1 /* $NetBSD: sched_main.c,v 1.11 2021/12/19 12:42:58 riastradh Exp $ */
2
3 /*
4 * Copyright 2015 Advanced Micro Devices, Inc.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26 /**
27 * DOC: Overview
28 *
29 * The GPU scheduler provides entities which allow userspace to push jobs
30 * into software queues which are then scheduled on a hardware run queue.
31 * The software queues have a priority among them. The scheduler selects the entities
32 * from the run queue using a FIFO. The scheduler provides dependency handling
33 * features among jobs. The driver is supposed to provide callback functions for
34 * backend operations to the scheduler like submitting a job to hardware run queue,
35 * returning the dependencies of a job etc.
36 *
37 * The organisation of the scheduler is the following:
38 *
39 * 1. Each hw run queue has one scheduler
40 * 2. Each scheduler has multiple run queues with different priorities
41 * (e.g., HIGH_HW,HIGH_SW, KERNEL, NORMAL)
42 * 3. Each scheduler run queue has a queue of entities to schedule
43 * 4. Entities themselves maintain a queue of jobs that will be scheduled on
44 * the hardware.
45 *
46 * The jobs in a entity are always scheduled in the order that they were pushed.
47 */
48
49 #include <sys/cdefs.h>
50 __KERNEL_RCSID(0, "$NetBSD: sched_main.c,v 1.11 2021/12/19 12:42:58 riastradh Exp $");
51
52 #include <linux/kthread.h>
53 #include <linux/wait.h>
54 #include <linux/sched.h>
55 #include <linux/completion.h>
56 #include <uapi/linux/sched/types.h>
57
58 #include <drm/drm_print.h>
59 #include <drm/gpu_scheduler.h>
60 #include <drm/spsc_queue.h>
61
62 #define CREATE_TRACE_POINTS
63 #include "gpu_scheduler_trace.h"
64
65 #define to_drm_sched_job(sched_job) \
66 container_of((sched_job), struct drm_sched_job, queue_node)
67
68 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
69
70 /**
71 * drm_sched_rq_init - initialize a given run queue struct
72 *
73 * @rq: scheduler run queue
74 *
75 * Initializes a scheduler runqueue.
76 */
drm_sched_rq_init(struct drm_gpu_scheduler * sched,struct drm_sched_rq * rq)77 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
78 struct drm_sched_rq *rq)
79 {
80 spin_lock_init(&rq->lock);
81 INIT_LIST_HEAD(&rq->entities);
82 rq->current_entity = NULL;
83 rq->sched = sched;
84 }
85
86 /**
87 * drm_sched_rq_add_entity - add an entity
88 *
89 * @rq: scheduler run queue
90 * @entity: scheduler entity
91 *
92 * Adds a scheduler entity to the run queue.
93 */
drm_sched_rq_add_entity(struct drm_sched_rq * rq,struct drm_sched_entity * entity)94 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
95 struct drm_sched_entity *entity)
96 {
97 if (!list_empty(&entity->list))
98 return;
99 spin_lock(&rq->lock);
100 atomic_inc(&rq->sched->score);
101 list_add_tail(&entity->list, &rq->entities);
102 spin_unlock(&rq->lock);
103 }
104
105 /**
106 * drm_sched_rq_remove_entity - remove an entity
107 *
108 * @rq: scheduler run queue
109 * @entity: scheduler entity
110 *
111 * Removes a scheduler entity from the run queue.
112 */
drm_sched_rq_remove_entity(struct drm_sched_rq * rq,struct drm_sched_entity * entity)113 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
114 struct drm_sched_entity *entity)
115 {
116 if (list_empty(&entity->list))
117 return;
118 spin_lock(&rq->lock);
119 atomic_dec(&rq->sched->score);
120 list_del_init(&entity->list);
121 if (rq->current_entity == entity)
122 rq->current_entity = NULL;
123 spin_unlock(&rq->lock);
124 }
125
126 /**
127 * drm_sched_rq_select_entity - Select an entity which could provide a job to run
128 *
129 * @rq: scheduler run queue to check.
130 *
131 * Try to find a ready entity, returns NULL if none found.
132 */
133 static struct drm_sched_entity *
drm_sched_rq_select_entity(struct drm_sched_rq * rq)134 drm_sched_rq_select_entity(struct drm_sched_rq *rq)
135 {
136 struct drm_sched_entity *entity;
137
138 spin_lock(&rq->lock);
139
140 entity = rq->current_entity;
141 if (entity) {
142 list_for_each_entry_continue(entity, &rq->entities, list) {
143 if (drm_sched_entity_is_ready(entity)) {
144 rq->current_entity = entity;
145 reinit_completion(&entity->entity_idle);
146 spin_unlock(&rq->lock);
147 return entity;
148 }
149 }
150 }
151
152 list_for_each_entry(entity, &rq->entities, list) {
153
154 if (drm_sched_entity_is_ready(entity)) {
155 rq->current_entity = entity;
156 reinit_completion(&entity->entity_idle);
157 spin_unlock(&rq->lock);
158 return entity;
159 }
160
161 if (entity == rq->current_entity)
162 break;
163 }
164
165 spin_unlock(&rq->lock);
166
167 return NULL;
168 }
169
170 /**
171 * drm_sched_dependency_optimized
172 *
173 * @fence: the dependency fence
174 * @entity: the entity which depends on the above fence
175 *
176 * Returns true if the dependency can be optimized and false otherwise
177 */
drm_sched_dependency_optimized(struct dma_fence * fence,struct drm_sched_entity * entity)178 bool drm_sched_dependency_optimized(struct dma_fence* fence,
179 struct drm_sched_entity *entity)
180 {
181 struct drm_gpu_scheduler *sched = entity->rq->sched;
182 struct drm_sched_fence *s_fence;
183
184 if (!fence || dma_fence_is_signaled(fence))
185 return false;
186 if (fence->context == entity->fence_context)
187 return true;
188 s_fence = to_drm_sched_fence(fence);
189 if (s_fence && s_fence->sched == sched)
190 return true;
191
192 return false;
193 }
194 EXPORT_SYMBOL(drm_sched_dependency_optimized);
195
196 /**
197 * drm_sched_start_timeout - start timeout for reset worker
198 *
199 * @sched: scheduler instance to start the worker for
200 *
201 * Start the timeout for the given scheduler.
202 */
drm_sched_start_timeout(struct drm_gpu_scheduler * sched)203 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
204 {
205 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
206 !list_empty(&sched->ring_mirror_list))
207 schedule_delayed_work(&sched->work_tdr, sched->timeout);
208 }
209
210 /**
211 * drm_sched_fault - immediately start timeout handler
212 *
213 * @sched: scheduler where the timeout handling should be started.
214 *
215 * Start timeout handling immediately when the driver detects a hardware fault.
216 */
drm_sched_fault(struct drm_gpu_scheduler * sched)217 void drm_sched_fault(struct drm_gpu_scheduler *sched)
218 {
219 mod_delayed_work(system_wq, &sched->work_tdr, 0);
220 }
221 EXPORT_SYMBOL(drm_sched_fault);
222
223 /**
224 * drm_sched_suspend_timeout - Suspend scheduler job timeout
225 *
226 * @sched: scheduler instance for which to suspend the timeout
227 *
228 * Suspend the delayed work timeout for the scheduler. This is done by
229 * modifying the delayed work timeout to an arbitrary large value,
230 * MAX_SCHEDULE_TIMEOUT in this case. Note that this function can be
231 * called from an IRQ context.
232 *
233 * Returns the timeout remaining
234 *
235 */
drm_sched_suspend_timeout(struct drm_gpu_scheduler * sched)236 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
237 {
238 #ifdef __NetBSD__
239 /* XXX Currently nothing uses the return value. */
240 cancel_delayed_work(&sched->work_tdr);
241 return -123456789;
242 #else
243 unsigned long sched_timeout, now = jiffies;
244
245 sched_timeout = sched->work_tdr.timer.expires;
246
247 /*
248 * Modify the timeout to an arbitrarily large value. This also prevents
249 * the timeout to be restarted when new submissions arrive
250 */
251 if (mod_delayed_work(system_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
252 && time_after(sched_timeout, now))
253 return sched_timeout - now;
254 else
255 return sched->timeout;
256 #endif
257 }
258 EXPORT_SYMBOL(drm_sched_suspend_timeout);
259
260 /**
261 * drm_sched_resume_timeout - Resume scheduler job timeout
262 *
263 * @sched: scheduler instance for which to resume the timeout
264 * @remaining: remaining timeout
265 *
266 * Resume the delayed work timeout for the scheduler. Note that
267 * this function can be called from an IRQ context.
268 */
drm_sched_resume_timeout(struct drm_gpu_scheduler * sched,unsigned long remaining)269 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
270 unsigned long remaining)
271 {
272 unsigned long flags;
273
274 spin_lock_irqsave(&sched->job_list_lock, flags);
275
276 if (list_empty(&sched->ring_mirror_list))
277 cancel_delayed_work(&sched->work_tdr);
278 else
279 mod_delayed_work(system_wq, &sched->work_tdr, remaining);
280
281 spin_unlock_irqrestore(&sched->job_list_lock, flags);
282 }
283 EXPORT_SYMBOL(drm_sched_resume_timeout);
284
drm_sched_job_begin(struct drm_sched_job * s_job)285 static void drm_sched_job_begin(struct drm_sched_job *s_job)
286 {
287 struct drm_gpu_scheduler *sched = s_job->sched;
288 unsigned long flags;
289
290 spin_lock_irqsave(&sched->job_list_lock, flags);
291 list_add_tail(&s_job->node, &sched->ring_mirror_list);
292 drm_sched_start_timeout(sched);
293 spin_unlock_irqrestore(&sched->job_list_lock, flags);
294 }
295
drm_sched_job_timedout(struct work_struct * work)296 static void drm_sched_job_timedout(struct work_struct *work)
297 {
298 struct drm_gpu_scheduler *sched;
299 struct drm_sched_job *job;
300 unsigned long flags;
301
302 sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
303
304 /* Protects against concurrent deletion in drm_sched_get_cleanup_job */
305 spin_lock_irqsave(&sched->job_list_lock, flags);
306 job = list_first_entry_or_null(&sched->ring_mirror_list,
307 struct drm_sched_job, node);
308
309 if (job) {
310 /*
311 * Remove the bad job so it cannot be freed by concurrent
312 * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
313 * is parked at which point it's safe.
314 */
315 list_del_init(&job->node);
316 spin_unlock_irqrestore(&sched->job_list_lock, flags);
317
318 job->sched->ops->timedout_job(job);
319
320 /*
321 * Guilty job did complete and hence needs to be manually removed
322 * See drm_sched_stop doc.
323 */
324 if (sched->free_guilty) {
325 job->sched->ops->free_job(job);
326 sched->free_guilty = false;
327 }
328 } else {
329 spin_unlock_irqrestore(&sched->job_list_lock, flags);
330 }
331
332 spin_lock_irqsave(&sched->job_list_lock, flags);
333 drm_sched_start_timeout(sched);
334 spin_unlock_irqrestore(&sched->job_list_lock, flags);
335 }
336
337 /**
338 * drm_sched_increase_karma - Update sched_entity guilty flag
339 *
340 * @bad: The job guilty of time out
341 *
342 * Increment on every hang caused by the 'bad' job. If this exceeds the hang
343 * limit of the scheduler then the respective sched entity is marked guilty and
344 * jobs from it will not be scheduled further
345 */
drm_sched_increase_karma(struct drm_sched_job * bad)346 void drm_sched_increase_karma(struct drm_sched_job *bad)
347 {
348 int i;
349 struct drm_sched_entity *tmp;
350 struct drm_sched_entity *entity;
351 struct drm_gpu_scheduler *sched = bad->sched;
352
353 /* don't increase @bad's karma if it's from KERNEL RQ,
354 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
355 * corrupt but keep in mind that kernel jobs always considered good.
356 */
357 if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
358 atomic_inc(&bad->karma);
359 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
360 i++) {
361 struct drm_sched_rq *rq = &sched->sched_rq[i];
362
363 spin_lock(&rq->lock);
364 list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
365 if (bad->s_fence->scheduled.context ==
366 entity->fence_context) {
367 if (atomic_read(&bad->karma) >
368 bad->sched->hang_limit)
369 if (entity->guilty)
370 atomic_set(entity->guilty, 1);
371 break;
372 }
373 }
374 spin_unlock(&rq->lock);
375 if (&entity->list != &rq->entities)
376 break;
377 }
378 }
379 }
380 EXPORT_SYMBOL(drm_sched_increase_karma);
381
382 /**
383 * drm_sched_stop - stop the scheduler
384 *
385 * @sched: scheduler instance
386 * @bad: job which caused the time out
387 *
388 * Stop the scheduler and also removes and frees all completed jobs.
389 * Note: bad job will not be freed as it might be used later and so it's
390 * callers responsibility to release it manually if it's not part of the
391 * mirror list any more.
392 *
393 */
drm_sched_stop(struct drm_gpu_scheduler * sched,struct drm_sched_job * bad)394 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
395 {
396 struct drm_sched_job *s_job, *tmp;
397 unsigned long flags;
398
399 kthread_park(sched->thread);
400
401 /*
402 * Reinsert back the bad job here - now it's safe as
403 * drm_sched_get_cleanup_job cannot race against us and release the
404 * bad job at this point - we parked (waited for) any in progress
405 * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
406 * now until the scheduler thread is unparked.
407 */
408 if (bad && bad->sched == sched)
409 /*
410 * Add at the head of the queue to reflect it was the earliest
411 * job extracted.
412 */
413 list_add(&bad->node, &sched->ring_mirror_list);
414
415 /*
416 * Iterate the job list from later to earlier one and either deactive
417 * their HW callbacks or remove them from mirror list if they already
418 * signaled.
419 * This iteration is thread safe as sched thread is stopped.
420 */
421 list_for_each_entry_safe_reverse(s_job, tmp, &sched->ring_mirror_list, node) {
422 if (s_job->s_fence->parent &&
423 dma_fence_remove_callback(s_job->s_fence->parent,
424 &s_job->cb)) {
425 atomic_dec(&sched->hw_rq_count);
426 } else {
427 /*
428 * remove job from ring_mirror_list.
429 * Locking here is for concurrent resume timeout
430 */
431 spin_lock_irqsave(&sched->job_list_lock, flags);
432 list_del_init(&s_job->node);
433 spin_unlock_irqrestore(&sched->job_list_lock, flags);
434
435 /*
436 * Wait for job's HW fence callback to finish using s_job
437 * before releasing it.
438 *
439 * Job is still alive so fence refcount at least 1
440 */
441 dma_fence_wait(&s_job->s_fence->finished, false);
442
443 /*
444 * We must keep bad job alive for later use during
445 * recovery by some of the drivers but leave a hint
446 * that the guilty job must be released.
447 */
448 if (bad != s_job)
449 sched->ops->free_job(s_job);
450 else
451 sched->free_guilty = true;
452 }
453 }
454
455 /*
456 * Stop pending timer in flight as we rearm it in drm_sched_start. This
457 * avoids the pending timeout work in progress to fire right away after
458 * this TDR finished and before the newly restarted jobs had a
459 * chance to complete.
460 */
461 cancel_delayed_work(&sched->work_tdr);
462 }
463
464 EXPORT_SYMBOL(drm_sched_stop);
465
466 /**
467 * drm_sched_job_recovery - recover jobs after a reset
468 *
469 * @sched: scheduler instance
470 * @full_recovery: proceed with complete sched restart
471 *
472 */
drm_sched_start(struct drm_gpu_scheduler * sched,bool full_recovery)473 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
474 {
475 struct drm_sched_job *s_job, *tmp;
476 unsigned long flags;
477 int r;
478
479 /*
480 * Locking the list is not required here as the sched thread is parked
481 * so no new jobs are being inserted or removed. Also concurrent
482 * GPU recovers can't run in parallel.
483 */
484 list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
485 struct dma_fence *fence = s_job->s_fence->parent;
486
487 atomic_inc(&sched->hw_rq_count);
488
489 if (!full_recovery)
490 continue;
491
492 if (fence) {
493 r = dma_fence_add_callback(fence, &s_job->cb,
494 drm_sched_process_job);
495 if (r == -ENOENT)
496 drm_sched_process_job(fence, &s_job->cb);
497 else if (r)
498 DRM_ERROR("fence add callback failed (%d)\n",
499 r);
500 } else
501 drm_sched_process_job(NULL, &s_job->cb);
502 }
503
504 if (full_recovery) {
505 spin_lock_irqsave(&sched->job_list_lock, flags);
506 drm_sched_start_timeout(sched);
507 spin_unlock_irqrestore(&sched->job_list_lock, flags);
508 }
509
510 kthread_unpark(sched->thread);
511 }
512 EXPORT_SYMBOL(drm_sched_start);
513
514 /**
515 * drm_sched_resubmit_jobs - helper to relunch job from mirror ring list
516 *
517 * @sched: scheduler instance
518 *
519 */
drm_sched_resubmit_jobs(struct drm_gpu_scheduler * sched)520 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
521 {
522 struct drm_sched_job *s_job, *tmp;
523 uint64_t guilty_context;
524 bool found_guilty = false;
525 struct dma_fence *fence;
526
527 list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
528 struct drm_sched_fence *s_fence = s_job->s_fence;
529
530 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
531 found_guilty = true;
532 guilty_context = s_job->s_fence->scheduled.context;
533 }
534
535 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
536 dma_fence_set_error(&s_fence->finished, -ECANCELED);
537
538 dma_fence_put(s_job->s_fence->parent);
539 fence = sched->ops->run_job(s_job);
540
541 if (IS_ERR_OR_NULL(fence)) {
542 if (IS_ERR(fence))
543 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
544
545 s_job->s_fence->parent = NULL;
546 } else {
547 s_job->s_fence->parent = fence;
548 }
549
550
551 }
552 }
553 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
554
555 /**
556 * drm_sched_job_init - init a scheduler job
557 *
558 * @job: scheduler job to init
559 * @entity: scheduler entity to use
560 * @owner: job owner for debugging
561 *
562 * Refer to drm_sched_entity_push_job() documentation
563 * for locking considerations.
564 *
565 * Returns 0 for success, negative error code otherwise.
566 */
drm_sched_job_init(struct drm_sched_job * job,struct drm_sched_entity * entity,void * owner)567 int drm_sched_job_init(struct drm_sched_job *job,
568 struct drm_sched_entity *entity,
569 void *owner)
570 {
571 struct drm_gpu_scheduler *sched;
572
573 drm_sched_entity_select_rq(entity);
574 if (!entity->rq)
575 return -ENOENT;
576
577 sched = entity->rq->sched;
578
579 job->sched = sched;
580 job->entity = entity;
581 job->s_priority = entity->rq - sched->sched_rq;
582 job->s_fence = drm_sched_fence_create(entity, owner);
583 if (!job->s_fence)
584 return -ENOMEM;
585 job->id = atomic64_inc_return(&sched->job_id_count);
586
587 INIT_LIST_HEAD(&job->node);
588
589 return 0;
590 }
591 EXPORT_SYMBOL(drm_sched_job_init);
592
593 /**
594 * drm_sched_job_cleanup - clean up scheduler job resources
595 *
596 * @job: scheduler job to clean up
597 */
drm_sched_job_cleanup(struct drm_sched_job * job)598 void drm_sched_job_cleanup(struct drm_sched_job *job)
599 {
600 dma_fence_put(&job->s_fence->finished);
601 job->s_fence = NULL;
602 }
603 EXPORT_SYMBOL(drm_sched_job_cleanup);
604
605 /**
606 * drm_sched_ready - is the scheduler ready
607 *
608 * @sched: scheduler instance
609 *
610 * Return true if we can push more jobs to the hw, otherwise false.
611 */
drm_sched_ready(struct drm_gpu_scheduler * sched)612 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
613 {
614 return atomic_read(&sched->hw_rq_count) <
615 sched->hw_submission_limit;
616 }
617
618 /**
619 * drm_sched_wakeup - Wake up the scheduler when it is ready
620 *
621 * @sched: scheduler instance
622 *
623 */
drm_sched_wakeup(struct drm_gpu_scheduler * sched)624 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
625 {
626 assert_spin_locked(&sched->job_list_lock);
627 if (drm_sched_ready(sched))
628 DRM_SPIN_WAKEUP_ONE(&sched->wake_up_worker,
629 &sched->job_list_lock);
630 }
631
632 /**
633 * drm_sched_select_entity - Select next entity to process
634 *
635 * @sched: scheduler instance
636 *
637 * Returns the entity to process or NULL if none are found.
638 */
639 static struct drm_sched_entity *
drm_sched_select_entity(struct drm_gpu_scheduler * sched)640 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
641 {
642 struct drm_sched_entity *entity;
643 int i;
644
645 if (!drm_sched_ready(sched))
646 return NULL;
647
648 /* Kernel run queue has higher priority than normal run queue*/
649 for (i = DRM_SCHED_PRIORITY_MAX - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
650 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
651 if (entity)
652 break;
653 }
654
655 return entity;
656 }
657
658 /**
659 * drm_sched_process_job - process a job
660 *
661 * @f: fence
662 * @cb: fence callbacks
663 *
664 * Called after job has finished execution.
665 */
drm_sched_process_job(struct dma_fence * f,struct dma_fence_cb * cb)666 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
667 {
668 struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
669 struct drm_sched_fence *s_fence = s_job->s_fence;
670 struct drm_gpu_scheduler *sched = s_fence->sched;
671
672 atomic_dec(&sched->hw_rq_count);
673 atomic_dec(&sched->score);
674
675 trace_drm_sched_process_job(s_fence);
676
677 drm_sched_fence_finished(s_fence);
678 spin_lock(&sched->job_list_lock);
679 DRM_SPIN_WAKEUP_ONE(&sched->wake_up_worker, &sched->job_list_lock);
680 spin_unlock(&sched->job_list_lock);
681 }
682
683 /**
684 * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
685 *
686 * @sched: scheduler instance
687 *
688 * Returns the next finished job from the mirror list (if there is one)
689 * ready for it to be destroyed.
690 */
691 static struct drm_sched_job *
drm_sched_get_cleanup_job(struct drm_gpu_scheduler * sched)692 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
693 {
694 struct drm_sched_job *job;
695
696 assert_spin_locked(&sched->job_list_lock);
697
698 /*
699 * Don't destroy jobs while the timeout worker is running OR thread
700 * is being parked and hence assumed to not touch ring_mirror_list
701 */
702 if ((sched->timeout != MAX_SCHEDULE_TIMEOUT &&
703 !cancel_delayed_work(&sched->work_tdr)) ||
704 __kthread_should_park(sched->thread))
705 return NULL;
706
707 job = list_first_entry_or_null(&sched->ring_mirror_list,
708 struct drm_sched_job, node);
709
710 if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
711 /* remove job from ring_mirror_list */
712 list_del_init(&job->node);
713 } else {
714 job = NULL;
715 /* queue timeout for next job */
716 drm_sched_start_timeout(sched);
717 }
718
719 return job;
720 }
721
722 /**
723 * drm_sched_blocked - check if the scheduler is blocked
724 *
725 * @sched: scheduler instance
726 *
727 * Returns true if blocked, otherwise false.
728 */
drm_sched_blocked(struct drm_gpu_scheduler * sched)729 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
730 {
731 if (kthread_should_park()) {
732 kthread_parkme();
733 return true;
734 }
735
736 return false;
737 }
738
739 /**
740 * drm_sched_main - main scheduler thread
741 *
742 * @param: scheduler instance
743 *
744 * Returns 0.
745 */
drm_sched_main(void * param)746 static int drm_sched_main(void *param)
747 {
748 struct sched_param sparam = {.sched_priority = 1};
749 struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
750 int r;
751
752 sched_setscheduler(current, SCHED_FIFO, &sparam);
753
754 /* Wait for sched->thread to be initialized by drm_sched_init. */
755 spin_lock(&sched->job_list_lock);
756 DRM_SPIN_WAIT_UNTIL(r, &sched->wake_up_worker, &sched->job_list_lock,
757 sched->thread != NULL);
758 spin_unlock(&sched->job_list_lock);
759
760 while (!kthread_should_stop()) {
761 struct drm_sched_entity *entity = NULL;
762 struct drm_sched_fence *s_fence;
763 struct drm_sched_job *sched_job;
764 struct dma_fence *fence;
765 struct drm_sched_job *cleanup_job = NULL;
766
767 spin_lock(&sched->job_list_lock);
768 DRM_SPIN_WAIT_UNTIL(r, &sched->wake_up_worker,
769 &sched->job_list_lock,
770 ((cleanup_job = drm_sched_get_cleanup_job(sched)) ||
771 (!drm_sched_blocked(sched) &&
772 (entity = drm_sched_select_entity(sched))) ||
773 kthread_should_stop()));
774 spin_unlock(&sched->job_list_lock);
775
776 if (cleanup_job) {
777 sched->ops->free_job(cleanup_job);
778 /* queue timeout for next job */
779 drm_sched_start_timeout(sched);
780 }
781
782 if (!entity)
783 continue;
784
785 sched_job = drm_sched_entity_pop_job(entity);
786
787 complete(&entity->entity_idle);
788
789 if (!sched_job)
790 continue;
791
792 s_fence = sched_job->s_fence;
793
794 atomic_inc(&sched->hw_rq_count);
795 drm_sched_job_begin(sched_job);
796
797 fence = sched->ops->run_job(sched_job);
798 drm_sched_fence_scheduled(s_fence);
799
800 if (!IS_ERR_OR_NULL(fence)) {
801 s_fence->parent = dma_fence_get(fence);
802 r = dma_fence_add_callback(fence, &sched_job->cb,
803 drm_sched_process_job);
804 if (r == -ENOENT)
805 drm_sched_process_job(fence, &sched_job->cb);
806 else if (r)
807 DRM_ERROR("fence add callback failed (%d)\n",
808 r);
809 dma_fence_put(fence);
810 } else {
811 if (IS_ERR(fence))
812 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
813
814 drm_sched_process_job(NULL, &sched_job->cb);
815 }
816
817 spin_lock(&sched->job_list_lock);
818 DRM_SPIN_WAKEUP_ONE(&sched->job_scheduled,
819 &sched->job_list_lock);
820 spin_unlock(&sched->job_list_lock);
821 }
822 return 0;
823 }
824
825 /**
826 * drm_sched_init - Init a gpu scheduler instance
827 *
828 * @sched: scheduler instance
829 * @ops: backend operations for this scheduler
830 * @hw_submission: number of hw submissions that can be in flight
831 * @hang_limit: number of times to allow a job to hang before dropping it
832 * @timeout: timeout value in jiffies for the scheduler
833 * @name: name used for debugging
834 *
835 * Return 0 on success, otherwise error code.
836 */
drm_sched_init(struct drm_gpu_scheduler * sched,const struct drm_sched_backend_ops * ops,unsigned hw_submission,unsigned hang_limit,long timeout,const char * name)837 int drm_sched_init(struct drm_gpu_scheduler *sched,
838 const struct drm_sched_backend_ops *ops,
839 unsigned hw_submission,
840 unsigned hang_limit,
841 long timeout,
842 const char *name)
843 {
844 int i, ret;
845 sched->ops = ops;
846 sched->hw_submission_limit = hw_submission;
847 sched->name = name;
848 sched->timeout = timeout;
849 sched->hang_limit = hang_limit;
850 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++)
851 drm_sched_rq_init(sched, &sched->sched_rq[i]);
852
853 DRM_INIT_WAITQUEUE(&sched->wake_up_worker, "drmschedw");
854 DRM_INIT_WAITQUEUE(&sched->job_scheduled, "drmschedj");
855 INIT_LIST_HEAD(&sched->ring_mirror_list);
856 spin_lock_init(&sched->job_list_lock);
857 atomic_set(&sched->hw_rq_count, 0);
858 INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
859 atomic_set(&sched->score, 0);
860 atomic64_set(&sched->job_id_count, 0);
861
862 /* Each scheduler will run on a seperate kernel thread */
863 struct task_struct *thread =
864 kthread_run(drm_sched_main, sched, sched->name,
865 &sched->job_list_lock, &sched->wake_up_worker);
866 if (IS_ERR(thread)) {
867 ret = PTR_ERR(thread);
868 sched->thread = NULL;
869 DRM_ERROR("Failed to create scheduler for %s.\n", name);
870 return ret;
871 }
872 spin_lock(&sched->job_list_lock);
873 sched->thread = thread;
874 DRM_SPIN_WAKEUP_ALL(&sched->wake_up_worker, &sched->job_list_lock);
875 spin_unlock(&sched->job_list_lock);
876
877 sched->ready = true;
878 return 0;
879 }
880 EXPORT_SYMBOL(drm_sched_init);
881
882 /**
883 * drm_sched_fini - Destroy a gpu scheduler
884 *
885 * @sched: scheduler instance
886 *
887 * Tears down and cleans up the scheduler.
888 */
drm_sched_fini(struct drm_gpu_scheduler * sched)889 void drm_sched_fini(struct drm_gpu_scheduler *sched)
890 {
891 if (sched->thread)
892 kthread_stop(sched->thread);
893
894 sched->ready = false;
895
896 spin_lock_destroy(&sched->job_list_lock);
897 DRM_DESTROY_WAITQUEUE(&sched->job_scheduled);
898 DRM_DESTROY_WAITQUEUE(&sched->wake_up_worker);
899
900 for (int i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++) {
901 struct drm_sched_rq *rq = &sched->sched_rq[i];
902 spin_lock_destroy(&rq->lock);
903 }
904 }
905 EXPORT_SYMBOL(drm_sched_fini);
906