* don't change @bad's karma if it's from KERNEL RQ,
1429 	 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1430 	 * corrupt but keep in mind that kernel jobs always considered good.
1431 	 */
1432 	if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1433 		atomic_inc(&bad->karma);
1434 
1435 		for (i = DRM_SCHED_PRIORITY_HIGH; i < sched->num_rqs; i++) {
1436 			struct drm_sched_rq *rq = sched->sched_rq[i];
1437 
1438 			spin_lock(&rq->lock);
1439 			list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1440 				if (bad->s_fence->scheduled.context ==
1441 				    entity->fence_context) {
1442 					if (entity->guilty)
1443 						atomic_set(entity->guilty, 1);
1444 					break;
1445 				}
1446 			}
1447 			spin_unlock(&rq->lock);
1448 			if (&entity->list != &rq->entities)
1449 				break;
1450 		}
1451 	}
1452 }
1453 EXPORT_SYMBOL(drm_sched_increase_karma);
1454 
1455 /**
1456  * drm_sched_wqueue_ready - Is the scheduler ready for submission
1457  *
1458  * @sched: scheduler instance
1459  *
1460  * Returns true if submission is ready
1461  */
1462 bool drm_sched_wqueue_ready(struct drm_gpu_scheduler *sched)
1463 {
1464 	return sched->ready;
1465 }
1466 EXPORT_SYMBOL(drm_sched_wqueue_ready);
1467 
1468 /**
1469  * drm_sched_wqueue_stop - stop scheduler submission
1470  * @sched: scheduler instance
1471  *
1472  * Stops the scheduler from pulling new jobs from entities. It also stops
1473  * freeing jobs automatically through drm_sched_backend_ops.free_job().
1474  */
1475 void drm_sched_wqueue_stop(struct drm_gpu_scheduler *sched)
1476 {
1477 	WRITE_ONCE(sched->pause_submit, true);
1478 	cancel_work_sync(&sched->work_run_job);
1479 	cancel_work_sync(&sched->work_free_job);
1480 }
1481 EXPORT_SYMBOL(drm_sched_wqueue_stop);
1482 
1483 /**
1484  * drm_sched_wqueue_start - start scheduler submission
1485  * @sched: scheduler instance
1486  *
1487  * Restarts the scheduler after drm_sched_wqueue_stop() has stopped it.
1488  *
1489  * This function is not necessary for 'conventional' startup. The scheduler is
1490  * fully operational after drm_sched_init() succeeded.
1491  */
1492 void drm_sched_wqueue_start(struct drm_gpu_scheduler *sched)
1493 {
1494 	WRITE_ONCE(sched->pause_submit, false);
1495 	queue_work(sched->submit_wq, &sched->work_run_job);
1496 	queue_work(sched->submit_wq, &sched->work_free_job);
1497 }
1498 EXPORT_SYMBOL(drm_sched_wqueue_start);
1499