xref: /linux/drivers/gpu/drm/scheduler/sched_main.c (revision 7b1166dee847d5018c1f3cc781218e806078f752)
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/wait.h>
70 #include <linux/sched.h>
71 #include <linux/completion.h>
72 #include <linux/dma-resv.h>
73 #include <uapi/linux/sched/types.h>
74 
75 #include <drm/drm_print.h>
76 #include <drm/drm_gem.h>
77 #include <drm/drm_syncobj.h>
78 #include <drm/gpu_scheduler.h>
79 #include <drm/spsc_queue.h>
80 
81 #include "sched_internal.h"
82 
83 #define CREATE_TRACE_POINTS
84 #include "gpu_scheduler_trace.h"
85 
86 #ifdef CONFIG_LOCKDEP
87 static struct lockdep_map drm_sched_lockdep_map = {
88 	.name = "drm_sched_lockdep_map"
89 };
90 #endif
91 
92 int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
93 
94 /**
95  * DOC: sched_policy (int)
96  * Used to override default entities scheduling policy in a run queue.
97  */
98 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).");
99 module_param_named(sched_policy, drm_sched_policy, int, 0444);
100 
101 static u32 drm_sched_available_credits(struct drm_gpu_scheduler *sched)
102 {
103 	u32 credits;
104 
105 	WARN_ON(check_sub_overflow(sched->credit_limit,
106 				   atomic_read(&sched->credit_count),
107 				   &credits));
108 
109 	return credits;
110 }
111 
112 /**
113  * drm_sched_can_queue -- Can we queue more to the hardware?
114  * @sched: scheduler instance
115  * @entity: the scheduler entity
116  *
117  * Return true if we can push at least one more job from @entity, false
118  * otherwise.
119  */
120 static bool drm_sched_can_queue(struct drm_gpu_scheduler *sched,
121 				struct drm_sched_entity *entity)
122 {
123 	struct drm_sched_job *s_job;
124 
125 	s_job = drm_sched_entity_queue_peek(entity);
126 	if (!s_job)
127 		return false;
128 
129 	/* If a job exceeds the credit limit, truncate it to the credit limit
130 	 * itself to guarantee forward progress.
131 	 */
132 	if (s_job->credits > sched->credit_limit) {
133 		dev_WARN(sched->dev,
134 			 "Jobs may not exceed the credit limit, truncate.\n");
135 		s_job->credits = sched->credit_limit;
136 	}
137 
138 	return drm_sched_available_credits(sched) >= s_job->credits;
139 }
140 
141 static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
142 							    const struct rb_node *b)
143 {
144 	struct drm_sched_entity *ent_a =  rb_entry((a), struct drm_sched_entity, rb_tree_node);
145 	struct drm_sched_entity *ent_b =  rb_entry((b), struct drm_sched_entity, rb_tree_node);
146 
147 	return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
148 }
149 
150 static void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity,
151 					    struct drm_sched_rq *rq)
152 {
153 	if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
154 		rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
155 		RB_CLEAR_NODE(&entity->rb_tree_node);
156 	}
157 }
158 
159 void drm_sched_rq_update_fifo_locked(struct drm_sched_entity *entity,
160 				     struct drm_sched_rq *rq,
161 				     ktime_t ts)
162 {
163 	/*
164 	 * Both locks need to be grabbed, one to protect from entity->rq change
165 	 * for entity from within concurrent drm_sched_entity_select_rq and the
166 	 * other to update the rb tree structure.
167 	 */
168 	lockdep_assert_held(&entity->lock);
169 	lockdep_assert_held(&rq->lock);
170 
171 	drm_sched_rq_remove_fifo_locked(entity, rq);
172 
173 	entity->oldest_job_waiting = ts;
174 
175 	rb_add_cached(&entity->rb_tree_node, &rq->rb_tree_root,
176 		      drm_sched_entity_compare_before);
177 }
178 
179 /**
180  * drm_sched_rq_init - initialize a given run queue struct
181  *
182  * @sched: scheduler instance to associate with this run queue
183  * @rq: scheduler run queue
184  *
185  * Initializes a scheduler runqueue.
186  */
187 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
188 			      struct drm_sched_rq *rq)
189 {
190 	spin_lock_init(&rq->lock);
191 	INIT_LIST_HEAD(&rq->entities);
192 	rq->rb_tree_root = RB_ROOT_CACHED;
193 	rq->current_entity = NULL;
194 	rq->sched = sched;
195 }
196 
197 /**
198  * drm_sched_rq_add_entity - add an entity
199  *
200  * @rq: scheduler run queue
201  * @entity: scheduler entity
202  *
203  * Adds a scheduler entity to the run queue.
204  */
205 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
206 			     struct drm_sched_entity *entity)
207 {
208 	lockdep_assert_held(&entity->lock);
209 	lockdep_assert_held(&rq->lock);
210 
211 	if (!list_empty(&entity->list))
212 		return;
213 
214 	atomic_inc(rq->sched->score);
215 	list_add_tail(&entity->list, &rq->entities);
216 }
217 
218 /**
219  * drm_sched_rq_remove_entity - remove an entity
220  *
221  * @rq: scheduler run queue
222  * @entity: scheduler entity
223  *
224  * Removes a scheduler entity from the run queue.
225  */
226 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
227 				struct drm_sched_entity *entity)
228 {
229 	lockdep_assert_held(&entity->lock);
230 
231 	if (list_empty(&entity->list))
232 		return;
233 
234 	spin_lock(&rq->lock);
235 
236 	atomic_dec(rq->sched->score);
237 	list_del_init(&entity->list);
238 
239 	if (rq->current_entity == entity)
240 		rq->current_entity = NULL;
241 
242 	if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
243 		drm_sched_rq_remove_fifo_locked(entity, rq);
244 
245 	spin_unlock(&rq->lock);
246 }
247 
248 /**
249  * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
250  *
251  * @sched: the gpu scheduler
252  * @rq: scheduler run queue to check.
253  *
254  * Try to find the next ready entity.
255  *
256  * Return an entity if one is found; return an error-pointer (!NULL) if an
257  * entity was ready, but the scheduler had insufficient credits to accommodate
258  * its job; return NULL, if no ready entity was found.
259  */
260 static struct drm_sched_entity *
261 drm_sched_rq_select_entity_rr(struct drm_gpu_scheduler *sched,
262 			      struct drm_sched_rq *rq)
263 {
264 	struct drm_sched_entity *entity;
265 
266 	spin_lock(&rq->lock);
267 
268 	entity = rq->current_entity;
269 	if (entity) {
270 		list_for_each_entry_continue(entity, &rq->entities, list) {
271 			if (drm_sched_entity_is_ready(entity)) {
272 				/* If we can't queue yet, preserve the current
273 				 * entity in terms of fairness.
274 				 */
275 				if (!drm_sched_can_queue(sched, entity)) {
276 					spin_unlock(&rq->lock);
277 					return ERR_PTR(-ENOSPC);
278 				}
279 
280 				rq->current_entity = entity;
281 				reinit_completion(&entity->entity_idle);
282 				spin_unlock(&rq->lock);
283 				return entity;
284 			}
285 		}
286 	}
287 
288 	list_for_each_entry(entity, &rq->entities, list) {
289 		if (drm_sched_entity_is_ready(entity)) {
290 			/* If we can't queue yet, preserve the current entity in
291 			 * terms of fairness.
292 			 */
293 			if (!drm_sched_can_queue(sched, entity)) {
294 				spin_unlock(&rq->lock);
295 				return ERR_PTR(-ENOSPC);
296 			}
297 
298 			rq->current_entity = entity;
299 			reinit_completion(&entity->entity_idle);
300 			spin_unlock(&rq->lock);
301 			return entity;
302 		}
303 
304 		if (entity == rq->current_entity)
305 			break;
306 	}
307 
308 	spin_unlock(&rq->lock);
309 
310 	return NULL;
311 }
312 
313 /**
314  * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
315  *
316  * @sched: the gpu scheduler
317  * @rq: scheduler run queue to check.
318  *
319  * Find oldest waiting ready entity.
320  *
321  * Return an entity if one is found; return an error-pointer (!NULL) if an
322  * entity was ready, but the scheduler had insufficient credits to accommodate
323  * its job; return NULL, if no ready entity was found.
324  */
325 static struct drm_sched_entity *
326 drm_sched_rq_select_entity_fifo(struct drm_gpu_scheduler *sched,
327 				struct drm_sched_rq *rq)
328 {
329 	struct rb_node *rb;
330 
331 	spin_lock(&rq->lock);
332 	for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
333 		struct drm_sched_entity *entity;
334 
335 		entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
336 		if (drm_sched_entity_is_ready(entity)) {
337 			/* If we can't queue yet, preserve the current entity in
338 			 * terms of fairness.
339 			 */
340 			if (!drm_sched_can_queue(sched, entity)) {
341 				spin_unlock(&rq->lock);
342 				return ERR_PTR(-ENOSPC);
343 			}
344 
345 			reinit_completion(&entity->entity_idle);
346 			break;
347 		}
348 	}
349 	spin_unlock(&rq->lock);
350 
351 	return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
352 }
353 
354 /**
355  * drm_sched_run_job_queue - enqueue run-job work
356  * @sched: scheduler instance
357  */
358 static void drm_sched_run_job_queue(struct drm_gpu_scheduler *sched)
359 {
360 	if (!READ_ONCE(sched->pause_submit))
361 		queue_work(sched->submit_wq, &sched->work_run_job);
362 }
363 
364 /**
365  * __drm_sched_run_free_queue - enqueue free-job work
366  * @sched: scheduler instance
367  */
368 static void __drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
369 {
370 	if (!READ_ONCE(sched->pause_submit))
371 		queue_work(sched->submit_wq, &sched->work_free_job);
372 }
373 
374 /**
375  * drm_sched_run_free_queue - enqueue free-job work if ready
376  * @sched: scheduler instance
377  */
378 static void drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
379 {
380 	struct drm_sched_job *job;
381 
382 	spin_lock(&sched->job_list_lock);
383 	job = list_first_entry_or_null(&sched->pending_list,
384 				       struct drm_sched_job, list);
385 	if (job && dma_fence_is_signaled(&job->s_fence->finished))
386 		__drm_sched_run_free_queue(sched);
387 	spin_unlock(&sched->job_list_lock);
388 }
389 
390 /**
391  * drm_sched_job_done - complete a job
392  * @s_job: pointer to the job which is done
393  *
394  * Finish the job's fence and resubmit the work items.
395  */
396 static void drm_sched_job_done(struct drm_sched_job *s_job, int result)
397 {
398 	struct drm_sched_fence *s_fence = s_job->s_fence;
399 	struct drm_gpu_scheduler *sched = s_fence->sched;
400 
401 	atomic_sub(s_job->credits, &sched->credit_count);
402 	atomic_dec(sched->score);
403 
404 	trace_drm_sched_job_done(s_fence);
405 
406 	dma_fence_get(&s_fence->finished);
407 	drm_sched_fence_finished(s_fence, result);
408 	dma_fence_put(&s_fence->finished);
409 	__drm_sched_run_free_queue(sched);
410 }
411 
412 /**
413  * drm_sched_job_done_cb - the callback for a done job
414  * @f: fence
415  * @cb: fence callbacks
416  */
417 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
418 {
419 	struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
420 
421 	drm_sched_job_done(s_job, f->error);
422 }
423 
424 /**
425  * drm_sched_start_timeout - start timeout for reset worker
426  *
427  * @sched: scheduler instance to start the worker for
428  *
429  * Start the timeout for the given scheduler.
430  */
431 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
432 {
433 	lockdep_assert_held(&sched->job_list_lock);
434 
435 	if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
436 	    !list_empty(&sched->pending_list))
437 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
438 }
439 
440 static void drm_sched_start_timeout_unlocked(struct drm_gpu_scheduler *sched)
441 {
442 	spin_lock(&sched->job_list_lock);
443 	drm_sched_start_timeout(sched);
444 	spin_unlock(&sched->job_list_lock);
445 }
446 
447 /**
448  * drm_sched_tdr_queue_imm: - immediately start job timeout handler
449  *
450  * @sched: scheduler for which the timeout handling should be started.
451  *
452  * Start timeout handling immediately for the named scheduler.
453  */
454 void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched)
455 {
456 	spin_lock(&sched->job_list_lock);
457 	sched->timeout = 0;
458 	drm_sched_start_timeout(sched);
459 	spin_unlock(&sched->job_list_lock);
460 }
461 EXPORT_SYMBOL(drm_sched_tdr_queue_imm);
462 
463 /**
464  * drm_sched_fault - immediately start timeout handler
465  *
466  * @sched: scheduler where the timeout handling should be started.
467  *
468  * Start timeout handling immediately when the driver detects a hardware fault.
469  */
470 void drm_sched_fault(struct drm_gpu_scheduler *sched)
471 {
472 	if (sched->timeout_wq)
473 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
474 }
475 EXPORT_SYMBOL(drm_sched_fault);
476 
477 /**
478  * drm_sched_suspend_timeout - Suspend scheduler job timeout
479  *
480  * @sched: scheduler instance for which to suspend the timeout
481  *
482  * Suspend the delayed work timeout for the scheduler. This is done by
483  * modifying the delayed work timeout to an arbitrary large value,
484  * MAX_SCHEDULE_TIMEOUT in this case.
485  *
486  * Returns the timeout remaining
487  *
488  */
489 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
490 {
491 	unsigned long sched_timeout, now = jiffies;
492 
493 	sched_timeout = sched->work_tdr.timer.expires;
494 
495 	/*
496 	 * Modify the timeout to an arbitrarily large value. This also prevents
497 	 * the timeout to be restarted when new submissions arrive
498 	 */
499 	if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
500 			&& time_after(sched_timeout, now))
501 		return sched_timeout - now;
502 	else
503 		return sched->timeout;
504 }
505 EXPORT_SYMBOL(drm_sched_suspend_timeout);
506 
507 /**
508  * drm_sched_resume_timeout - Resume scheduler job timeout
509  *
510  * @sched: scheduler instance for which to resume the timeout
511  * @remaining: remaining timeout
512  *
513  * Resume the delayed work timeout for the scheduler.
514  */
515 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
516 		unsigned long remaining)
517 {
518 	spin_lock(&sched->job_list_lock);
519 
520 	if (list_empty(&sched->pending_list))
521 		cancel_delayed_work(&sched->work_tdr);
522 	else
523 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
524 
525 	spin_unlock(&sched->job_list_lock);
526 }
527 EXPORT_SYMBOL(drm_sched_resume_timeout);
528 
529 static void drm_sched_job_begin(struct drm_sched_job *s_job)
530 {
531 	struct drm_gpu_scheduler *sched = s_job->sched;
532 
533 	spin_lock(&sched->job_list_lock);
534 	list_add_tail(&s_job->list, &sched->pending_list);
535 	drm_sched_start_timeout(sched);
536 	spin_unlock(&sched->job_list_lock);
537 }
538 
539 static void drm_sched_job_timedout(struct work_struct *work)
540 {
541 	struct drm_gpu_scheduler *sched;
542 	struct drm_sched_job *job;
543 	enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
544 
545 	sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
546 
547 	/* Protects against concurrent deletion in drm_sched_get_finished_job */
548 	spin_lock(&sched->job_list_lock);
549 	job = list_first_entry_or_null(&sched->pending_list,
550 				       struct drm_sched_job, list);
551 
552 	if (job) {
553 		/*
554 		 * Remove the bad job so it cannot be freed by a concurrent
555 		 * &struct drm_sched_backend_ops.free_job. It will be
556 		 * reinserted after the scheduler's work items have been
557 		 * cancelled, at which point it's safe.
558 		 */
559 		list_del_init(&job->list);
560 		spin_unlock(&sched->job_list_lock);
561 
562 		status = job->sched->ops->timedout_job(job);
563 
564 		/*
565 		 * Guilty job did complete and hence needs to be manually removed
566 		 * See drm_sched_stop doc.
567 		 */
568 		if (sched->free_guilty) {
569 			job->sched->ops->free_job(job);
570 			sched->free_guilty = false;
571 		}
572 	} else {
573 		spin_unlock(&sched->job_list_lock);
574 	}
575 
576 	if (status != DRM_GPU_SCHED_STAT_ENODEV)
577 		drm_sched_start_timeout_unlocked(sched);
578 }
579 
580 /**
581  * drm_sched_stop - stop the scheduler
582  *
583  * @sched: scheduler instance
584  * @bad: job which caused the time out
585  *
586  * Stop the scheduler and also removes and frees all completed jobs.
587  * Note: bad job will not be freed as it might be used later and so it's
588  * callers responsibility to release it manually if it's not part of the
589  * pending list any more.
590  *
591  * This function is typically used for reset recovery (see the docu of
592  * drm_sched_backend_ops.timedout_job() for details). Do not call it for
593  * scheduler teardown, i.e., before calling drm_sched_fini().
594  */
595 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
596 {
597 	struct drm_sched_job *s_job, *tmp;
598 
599 	drm_sched_wqueue_stop(sched);
600 
601 	/*
602 	 * Reinsert back the bad job here - now it's safe as
603 	 * drm_sched_get_finished_job() cannot race against us and release the
604 	 * bad job at this point - we parked (waited for) any in progress
605 	 * (earlier) cleanups and drm_sched_get_finished_job() will not be
606 	 * called now until the scheduler's work items are submitted again.
607 	 */
608 	if (bad && bad->sched == sched)
609 		/*
610 		 * Add at the head of the queue to reflect it was the earliest
611 		 * job extracted.
612 		 */
613 		list_add(&bad->list, &sched->pending_list);
614 
615 	/*
616 	 * Iterate the job list from later to  earlier one and either deactive
617 	 * their HW callbacks or remove them from pending list if they already
618 	 * signaled.
619 	 * This iteration is thread safe as the scheduler's work items have been
620 	 * cancelled.
621 	 */
622 	list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
623 					 list) {
624 		if (s_job->s_fence->parent &&
625 		    dma_fence_remove_callback(s_job->s_fence->parent,
626 					      &s_job->cb)) {
627 			dma_fence_put(s_job->s_fence->parent);
628 			s_job->s_fence->parent = NULL;
629 			atomic_sub(s_job->credits, &sched->credit_count);
630 		} else {
631 			/*
632 			 * remove job from pending_list.
633 			 * Locking here is for concurrent resume timeout
634 			 */
635 			spin_lock(&sched->job_list_lock);
636 			list_del_init(&s_job->list);
637 			spin_unlock(&sched->job_list_lock);
638 
639 			/*
640 			 * Wait for job's HW fence callback to finish using s_job
641 			 * before releasing it.
642 			 *
643 			 * Job is still alive so fence refcount at least 1
644 			 */
645 			dma_fence_wait(&s_job->s_fence->finished, false);
646 
647 			/*
648 			 * We must keep bad job alive for later use during
649 			 * recovery by some of the drivers but leave a hint
650 			 * that the guilty job must be released.
651 			 */
652 			if (bad != s_job)
653 				sched->ops->free_job(s_job);
654 			else
655 				sched->free_guilty = true;
656 		}
657 	}
658 
659 	/*
660 	 * Stop pending timer in flight as we rearm it in  drm_sched_start. This
661 	 * avoids the pending timeout work in progress to fire right away after
662 	 * this TDR finished and before the newly restarted jobs had a
663 	 * chance to complete.
664 	 */
665 	cancel_delayed_work(&sched->work_tdr);
666 }
667 EXPORT_SYMBOL(drm_sched_stop);
668 
669 /**
670  * drm_sched_start - recover jobs after a reset
671  *
672  * @sched: scheduler instance
673  * @errno: error to set on the pending fences
674  *
675  * This function is typically used for reset recovery (see the docu of
676  * drm_sched_backend_ops.timedout_job() for details). Do not call it for
677  * scheduler startup. The scheduler itself is fully operational after
678  * drm_sched_init() succeeded.
679  */
680 void drm_sched_start(struct drm_gpu_scheduler *sched, int errno)
681 {
682 	struct drm_sched_job *s_job, *tmp;
683 
684 	/*
685 	 * Locking the list is not required here as the scheduler's work items
686 	 * are currently not running, so no new jobs are being inserted or
687 	 * removed. Also concurrent GPU recovers can't run in parallel.
688 	 */
689 	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
690 		struct dma_fence *fence = s_job->s_fence->parent;
691 
692 		atomic_add(s_job->credits, &sched->credit_count);
693 
694 		if (!fence) {
695 			drm_sched_job_done(s_job, errno ?: -ECANCELED);
696 			continue;
697 		}
698 
699 		if (dma_fence_add_callback(fence, &s_job->cb,
700 					   drm_sched_job_done_cb))
701 			drm_sched_job_done(s_job, fence->error ?: errno);
702 	}
703 
704 	drm_sched_start_timeout_unlocked(sched);
705 	drm_sched_wqueue_start(sched);
706 }
707 EXPORT_SYMBOL(drm_sched_start);
708 
709 /**
710  * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
711  *
712  * @sched: scheduler instance
713  *
714  * Re-submitting jobs was a concept AMD came up as cheap way to implement
715  * recovery after a job timeout.
716  *
717  * This turned out to be not working very well. First of all there are many
718  * problem with the dma_fence implementation and requirements. Either the
719  * implementation is risking deadlocks with core memory management or violating
720  * documented implementation details of the dma_fence object.
721  *
722  * Drivers can still save and restore their state for recovery operations, but
723  * we shouldn't make this a general scheduler feature around the dma_fence
724  * interface.
725  */
726 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
727 {
728 	struct drm_sched_job *s_job, *tmp;
729 	uint64_t guilty_context;
730 	bool found_guilty = false;
731 	struct dma_fence *fence;
732 
733 	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
734 		struct drm_sched_fence *s_fence = s_job->s_fence;
735 
736 		if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
737 			found_guilty = true;
738 			guilty_context = s_job->s_fence->scheduled.context;
739 		}
740 
741 		if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
742 			dma_fence_set_error(&s_fence->finished, -ECANCELED);
743 
744 		fence = sched->ops->run_job(s_job);
745 
746 		if (IS_ERR_OR_NULL(fence)) {
747 			if (IS_ERR(fence))
748 				dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
749 
750 			s_job->s_fence->parent = NULL;
751 		} else {
752 
753 			s_job->s_fence->parent = dma_fence_get(fence);
754 
755 			/* Drop for orignal kref_init */
756 			dma_fence_put(fence);
757 		}
758 	}
759 }
760 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
761 
762 /**
763  * drm_sched_job_init - init a scheduler job
764  * @job: scheduler job to init
765  * @entity: scheduler entity to use
766  * @credits: the number of credits this job contributes to the schedulers
767  * credit limit
768  * @owner: job owner for debugging
769  * @drm_client_id: &struct drm_file.client_id of the owner (used by trace
770  * events)
771  *
772  * Refer to drm_sched_entity_push_job() documentation
773  * for locking considerations.
774  *
775  * Drivers must make sure drm_sched_job_cleanup() if this function returns
776  * successfully, even when @job is aborted before drm_sched_job_arm() is called.
777  *
778  * Note that this function does not assign a valid value to each struct member
779  * of struct drm_sched_job. Take a look at that struct's documentation to see
780  * who sets which struct member with what lifetime.
781  *
782  * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
783  * has died, which can mean that there's no valid runqueue for a @entity.
784  * This function returns -ENOENT in this case (which probably should be -EIO as
785  * a more meanigful return value).
786  *
787  * Returns 0 for success, negative error code otherwise.
788  */
789 int drm_sched_job_init(struct drm_sched_job *job,
790 		       struct drm_sched_entity *entity,
791 		       u32 credits, void *owner,
792 		       uint64_t drm_client_id)
793 {
794 	if (!entity->rq) {
795 		/* This will most likely be followed by missing frames
796 		 * or worse--a blank screen--leave a trail in the
797 		 * logs, so this can be debugged easier.
798 		 */
799 		dev_err(job->sched->dev, "%s: entity has no rq!\n", __func__);
800 		return -ENOENT;
801 	}
802 
803 	if (unlikely(!credits)) {
804 		pr_err("*ERROR* %s: credits cannot be 0!\n", __func__);
805 		return -EINVAL;
806 	}
807 
808 	/*
809 	 * We don't know for sure how the user has allocated. Thus, zero the
810 	 * struct so that unallowed (i.e., too early) usage of pointers that
811 	 * this function does not set is guaranteed to lead to a NULL pointer
812 	 * exception instead of UB.
813 	 */
814 	memset(job, 0, sizeof(*job));
815 
816 	job->entity = entity;
817 	job->credits = credits;
818 	job->s_fence = drm_sched_fence_alloc(entity, owner, drm_client_id);
819 	if (!job->s_fence)
820 		return -ENOMEM;
821 
822 	INIT_LIST_HEAD(&job->list);
823 
824 	xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
825 
826 	return 0;
827 }
828 EXPORT_SYMBOL(drm_sched_job_init);
829 
830 /**
831  * drm_sched_job_arm - arm a scheduler job for execution
832  * @job: scheduler job to arm
833  *
834  * This arms a scheduler job for execution. Specifically it initializes the
835  * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
836  * or other places that need to track the completion of this job. It also
837  * initializes sequence numbers, which are fundamental for fence ordering.
838  *
839  * Refer to drm_sched_entity_push_job() documentation for locking
840  * considerations.
841  *
842  * Once this function was called, you *must* submit @job with
843  * drm_sched_entity_push_job().
844  *
845  * This can only be called if drm_sched_job_init() succeeded.
846  */
847 void drm_sched_job_arm(struct drm_sched_job *job)
848 {
849 	struct drm_gpu_scheduler *sched;
850 	struct drm_sched_entity *entity = job->entity;
851 
852 	BUG_ON(!entity);
853 	drm_sched_entity_select_rq(entity);
854 	sched = entity->rq->sched;
855 
856 	job->sched = sched;
857 	job->s_priority = entity->priority;
858 
859 	drm_sched_fence_init(job->s_fence, job->entity);
860 }
861 EXPORT_SYMBOL(drm_sched_job_arm);
862 
863 /**
864  * drm_sched_job_add_dependency - adds the fence as a job dependency
865  * @job: scheduler job to add the dependencies to
866  * @fence: the dma_fence to add to the list of dependencies.
867  *
868  * Note that @fence is consumed in both the success and error cases.
869  *
870  * Returns:
871  * 0 on success, or an error on failing to expand the array.
872  */
873 int drm_sched_job_add_dependency(struct drm_sched_job *job,
874 				 struct dma_fence *fence)
875 {
876 	struct dma_fence *entry;
877 	unsigned long index;
878 	u32 id = 0;
879 	int ret;
880 
881 	if (!fence)
882 		return 0;
883 
884 	/* Deduplicate if we already depend on a fence from the same context.
885 	 * This lets the size of the array of deps scale with the number of
886 	 * engines involved, rather than the number of BOs.
887 	 */
888 	xa_for_each(&job->dependencies, index, entry) {
889 		if (entry->context != fence->context)
890 			continue;
891 
892 		if (dma_fence_is_later(fence, entry)) {
893 			dma_fence_put(entry);
894 			xa_store(&job->dependencies, index, fence, GFP_KERNEL);
895 		} else {
896 			dma_fence_put(fence);
897 		}
898 		return 0;
899 	}
900 
901 	ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
902 	if (ret != 0)
903 		dma_fence_put(fence);
904 
905 	return ret;
906 }
907 EXPORT_SYMBOL(drm_sched_job_add_dependency);
908 
909 /**
910  * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
911  * @job: scheduler job to add the dependencies to
912  * @file: drm file private pointer
913  * @handle: syncobj handle to lookup
914  * @point: timeline point
915  *
916  * This adds the fence matching the given syncobj to @job.
917  *
918  * Returns:
919  * 0 on success, or an error on failing to expand the array.
920  */
921 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
922 					 struct drm_file *file,
923 					 u32 handle,
924 					 u32 point)
925 {
926 	struct dma_fence *fence;
927 	int ret;
928 
929 	ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
930 	if (ret)
931 		return ret;
932 
933 	return drm_sched_job_add_dependency(job, fence);
934 }
935 EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
936 
937 /**
938  * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
939  * @job: scheduler job to add the dependencies to
940  * @resv: the dma_resv object to get the fences from
941  * @usage: the dma_resv_usage to use to filter the fences
942  *
943  * This adds all fences matching the given usage from @resv to @job.
944  * Must be called with the @resv lock held.
945  *
946  * Returns:
947  * 0 on success, or an error on failing to expand the array.
948  */
949 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
950 					struct dma_resv *resv,
951 					enum dma_resv_usage usage)
952 {
953 	struct dma_resv_iter cursor;
954 	struct dma_fence *fence;
955 	int ret;
956 
957 	dma_resv_assert_held(resv);
958 
959 	dma_resv_for_each_fence(&cursor, resv, usage, fence) {
960 		/* Make sure to grab an additional ref on the added fence */
961 		dma_fence_get(fence);
962 		ret = drm_sched_job_add_dependency(job, fence);
963 		if (ret) {
964 			dma_fence_put(fence);
965 			return ret;
966 		}
967 	}
968 	return 0;
969 }
970 EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
971 
972 /**
973  * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
974  *   dependencies
975  * @job: scheduler job to add the dependencies to
976  * @obj: the gem object to add new dependencies from.
977  * @write: whether the job might write the object (so we need to depend on
978  * shared fences in the reservation object).
979  *
980  * This should be called after drm_gem_lock_reservations() on your array of
981  * GEM objects used in the job but before updating the reservations with your
982  * own fences.
983  *
984  * Returns:
985  * 0 on success, or an error on failing to expand the array.
986  */
987 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
988 					    struct drm_gem_object *obj,
989 					    bool write)
990 {
991 	return drm_sched_job_add_resv_dependencies(job, obj->resv,
992 						   dma_resv_usage_rw(write));
993 }
994 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
995 
996 /**
997  * drm_sched_job_has_dependency - check whether fence is the job's dependency
998  * @job: scheduler job to check
999  * @fence: fence to look for
1000  *
1001  * Returns:
1002  * True if @fence is found within the job's dependencies, or otherwise false.
1003  */
1004 bool drm_sched_job_has_dependency(struct drm_sched_job *job,
1005 				  struct dma_fence *fence)
1006 {
1007 	struct dma_fence *f;
1008 	unsigned long index;
1009 
1010 	xa_for_each(&job->dependencies, index, f) {
1011 		if (f == fence)
1012 			return true;
1013 	}
1014 
1015 	return false;
1016 }
1017 EXPORT_SYMBOL(drm_sched_job_has_dependency);
1018 
1019 /**
1020  * drm_sched_job_cleanup - clean up scheduler job resources
1021  * @job: scheduler job to clean up
1022  *
1023  * Cleans up the resources allocated with drm_sched_job_init().
1024  *
1025  * Drivers should call this from their error unwind code if @job is aborted
1026  * before drm_sched_job_arm() is called.
1027  *
1028  * drm_sched_job_arm() is a point of no return since it initializes the fences
1029  * and their sequence number etc. Once that function has been called, you *must*
1030  * submit it with drm_sched_entity_push_job() and cannot simply abort it by
1031  * calling drm_sched_job_cleanup().
1032  *
1033  * This function should be called in the &drm_sched_backend_ops.free_job callback.
1034  */
1035 void drm_sched_job_cleanup(struct drm_sched_job *job)
1036 {
1037 	struct dma_fence *fence;
1038 	unsigned long index;
1039 
1040 	if (kref_read(&job->s_fence->finished.refcount)) {
1041 		/* The job has been processed by the scheduler, i.e.,
1042 		 * drm_sched_job_arm() and drm_sched_entity_push_job() have
1043 		 * been called.
1044 		 */
1045 		dma_fence_put(&job->s_fence->finished);
1046 	} else {
1047 		/* The job was aborted before it has been committed to be run;
1048 		 * notably, drm_sched_job_arm() has not been called.
1049 		 */
1050 		drm_sched_fence_free(job->s_fence);
1051 	}
1052 
1053 	job->s_fence = NULL;
1054 
1055 	xa_for_each(&job->dependencies, index, fence) {
1056 		dma_fence_put(fence);
1057 	}
1058 	xa_destroy(&job->dependencies);
1059 
1060 }
1061 EXPORT_SYMBOL(drm_sched_job_cleanup);
1062 
1063 /**
1064  * drm_sched_wakeup - Wake up the scheduler if it is ready to queue
1065  * @sched: scheduler instance
1066  *
1067  * Wake up the scheduler if we can queue jobs.
1068  */
1069 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
1070 {
1071 	drm_sched_run_job_queue(sched);
1072 }
1073 
1074 /**
1075  * drm_sched_select_entity - Select next entity to process
1076  *
1077  * @sched: scheduler instance
1078  *
1079  * Return an entity to process or NULL if none are found.
1080  *
1081  * Note, that we break out of the for-loop when "entity" is non-null, which can
1082  * also be an error-pointer--this assures we don't process lower priority
1083  * run-queues. See comments in the respectively called functions.
1084  */
1085 static struct drm_sched_entity *
1086 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
1087 {
1088 	struct drm_sched_entity *entity;
1089 	int i;
1090 
1091 	/* Start with the highest priority.
1092 	 */
1093 	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1094 		entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
1095 			drm_sched_rq_select_entity_fifo(sched, sched->sched_rq[i]) :
1096 			drm_sched_rq_select_entity_rr(sched, sched->sched_rq[i]);
1097 		if (entity)
1098 			break;
1099 	}
1100 
1101 	return IS_ERR(entity) ? NULL : entity;
1102 }
1103 
1104 /**
1105  * drm_sched_get_finished_job - fetch the next finished job to be destroyed
1106  *
1107  * @sched: scheduler instance
1108  *
1109  * Returns the next finished job from the pending list (if there is one)
1110  * ready for it to be destroyed.
1111  */
1112 static struct drm_sched_job *
1113 drm_sched_get_finished_job(struct drm_gpu_scheduler *sched)
1114 {
1115 	struct drm_sched_job *job, *next;
1116 
1117 	spin_lock(&sched->job_list_lock);
1118 
1119 	job = list_first_entry_or_null(&sched->pending_list,
1120 				       struct drm_sched_job, list);
1121 
1122 	if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
1123 		/* remove job from pending_list */
1124 		list_del_init(&job->list);
1125 
1126 		/* cancel this job's TO timer */
1127 		cancel_delayed_work(&sched->work_tdr);
1128 		/* make the scheduled timestamp more accurate */
1129 		next = list_first_entry_or_null(&sched->pending_list,
1130 						typeof(*next), list);
1131 
1132 		if (next) {
1133 			if (test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT,
1134 				     &next->s_fence->scheduled.flags))
1135 				next->s_fence->scheduled.timestamp =
1136 					dma_fence_timestamp(&job->s_fence->finished);
1137 			/* start TO timer for next job */
1138 			drm_sched_start_timeout(sched);
1139 		}
1140 	} else {
1141 		job = NULL;
1142 	}
1143 
1144 	spin_unlock(&sched->job_list_lock);
1145 
1146 	return job;
1147 }
1148 
1149 /**
1150  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
1151  * @sched_list: list of drm_gpu_schedulers
1152  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
1153  *
1154  * Returns pointer of the sched with the least load or NULL if none of the
1155  * drm_gpu_schedulers are ready
1156  */
1157 struct drm_gpu_scheduler *
1158 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
1159 		     unsigned int num_sched_list)
1160 {
1161 	struct drm_gpu_scheduler *sched, *picked_sched = NULL;
1162 	int i;
1163 	unsigned int min_score = UINT_MAX, num_score;
1164 
1165 	for (i = 0; i < num_sched_list; ++i) {
1166 		sched = sched_list[i];
1167 
1168 		if (!sched->ready) {
1169 			DRM_WARN("scheduler %s is not ready, skipping",
1170 				 sched->name);
1171 			continue;
1172 		}
1173 
1174 		num_score = atomic_read(sched->score);
1175 		if (num_score < min_score) {
1176 			min_score = num_score;
1177 			picked_sched = sched;
1178 		}
1179 	}
1180 
1181 	return picked_sched;
1182 }
1183 EXPORT_SYMBOL(drm_sched_pick_best);
1184 
1185 /**
1186  * drm_sched_free_job_work - worker to call free_job
1187  *
1188  * @w: free job work
1189  */
1190 static void drm_sched_free_job_work(struct work_struct *w)
1191 {
1192 	struct drm_gpu_scheduler *sched =
1193 		container_of(w, struct drm_gpu_scheduler, work_free_job);
1194 	struct drm_sched_job *job;
1195 
1196 	job = drm_sched_get_finished_job(sched);
1197 	if (job)
1198 		sched->ops->free_job(job);
1199 
1200 	drm_sched_run_free_queue(sched);
1201 	drm_sched_run_job_queue(sched);
1202 }
1203 
1204 /**
1205  * drm_sched_run_job_work - worker to call run_job
1206  *
1207  * @w: run job work
1208  */
1209 static void drm_sched_run_job_work(struct work_struct *w)
1210 {
1211 	struct drm_gpu_scheduler *sched =
1212 		container_of(w, struct drm_gpu_scheduler, work_run_job);
1213 	struct drm_sched_entity *entity;
1214 	struct dma_fence *fence;
1215 	struct drm_sched_fence *s_fence;
1216 	struct drm_sched_job *sched_job;
1217 	int r;
1218 
1219 	/* Find entity with a ready job */
1220 	entity = drm_sched_select_entity(sched);
1221 	if (!entity)
1222 		return;	/* No more work */
1223 
1224 	sched_job = drm_sched_entity_pop_job(entity);
1225 	if (!sched_job) {
1226 		complete_all(&entity->entity_idle);
1227 		drm_sched_run_job_queue(sched);
1228 		return;
1229 	}
1230 
1231 	s_fence = sched_job->s_fence;
1232 
1233 	atomic_add(sched_job->credits, &sched->credit_count);
1234 	drm_sched_job_begin(sched_job);
1235 
1236 	trace_drm_sched_job_run(sched_job, entity);
1237 	/*
1238 	 * The run_job() callback must by definition return a fence whose
1239 	 * refcount has been incremented for the scheduler already.
1240 	 */
1241 	fence = sched->ops->run_job(sched_job);
1242 	complete_all(&entity->entity_idle);
1243 	drm_sched_fence_scheduled(s_fence, fence);
1244 
1245 	if (!IS_ERR_OR_NULL(fence)) {
1246 		r = dma_fence_add_callback(fence, &sched_job->cb,
1247 					   drm_sched_job_done_cb);
1248 		if (r == -ENOENT)
1249 			drm_sched_job_done(sched_job, fence->error);
1250 		else if (r)
1251 			DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
1252 
1253 		dma_fence_put(fence);
1254 	} else {
1255 		drm_sched_job_done(sched_job, IS_ERR(fence) ?
1256 				   PTR_ERR(fence) : 0);
1257 	}
1258 
1259 	wake_up(&sched->job_scheduled);
1260 	drm_sched_run_job_queue(sched);
1261 }
1262 
1263 /**
1264  * drm_sched_init - Init a gpu scheduler instance
1265  *
1266  * @sched: scheduler instance
1267  * @args: scheduler initialization arguments
1268  *
1269  * Return 0 on success, otherwise error code.
1270  */
1271 int drm_sched_init(struct drm_gpu_scheduler *sched, const struct drm_sched_init_args *args)
1272 {
1273 	int i;
1274 
1275 	sched->ops = args->ops;
1276 	sched->credit_limit = args->credit_limit;
1277 	sched->name = args->name;
1278 	sched->timeout = args->timeout;
1279 	sched->hang_limit = args->hang_limit;
1280 	sched->timeout_wq = args->timeout_wq ? args->timeout_wq : system_wq;
1281 	sched->score = args->score ? args->score : &sched->_score;
1282 	sched->dev = args->dev;
1283 
1284 	if (args->num_rqs > DRM_SCHED_PRIORITY_COUNT) {
1285 		/* This is a gross violation--tell drivers what the  problem is.
1286 		 */
1287 		dev_err(sched->dev, "%s: num_rqs cannot be greater than DRM_SCHED_PRIORITY_COUNT\n",
1288 			__func__);
1289 		return -EINVAL;
1290 	} else if (sched->sched_rq) {
1291 		/* Not an error, but warn anyway so drivers can
1292 		 * fine-tune their DRM calling order, and return all
1293 		 * is good.
1294 		 */
1295 		dev_warn(sched->dev, "%s: scheduler already initialized!\n", __func__);
1296 		return 0;
1297 	}
1298 
1299 	if (args->submit_wq) {
1300 		sched->submit_wq = args->submit_wq;
1301 		sched->own_submit_wq = false;
1302 	} else {
1303 #ifdef CONFIG_LOCKDEP
1304 		sched->submit_wq = alloc_ordered_workqueue_lockdep_map(args->name,
1305 								       WQ_MEM_RECLAIM,
1306 								       &drm_sched_lockdep_map);
1307 #else
1308 		sched->submit_wq = alloc_ordered_workqueue(args->name, WQ_MEM_RECLAIM);
1309 #endif
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