1 /*
2  * Copyright 2015 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 /**
25  * DOC: Overview
26  *
27  * The GPU scheduler provides entities which allow userspace to push jobs
28  * into software queues which are then scheduled on a hardware run queue.
29  * The software queues have a priority among them. The scheduler selects the entities
30  * from the run queue using a FIFO. The scheduler provides dependency handling
31  * features among jobs. The driver is supposed to provide callback functions for
32  * backend operations to the scheduler like submitting a job to hardware run queue,
33  * returning the dependencies of a job etc.
34  *
35  * The organisation of the scheduler is the following:
36  *
37  * 1. Each hw run queue has one scheduler
38  * 2. Each scheduler has multiple run queues with different priorities
39  *    (e.g., HIGH_HW,HIGH_SW, KERNEL, NORMAL)
40  * 3. Each scheduler run queue has a queue of entities to schedule
41  * 4. Entities themselves maintain a queue of jobs that will be scheduled on
42  *    the hardware.
43  *
44  * The jobs in a entity are always scheduled in the order that they were pushed.
45  *
46  * Note that once a job was taken from the entities queue and pushed to the
47  * hardware, i.e. the pending queue, the entity must not be referenced anymore
48  * through the jobs entity pointer.
49  */
50 
51 #include <linux/kthread.h>
52 #include <linux/wait.h>
53 #include <linux/sched.h>
54 #include <linux/completion.h>
55 #include <linux/dma-resv.h>
56 #include <uapi/linux/sched/types.h>
57 
58 #include <drm/drm_print.h>
59 #include <drm/drm_gem.h>
60 #include <drm/drm_syncobj.h>
61 #include <drm/gpu_scheduler.h>
62 #include <drm/spsc_queue.h>
63 
64 #define CREATE_TRACE_POINTS
65 #include "gpu_scheduler_trace.h"
66 
67 #define to_drm_sched_job(sched_job)		\
68 		container_of((sched_job), struct drm_sched_job, queue_node)
69 
70 int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
71 
72 /**
73  * DOC: sched_policy (int)
74  * Used to override default entities scheduling policy in a run queue.
75  */
76 MODULE_PARM_DESC(sched_policy, "Specify the scheduling policy for entities on a run-queue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin, " __stringify(DRM_SCHED_POLICY_FIFO) " = FIFO (default).");
77 module_param_named(sched_policy, drm_sched_policy, int, 0444);
78 
79 static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
80 							    const struct rb_node *b)
81 {
82 	struct drm_sched_entity *ent_a =  rb_entry((a), struct drm_sched_entity, rb_tree_node);
83 	struct drm_sched_entity *ent_b =  rb_entry((b), struct drm_sched_entity, rb_tree_node);
84 
85 	return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
86 }
87 
88 static inline void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity)
89 {
90 	struct drm_sched_rq *rq = entity->rq;
91 
92 	if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
93 		rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
94 		RB_CLEAR_NODE(&entity->rb_tree_node);
95 	}
96 }
97 
98 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts)
99 {
100 	/*
101 	 * Both locks need to be grabbed, one to protect from entity->rq change
102 	 * for entity from within concurrent drm_sched_entity_select_rq and the
103 	 * other to update the rb tree structure.
104 	 */
105 	spin_lock(&entity->rq_lock);
106 	spin_lock(&entity->rq->lock);
107 
108 	drm_sched_rq_remove_fifo_locked(entity);
109 
110 	entity->oldest_job_waiting = ts;
111 
112 	rb_add_cached(&entity->rb_tree_node, &entity->rq->rb_tree_root,
113 		      drm_sched_entity_compare_before);
114 
115 	spin_unlock(&entity->rq->lock);
116 	spin_unlock(&entity->rq_lock);
117 }
118 
119 /**
120  * drm_sched_rq_init - initialize a given run queue struct
121  *
122  * @sched: scheduler instance to associate with this run queue
123  * @rq: scheduler run queue
124  *
125  * Initializes a scheduler runqueue.
126  */
127 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
128 			      struct drm_sched_rq *rq)
129 {
130 	spin_lock_init(&rq->lock);
131 	INIT_LIST_HEAD(&rq->entities);
132 	rq->rb_tree_root = RB_ROOT_CACHED;
133 	rq->current_entity = NULL;
134 	rq->sched = sched;
135 }
136 
137 /**
138  * drm_sched_rq_add_entity - add an entity
139  *
140  * @rq: scheduler run queue
141  * @entity: scheduler entity
142  *
143  * Adds a scheduler entity to the run queue.
144  */
145 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
146 			     struct drm_sched_entity *entity)
147 {
148 	if (!list_empty(&entity->list))
149 		return;
150 
151 	spin_lock(&rq->lock);
152 
153 	atomic_inc(rq->sched->score);
154 	list_add_tail(&entity->list, &rq->entities);
155 
156 	spin_unlock(&rq->lock);
157 }
158 
159 /**
160  * drm_sched_rq_remove_entity - remove an entity
161  *
162  * @rq: scheduler run queue
163  * @entity: scheduler entity
164  *
165  * Removes a scheduler entity from the run queue.
166  */
167 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
168 				struct drm_sched_entity *entity)
169 {
170 	if (list_empty(&entity->list))
171 		return;
172 
173 	spin_lock(&rq->lock);
174 
175 	atomic_dec(rq->sched->score);
176 	list_del_init(&entity->list);
177 
178 	if (rq->current_entity == entity)
179 		rq->current_entity = NULL;
180 
181 	if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
182 		drm_sched_rq_remove_fifo_locked(entity);
183 
184 	spin_unlock(&rq->lock);
185 }
186 
187 /**
188  * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
189  *
190  * @rq: scheduler run queue to check.
191  *
192  * Try to find a ready entity, returns NULL if none found.
193  */
194 static struct drm_sched_entity *
195 drm_sched_rq_select_entity_rr(struct drm_sched_rq *rq)
196 {
197 	struct drm_sched_entity *entity;
198 
199 	spin_lock(&rq->lock);
200 
201 	entity = rq->current_entity;
202 	if (entity) {
203 		list_for_each_entry_continue(entity, &rq->entities, list) {
204 			if (drm_sched_entity_is_ready(entity)) {
205 				rq->current_entity = entity;
206 				reinit_completion(&entity->entity_idle);
207 				spin_unlock(&rq->lock);
208 				return entity;
209 			}
210 		}
211 	}
212 
213 	list_for_each_entry(entity, &rq->entities, list) {
214 
215 		if (drm_sched_entity_is_ready(entity)) {
216 			rq->current_entity = entity;
217 			reinit_completion(&entity->entity_idle);
218 			spin_unlock(&rq->lock);
219 			return entity;
220 		}
221 
222 		if (entity == rq->current_entity)
223 			break;
224 	}
225 
226 	spin_unlock(&rq->lock);
227 
228 	return NULL;
229 }
230 
231 /**
232  * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
233  *
234  * @rq: scheduler run queue to check.
235  *
236  * Find oldest waiting ready entity, returns NULL if none found.
237  */
238 static struct drm_sched_entity *
239 drm_sched_rq_select_entity_fifo(struct drm_sched_rq *rq)
240 {
241 	struct rb_node *rb;
242 
243 	spin_lock(&rq->lock);
244 	for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
245 		struct drm_sched_entity *entity;
246 
247 		entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
248 		if (drm_sched_entity_is_ready(entity)) {
249 			rq->current_entity = entity;
250 			reinit_completion(&entity->entity_idle);
251 			break;
252 		}
253 	}
254 	spin_unlock(&rq->lock);
255 
256 	return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
257 }
258 
259 /**
260  * drm_sched_job_done - complete a job
261  * @s_job: pointer to the job which is done
262  *
263  * Finish the job's fence and wake up the worker thread.
264  */
265 static void drm_sched_job_done(struct drm_sched_job *s_job, int result)
266 {
267 	struct drm_sched_fence *s_fence = s_job->s_fence;
268 	struct drm_gpu_scheduler *sched = s_fence->sched;
269 
270 	atomic_dec(&sched->hw_rq_count);
271 	atomic_dec(sched->score);
272 
273 	trace_drm_sched_process_job(s_fence);
274 
275 	dma_fence_get(&s_fence->finished);
276 	drm_sched_fence_finished(s_fence, result);
277 	dma_fence_put(&s_fence->finished);
278 	wake_up_interruptible(&sched->wake_up_worker);
279 }
280 
281 /**
282  * drm_sched_job_done_cb - the callback for a done job
283  * @f: fence
284  * @cb: fence callbacks
285  */
286 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
287 {
288 	struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
289 
290 	drm_sched_job_done(s_job, f->error);
291 }
292 
293 /**
294  * drm_sched_start_timeout - start timeout for reset worker
295  *
296  * @sched: scheduler instance to start the worker for
297  *
298  * Start the timeout for the given scheduler.
299  */
300 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
301 {
302 	if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
303 	    !list_empty(&sched->pending_list))
304 		queue_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
305 }
306 
307 /**
308  * drm_sched_fault - immediately start timeout handler
309  *
310  * @sched: scheduler where the timeout handling should be started.
311  *
312  * Start timeout handling immediately when the driver detects a hardware fault.
313  */
314 void drm_sched_fault(struct drm_gpu_scheduler *sched)
315 {
316 	if (sched->timeout_wq)
317 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
318 }
319 EXPORT_SYMBOL(drm_sched_fault);
320 
321 /**
322  * drm_sched_suspend_timeout - Suspend scheduler job timeout
323  *
324  * @sched: scheduler instance for which to suspend the timeout
325  *
326  * Suspend the delayed work timeout for the scheduler. This is done by
327  * modifying the delayed work timeout to an arbitrary large value,
328  * MAX_SCHEDULE_TIMEOUT in this case.
329  *
330  * Returns the timeout remaining
331  *
332  */
333 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
334 {
335 	unsigned long sched_timeout, now = jiffies;
336 
337 	sched_timeout = sched->work_tdr.timer.expires;
338 
339 	/*
340 	 * Modify the timeout to an arbitrarily large value. This also prevents
341 	 * the timeout to be restarted when new submissions arrive
342 	 */
343 	if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
344 			&& time_after(sched_timeout, now))
345 		return sched_timeout - now;
346 	else
347 		return sched->timeout;
348 }
349 EXPORT_SYMBOL(drm_sched_suspend_timeout);
350 
351 /**
352  * drm_sched_resume_timeout - Resume scheduler job timeout
353  *
354  * @sched: scheduler instance for which to resume the timeout
355  * @remaining: remaining timeout
356  *
357  * Resume the delayed work timeout for the scheduler.
358  */
359 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
360 		unsigned long remaining)
361 {
362 	spin_lock(&sched->job_list_lock);
363 
364 	if (list_empty(&sched->pending_list))
365 		cancel_delayed_work(&sched->work_tdr);
366 	else
367 		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
368 
369 	spin_unlock(&sched->job_list_lock);
370 }
371 EXPORT_SYMBOL(drm_sched_resume_timeout);
372 
373 static void drm_sched_job_begin(struct drm_sched_job *s_job)
374 {
375 	struct drm_gpu_scheduler *sched = s_job->sched;
376 
377 	spin_lock(&sched->job_list_lock);
378 	list_add_tail(&s_job->list, &sched->pending_list);
379 	drm_sched_start_timeout(sched);
380 	spin_unlock(&sched->job_list_lock);
381 }
382 
383 static void drm_sched_job_timedout(struct work_struct *work)
384 {
385 	struct drm_gpu_scheduler *sched;
386 	struct drm_sched_job *job;
387 	enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
388 
389 	sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
390 
391 	/* Protects against concurrent deletion in drm_sched_get_cleanup_job */
392 	spin_lock(&sched->job_list_lock);
393 	job = list_first_entry_or_null(&sched->pending_list,
394 				       struct drm_sched_job, list);
395 
396 	if (job) {
397 		/*
398 		 * Remove the bad job so it cannot be freed by concurrent
399 		 * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
400 		 * is parked at which point it's safe.
401 		 */
402 		list_del_init(&job->list);
403 		spin_unlock(&sched->job_list_lock);
404 
405 		status = job->sched->ops->timedout_job(job);
406 
407 		/*
408 		 * Guilty job did complete and hence needs to be manually removed
409 		 * See drm_sched_stop doc.
410 		 */
411 		if (sched->free_guilty) {
412 			job->sched->ops->free_job(job);
413 			sched->free_guilty = false;
414 		}
415 	} else {
416 		spin_unlock(&sched->job_list_lock);
417 	}
418 
419 	if (status != DRM_GPU_SCHED_STAT_ENODEV) {
420 		spin_lock(&sched->job_list_lock);
421 		drm_sched_start_timeout(sched);
422 		spin_unlock(&sched->job_list_lock);
423 	}
424 }
425 
426 /**
427  * drm_sched_stop - stop the scheduler
428  *
429  * @sched: scheduler instance
430  * @bad: job which caused the time out
431  *
432  * Stop the scheduler and also removes and frees all completed jobs.
433  * Note: bad job will not be freed as it might be used later and so it's
434  * callers responsibility to release it manually if it's not part of the
435  * pending list any more.
436  *
437  */
438 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
439 {
440 	struct drm_sched_job *s_job, *tmp;
441 
442 	kthread_park(sched->thread);
443 
444 	/*
445 	 * Reinsert back the bad job here - now it's safe as
446 	 * drm_sched_get_cleanup_job cannot race against us and release the
447 	 * bad job at this point - we parked (waited for) any in progress
448 	 * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
449 	 * now until the scheduler thread is unparked.
450 	 */
451 	if (bad && bad->sched == sched)
452 		/*
453 		 * Add at the head of the queue to reflect it was the earliest
454 		 * job extracted.
455 		 */
456 		list_add(&bad->list, &sched->pending_list);
457 
458 	/*
459 	 * Iterate the job list from later to  earlier one and either deactive
460 	 * their HW callbacks or remove them from pending list if they already
461 	 * signaled.
462 	 * This iteration is thread safe as sched thread is stopped.
463 	 */
464 	list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
465 					 list) {
466 		if (s_job->s_fence->parent &&
467 		    dma_fence_remove_callback(s_job->s_fence->parent,
468 					      &s_job->cb)) {
469 			dma_fence_put(s_job->s_fence->parent);
470 			s_job->s_fence->parent = NULL;
471 			atomic_dec(&sched->hw_rq_count);
472 		} else {
473 			/*
474 			 * remove job from pending_list.
475 			 * Locking here is for concurrent resume timeout
476 			 */
477 			spin_lock(&sched->job_list_lock);
478 			list_del_init(&s_job->list);
479 			spin_unlock(&sched->job_list_lock);
480 
481 			/*
482 			 * Wait for job's HW fence callback to finish using s_job
483 			 * before releasing it.
484 			 *
485 			 * Job is still alive so fence refcount at least 1
486 			 */
487 			dma_fence_wait(&s_job->s_fence->finished, false);
488 
489 			/*
490 			 * We must keep bad job alive for later use during
491 			 * recovery by some of the drivers but leave a hint
492 			 * that the guilty job must be released.
493 			 */
494 			if (bad != s_job)
495 				sched->ops->free_job(s_job);
496 			else
497 				sched->free_guilty = true;
498 		}
499 	}
500 
501 	/*
502 	 * Stop pending timer in flight as we rearm it in  drm_sched_start. This
503 	 * avoids the pending timeout work in progress to fire right away after
504 	 * this TDR finished and before the newly restarted jobs had a
505 	 * chance to complete.
506 	 */
507 	cancel_delayed_work(&sched->work_tdr);
508 }
509 
510 EXPORT_SYMBOL(drm_sched_stop);
511 
512 /**
513  * drm_sched_start - recover jobs after a reset
514  *
515  * @sched: scheduler instance
516  * @full_recovery: proceed with complete sched restart
517  *
518  */
519 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
520 {
521 	struct drm_sched_job *s_job, *tmp;
522 	int r;
523 
524 	/*
525 	 * Locking the list is not required here as the sched thread is parked
526 	 * so no new jobs are being inserted or removed. Also concurrent
527 	 * GPU recovers can't run in parallel.
528 	 */
529 	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
530 		struct dma_fence *fence = s_job->s_fence->parent;
531 
532 		atomic_inc(&sched->hw_rq_count);
533 
534 		if (!full_recovery)
535 			continue;
536 
537 		if (fence) {
538 			r = dma_fence_add_callback(fence, &s_job->cb,
539 						   drm_sched_job_done_cb);
540 			if (r == -ENOENT)
541 				drm_sched_job_done(s_job, fence->error);
542 			else if (r)
543 				DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
544 					  r);
545 		} else
546 			drm_sched_job_done(s_job, -ECANCELED);
547 	}
548 
549 	if (full_recovery) {
550 		spin_lock(&sched->job_list_lock);
551 		drm_sched_start_timeout(sched);
552 		spin_unlock(&sched->job_list_lock);
553 	}
554 
555 	kthread_unpark(sched->thread);
556 }
557 EXPORT_SYMBOL(drm_sched_start);
558 
559 /**
560  * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
561  *
562  * @sched: scheduler instance
563  *
564  * Re-submitting jobs was a concept AMD came up as cheap way to implement
565  * recovery after a job timeout.
566  *
567  * This turned out to be not working very well. First of all there are many
568  * problem with the dma_fence implementation and requirements. Either the
569  * implementation is risking deadlocks with core memory management or violating
570  * documented implementation details of the dma_fence object.
571  *
572  * Drivers can still save and restore their state for recovery operations, but
573  * we shouldn't make this a general scheduler feature around the dma_fence
574  * interface.
575  */
576 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
577 {
578 	struct drm_sched_job *s_job, *tmp;
579 	uint64_t guilty_context;
580 	bool found_guilty = false;
581 	struct dma_fence *fence;
582 
583 	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
584 		struct drm_sched_fence *s_fence = s_job->s_fence;
585 
586 		if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
587 			found_guilty = true;
588 			guilty_context = s_job->s_fence->scheduled.context;
589 		}
590 
591 		if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
592 			dma_fence_set_error(&s_fence->finished, -ECANCELED);
593 
594 		fence = sched->ops->run_job(s_job);
595 
596 		if (IS_ERR_OR_NULL(fence)) {
597 			if (IS_ERR(fence))
598 				dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
599 
600 			s_job->s_fence->parent = NULL;
601 		} else {
602 
603 			s_job->s_fence->parent = dma_fence_get(fence);
604 
605 			/* Drop for orignal kref_init */
606 			dma_fence_put(fence);
607 		}
608 	}
609 }
610 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
611 
612 /**
613  * drm_sched_job_init - init a scheduler job
614  * @job: scheduler job to init
615  * @entity: scheduler entity to use
616  * @owner: job owner for debugging
617  *
618  * Refer to drm_sched_entity_push_job() documentation
619  * for locking considerations.
620  *
621  * Drivers must make sure drm_sched_job_cleanup() if this function returns
622  * successfully, even when @job is aborted before drm_sched_job_arm() is called.
623  *
624  * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
625  * has died, which can mean that there's no valid runqueue for a @entity.
626  * This function returns -ENOENT in this case (which probably should be -EIO as
627  * a more meanigful return value).
628  *
629  * Returns 0 for success, negative error code otherwise.
630  */
631 int drm_sched_job_init(struct drm_sched_job *job,
632 		       struct drm_sched_entity *entity,
633 		       void *owner)
634 {
635 	if (!entity->rq)
636 		return -ENOENT;
637 
638 	job->entity = entity;
639 	job->s_fence = drm_sched_fence_alloc(entity, owner);
640 	if (!job->s_fence)
641 		return -ENOMEM;
642 
643 	INIT_LIST_HEAD(&job->list);
644 
645 	xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
646 
647 	return 0;
648 }
649 EXPORT_SYMBOL(drm_sched_job_init);
650 
651 /**
652  * drm_sched_job_arm - arm a scheduler job for execution
653  * @job: scheduler job to arm
654  *
655  * This arms a scheduler job for execution. Specifically it initializes the
656  * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
657  * or other places that need to track the completion of this job.
658  *
659  * Refer to drm_sched_entity_push_job() documentation for locking
660  * considerations.
661  *
662  * This can only be called if drm_sched_job_init() succeeded.
663  */
664 void drm_sched_job_arm(struct drm_sched_job *job)
665 {
666 	struct drm_gpu_scheduler *sched;
667 	struct drm_sched_entity *entity = job->entity;
668 
669 	BUG_ON(!entity);
670 	drm_sched_entity_select_rq(entity);
671 	sched = entity->rq->sched;
672 
673 	job->sched = sched;
674 	job->s_priority = entity->rq - sched->sched_rq;
675 	job->id = atomic64_inc_return(&sched->job_id_count);
676 
677 	drm_sched_fence_init(job->s_fence, job->entity);
678 }
679 EXPORT_SYMBOL(drm_sched_job_arm);
680 
681 /**
682  * drm_sched_job_add_dependency - adds the fence as a job dependency
683  * @job: scheduler job to add the dependencies to
684  * @fence: the dma_fence to add to the list of dependencies.
685  *
686  * Note that @fence is consumed in both the success and error cases.
687  *
688  * Returns:
689  * 0 on success, or an error on failing to expand the array.
690  */
691 int drm_sched_job_add_dependency(struct drm_sched_job *job,
692 				 struct dma_fence *fence)
693 {
694 	struct dma_fence *entry;
695 	unsigned long index;
696 	u32 id = 0;
697 	int ret;
698 
699 	if (!fence)
700 		return 0;
701 
702 	/* Deduplicate if we already depend on a fence from the same context.
703 	 * This lets the size of the array of deps scale with the number of
704 	 * engines involved, rather than the number of BOs.
705 	 */
706 	xa_for_each(&job->dependencies, index, entry) {
707 		if (entry->context != fence->context)
708 			continue;
709 
710 		if (dma_fence_is_later(fence, entry)) {
711 			dma_fence_put(entry);
712 			xa_store(&job->dependencies, index, fence, GFP_KERNEL);
713 		} else {
714 			dma_fence_put(fence);
715 		}
716 		return 0;
717 	}
718 
719 	ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
720 	if (ret != 0)
721 		dma_fence_put(fence);
722 
723 	return ret;
724 }
725 EXPORT_SYMBOL(drm_sched_job_add_dependency);
726 
727 /**
728  * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
729  * @job: scheduler job to add the dependencies to
730  * @file: drm file private pointer
731  * @handle: syncobj handle to lookup
732  * @point: timeline point
733  *
734  * This adds the fence matching the given syncobj to @job.
735  *
736  * Returns:
737  * 0 on success, or an error on failing to expand the array.
738  */
739 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
740 					 struct drm_file *file,
741 					 u32 handle,
742 					 u32 point)
743 {
744 	struct dma_fence *fence;
745 	int ret;
746 
747 	ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
748 	if (ret)
749 		return ret;
750 
751 	return drm_sched_job_add_dependency(job, fence);
752 }
753 EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
754 
755 /**
756  * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
757  * @job: scheduler job to add the dependencies to
758  * @resv: the dma_resv object to get the fences from
759  * @usage: the dma_resv_usage to use to filter the fences
760  *
761  * This adds all fences matching the given usage from @resv to @job.
762  * Must be called with the @resv lock held.
763  *
764  * Returns:
765  * 0 on success, or an error on failing to expand the array.
766  */
767 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
768 					struct dma_resv *resv,
769 					enum dma_resv_usage usage)
770 {
771 	struct dma_resv_iter cursor;
772 	struct dma_fence *fence;
773 	int ret;
774 
775 	dma_resv_assert_held(resv);
776 
777 	dma_resv_for_each_fence(&cursor, resv, usage, fence) {
778 		/* Make sure to grab an additional ref on the added fence */
779 		dma_fence_get(fence);
780 		ret = drm_sched_job_add_dependency(job, fence);
781 		if (ret) {
782 			dma_fence_put(fence);
783 			return ret;
784 		}
785 	}
786 	return 0;
787 }
788 EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
789 
790 /**
791  * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
792  *   dependencies
793  * @job: scheduler job to add the dependencies to
794  * @obj: the gem object to add new dependencies from.
795  * @write: whether the job might write the object (so we need to depend on
796  * shared fences in the reservation object).
797  *
798  * This should be called after drm_gem_lock_reservations() on your array of
799  * GEM objects used in the job but before updating the reservations with your
800  * own fences.
801  *
802  * Returns:
803  * 0 on success, or an error on failing to expand the array.
804  */
805 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
806 					    struct drm_gem_object *obj,
807 					    bool write)
808 {
809 	return drm_sched_job_add_resv_dependencies(job, obj->resv,
810 						   dma_resv_usage_rw(write));
811 }
812 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
813 
814 /**
815  * drm_sched_job_cleanup - clean up scheduler job resources
816  * @job: scheduler job to clean up
817  *
818  * Cleans up the resources allocated with drm_sched_job_init().
819  *
820  * Drivers should call this from their error unwind code if @job is aborted
821  * before drm_sched_job_arm() is called.
822  *
823  * After that point of no return @job is committed to be executed by the
824  * scheduler, and this function should be called from the
825  * &drm_sched_backend_ops.free_job callback.
826  */
827 void drm_sched_job_cleanup(struct drm_sched_job *job)
828 {
829 	struct dma_fence *fence;
830 	unsigned long index;
831 
832 	if (kref_read(&job->s_fence->finished.refcount)) {
833 		/* drm_sched_job_arm() has been called */
834 		dma_fence_put(&job->s_fence->finished);
835 	} else {
836 		/* aborted job before committing to run it */
837 		drm_sched_fence_free(job->s_fence);
838 	}
839 
840 	job->s_fence = NULL;
841 
842 	xa_for_each(&job->dependencies, index, fence) {
843 		dma_fence_put(fence);
844 	}
845 	xa_destroy(&job->dependencies);
846 
847 }
848 EXPORT_SYMBOL(drm_sched_job_cleanup);
849 
850 /**
851  * drm_sched_can_queue -- Can we queue more to the hardware?
852  * @sched: scheduler instance
853  *
854  * Return true if we can push more jobs to the hw, otherwise false.
855  */
856 static bool drm_sched_can_queue(struct drm_gpu_scheduler *sched)
857 {
858 	return atomic_read(&sched->hw_rq_count) <
859 		sched->hw_submission_limit;
860 }
861 
862 /**
863  * drm_sched_wakeup_if_can_queue - Wake up the scheduler
864  * @sched: scheduler instance
865  *
866  * Wake up the scheduler if we can queue jobs.
867  */
868 void drm_sched_wakeup_if_can_queue(struct drm_gpu_scheduler *sched)
869 {
870 	if (drm_sched_can_queue(sched))
871 		wake_up_interruptible(&sched->wake_up_worker);
872 }
873 
874 /**
875  * drm_sched_select_entity - Select next entity to process
876  *
877  * @sched: scheduler instance
878  *
879  * Returns the entity to process or NULL if none are found.
880  */
881 static struct drm_sched_entity *
882 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
883 {
884 	struct drm_sched_entity *entity;
885 	int i;
886 
887 	if (!drm_sched_can_queue(sched))
888 		return NULL;
889 
890 	/* Kernel run queue has higher priority than normal run queue*/
891 	for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
892 		entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
893 			drm_sched_rq_select_entity_fifo(&sched->sched_rq[i]) :
894 			drm_sched_rq_select_entity_rr(&sched->sched_rq[i]);
895 		if (entity)
896 			break;
897 	}
898 
899 	return entity;
900 }
901 
902 /**
903  * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
904  *
905  * @sched: scheduler instance
906  *
907  * Returns the next finished job from the pending list (if there is one)
908  * ready for it to be destroyed.
909  */
910 static struct drm_sched_job *
911 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
912 {
913 	struct drm_sched_job *job, *next;
914 
915 	spin_lock(&sched->job_list_lock);
916 
917 	job = list_first_entry_or_null(&sched->pending_list,
918 				       struct drm_sched_job, list);
919 
920 	if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
921 		/* remove job from pending_list */
922 		list_del_init(&job->list);
923 
924 		/* cancel this job's TO timer */
925 		cancel_delayed_work(&sched->work_tdr);
926 		/* make the scheduled timestamp more accurate */
927 		next = list_first_entry_or_null(&sched->pending_list,
928 						typeof(*next), list);
929 
930 		if (next) {
931 			next->s_fence->scheduled.timestamp =
932 				job->s_fence->finished.timestamp;
933 			/* start TO timer for next job */
934 			drm_sched_start_timeout(sched);
935 		}
936 	} else {
937 		job = NULL;
938 	}
939 
940 	spin_unlock(&sched->job_list_lock);
941 
942 	return job;
943 }
944 
945 /**
946  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
947  * @sched_list: list of drm_gpu_schedulers
948  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
949  *
950  * Returns pointer of the sched with the least load or NULL if none of the
951  * drm_gpu_schedulers are ready
952  */
953 struct drm_gpu_scheduler *
954 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
955 		     unsigned int num_sched_list)
956 {
957 	struct drm_gpu_scheduler *sched, *picked_sched = NULL;
958 	int i;
959 	unsigned int min_score = UINT_MAX, num_score;
960 
961 	for (i = 0; i < num_sched_list; ++i) {
962 		sched = sched_list[i];
963 
964 		if (!sched->ready) {
965 			DRM_WARN("scheduler %s is not ready, skipping",
966 				 sched->name);
967 			continue;
968 		}
969 
970 		num_score = atomic_read(sched->score);
971 		if (num_score < min_score) {
972 			min_score = num_score;
973 			picked_sched = sched;
974 		}
975 	}
976 
977 	return picked_sched;
978 }
979 EXPORT_SYMBOL(drm_sched_pick_best);
980 
981 /**
982  * drm_sched_blocked - check if the scheduler is blocked
983  *
984  * @sched: scheduler instance
985  *
986  * Returns true if blocked, otherwise false.
987  */
988 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
989 {
990 	if (kthread_should_park()) {
991 		kthread_parkme();
992 		return true;
993 	}
994 
995 	return false;
996 }
997 
998 /**
999  * drm_sched_main - main scheduler thread
1000  *
1001  * @param: scheduler instance
1002  *
1003  * Returns 0.
1004  */
1005 static int drm_sched_main(void *param)
1006 {
1007 	struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
1008 	int r;
1009 
1010 	sched_set_fifo_low(current);
1011 
1012 	while (!kthread_should_stop()) {
1013 		struct drm_sched_entity *entity = NULL;
1014 		struct drm_sched_fence *s_fence;
1015 		struct drm_sched_job *sched_job;
1016 		struct dma_fence *fence;
1017 		struct drm_sched_job *cleanup_job = NULL;
1018 
1019 		wait_event_interruptible(sched->wake_up_worker,
1020 					 (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
1021 					 (!drm_sched_blocked(sched) &&
1022 					  (entity = drm_sched_select_entity(sched))) ||
1023 					 kthread_should_stop());
1024 
1025 		if (cleanup_job)
1026 			sched->ops->free_job(cleanup_job);
1027 
1028 		if (!entity)
1029 			continue;
1030 
1031 		sched_job = drm_sched_entity_pop_job(entity);
1032 
1033 		if (!sched_job) {
1034 			complete_all(&entity->entity_idle);
1035 			continue;
1036 		}
1037 
1038 		s_fence = sched_job->s_fence;
1039 
1040 		atomic_inc(&sched->hw_rq_count);
1041 		drm_sched_job_begin(sched_job);
1042 
1043 		trace_drm_run_job(sched_job, entity);
1044 		fence = sched->ops->run_job(sched_job);
1045 		complete_all(&entity->entity_idle);
1046 		drm_sched_fence_scheduled(s_fence, fence);
1047 
1048 		if (!IS_ERR_OR_NULL(fence)) {
1049 			/* Drop for original kref_init of the fence */
1050 			dma_fence_put(fence);
1051 
1052 			r = dma_fence_add_callback(fence, &sched_job->cb,
1053 						   drm_sched_job_done_cb);
1054 			if (r == -ENOENT)
1055 				drm_sched_job_done(sched_job, fence->error);
1056 			else if (r)
1057 				DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
1058 					  r);
1059 		} else {
1060 			drm_sched_job_done(sched_job, IS_ERR(fence) ?
1061 					   PTR_ERR(fence) : 0);
1062 		}
1063 
1064 		wake_up(&sched->job_scheduled);
1065 	}
1066 	return 0;
1067 }
1068 
1069 /**
1070  * drm_sched_init - Init a gpu scheduler instance
1071  *
1072  * @sched: scheduler instance
1073  * @ops: backend operations for this scheduler
1074  * @hw_submission: number of hw submissions that can be in flight
1075  * @hang_limit: number of times to allow a job to hang before dropping it
1076  * @timeout: timeout value in jiffies for the scheduler
1077  * @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
1078  *		used
1079  * @score: optional score atomic shared with other schedulers
1080  * @name: name used for debugging
1081  * @dev: target &struct device
1082  *
1083  * Return 0 on success, otherwise error code.
1084  */
1085 int drm_sched_init(struct drm_gpu_scheduler *sched,
1086 		   const struct drm_sched_backend_ops *ops,
1087 		   unsigned hw_submission, unsigned hang_limit,
1088 		   long timeout, struct workqueue_struct *timeout_wq,
1089 		   atomic_t *score, const char *name, struct device *dev)
1090 {
1091 	int i, ret;
1092 	sched->ops = ops;
1093 	sched->hw_submission_limit = hw_submission;
1094 	sched->name = name;
1095 	sched->timeout = timeout;
1096 	sched->timeout_wq = timeout_wq ? : system_wq;
1097 	sched->hang_limit = hang_limit;
1098 	sched->score = score ? score : &sched->_score;
1099 	sched->dev = dev;
1100 	for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
1101 		drm_sched_rq_init(sched, &sched->sched_rq[i]);
1102 
1103 	init_waitqueue_head(&sched->wake_up_worker);
1104 	init_waitqueue_head(&sched->job_scheduled);
1105 	INIT_LIST_HEAD(&sched->pending_list);
1106 	spin_lock_init(&sched->job_list_lock);
1107 	atomic_set(&sched->hw_rq_count, 0);
1108 	INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1109 	atomic_set(&sched->_score, 0);
1110 	atomic64_set(&sched->job_id_count, 0);
1111 
1112 	/* Each scheduler will run on a seperate kernel thread */
1113 	sched->thread = kthread_run(drm_sched_main, sched, sched->name);
1114 	if (IS_ERR(sched->thread)) {
1115 		ret = PTR_ERR(sched->thread);
1116 		sched->thread = NULL;
1117 		DRM_DEV_ERROR(sched->dev, "Failed to create scheduler for %s.\n", name);
1118 		return ret;
1119 	}
1120 
1121 	sched->ready = true;
1122 	return 0;
1123 }
1124 EXPORT_SYMBOL(drm_sched_init);
1125 
1126 /**
1127  * drm_sched_fini - Destroy a gpu scheduler
1128  *
1129  * @sched: scheduler instance
1130  *
1131  * Tears down and cleans up the scheduler.
1132  */
1133 void drm_sched_fini(struct drm_gpu_scheduler *sched)
1134 {
1135 	struct drm_sched_entity *s_entity;
1136 	int i;
1137 
1138 	if (sched->thread)
1139 		kthread_stop(sched->thread);
1140 
1141 	for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
1142 		struct drm_sched_rq *rq = &sched->sched_rq[i];
1143 
1144 		spin_lock(&rq->lock);
1145 		list_for_each_entry(s_entity, &rq->entities, list)
1146 			/*
1147 			 * Prevents reinsertion and marks job_queue as idle,
1148 			 * it will removed from rq in drm_sched_entity_fini
1149 			 * eventually
1150 			 */
1151 			s_entity->stopped = true;
1152 		spin_unlock(&rq->lock);
1153 
1154 	}
1155 
1156 	/* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1157 	wake_up_all(&sched->job_scheduled);
1158 
1159 	/* Confirm no work left behind accessing device structures */
1160 	cancel_delayed_work_sync(&sched->work_tdr);
1161 
1162 	sched->ready = false;
1163 }
1164 EXPORT_SYMBOL(drm_sched_fini);
1165 
1166 /**
1167  * drm_sched_increase_karma - Update sched_entity guilty flag
1168  *
1169  * @bad: The job guilty of time out
1170  *
1171  * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1172  * limit of the scheduler then the respective sched entity is marked guilty and
1173  * jobs from it will not be scheduled further
1174  */
1175 void drm_sched_increase_karma(struct drm_sched_job *bad)
1176 {
1177 	int i;
1178 	struct drm_sched_entity *tmp;
1179 	struct drm_sched_entity *entity;
1180 	struct drm_gpu_scheduler *sched = bad->sched;
1181 
1182 	/* don't change @bad's karma if it's from KERNEL RQ,
1183 	 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1184 	 * corrupt but keep in mind that kernel jobs always considered good.
1185 	 */
1186 	if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1187 		atomic_inc(&bad->karma);
1188 
1189 		for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
1190 		     i++) {
1191 			struct drm_sched_rq *rq = &sched->sched_rq[i];
1192 
1193 			spin_lock(&rq->lock);
1194 			list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1195 				if (bad->s_fence->scheduled.context ==
1196 				    entity->fence_context) {
1197 					if (entity->guilty)
1198 						atomic_set(entity->guilty, 1);
1199 					break;
1200 				}
1201 			}
1202 			spin_unlock(&rq->lock);
1203 			if (&entity->list != &rq->entities)
1204 				break;
1205 		}
1206 	}
1207 }
1208 EXPORT_SYMBOL(drm_sched_increase_karma);
1209