Loading...
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 * Optionally, drivers may register a callback (update_job_credits) provided by
69 * struct drm_sched_backend_ops to update the job's credits dynamically. The
70 * scheduler executes this callback every time the scheduler considers a job for
71 * execution and subsequently checks whether the job fits the scheduler's credit
72 * limit.
73 */
74
75#include <linux/wait.h>
76#include <linux/sched.h>
77#include <linux/completion.h>
78#include <linux/dma-resv.h>
79#include <uapi/linux/sched/types.h>
80
81#include <drm/drm_print.h>
82#include <drm/drm_gem.h>
83#include <drm/drm_syncobj.h>
84#include <drm/gpu_scheduler.h>
85#include <drm/spsc_queue.h>
86
87#define CREATE_TRACE_POINTS
88#include "gpu_scheduler_trace.h"
89
90#ifdef CONFIG_LOCKDEP
91static struct lockdep_map drm_sched_lockdep_map = {
92 .name = "drm_sched_lockdep_map"
93};
94#endif
95
96#define to_drm_sched_job(sched_job) \
97 container_of((sched_job), struct drm_sched_job, queue_node)
98
99int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
100
101/**
102 * DOC: sched_policy (int)
103 * Used to override default entities scheduling policy in a run queue.
104 */
105MODULE_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).");
106module_param_named(sched_policy, drm_sched_policy, int, 0444);
107
108static u32 drm_sched_available_credits(struct drm_gpu_scheduler *sched)
109{
110 u32 credits;
111
112 drm_WARN_ON(sched, check_sub_overflow(sched->credit_limit,
113 atomic_read(&sched->credit_count),
114 &credits));
115
116 return credits;
117}
118
119/**
120 * drm_sched_can_queue -- Can we queue more to the hardware?
121 * @sched: scheduler instance
122 * @entity: the scheduler entity
123 *
124 * Return true if we can push at least one more job from @entity, false
125 * otherwise.
126 */
127static bool drm_sched_can_queue(struct drm_gpu_scheduler *sched,
128 struct drm_sched_entity *entity)
129{
130 struct drm_sched_job *s_job;
131
132 s_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
133 if (!s_job)
134 return false;
135
136 if (sched->ops->update_job_credits) {
137 s_job->credits = sched->ops->update_job_credits(s_job);
138
139 drm_WARN(sched, !s_job->credits,
140 "Jobs with zero credits bypass job-flow control.\n");
141 }
142
143 /* If a job exceeds the credit limit, truncate it to the credit limit
144 * itself to guarantee forward progress.
145 */
146 if (drm_WARN(sched, s_job->credits > sched->credit_limit,
147 "Jobs may not exceed the credit limit, truncate.\n"))
148 s_job->credits = sched->credit_limit;
149
150 return drm_sched_available_credits(sched) >= s_job->credits;
151}
152
153static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
154 const struct rb_node *b)
155{
156 struct drm_sched_entity *ent_a = rb_entry((a), struct drm_sched_entity, rb_tree_node);
157 struct drm_sched_entity *ent_b = rb_entry((b), struct drm_sched_entity, rb_tree_node);
158
159 return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
160}
161
162static void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity,
163 struct drm_sched_rq *rq)
164{
165 if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
166 rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
167 RB_CLEAR_NODE(&entity->rb_tree_node);
168 }
169}
170
171void drm_sched_rq_update_fifo_locked(struct drm_sched_entity *entity,
172 struct drm_sched_rq *rq,
173 ktime_t ts)
174{
175 /*
176 * Both locks need to be grabbed, one to protect from entity->rq change
177 * for entity from within concurrent drm_sched_entity_select_rq and the
178 * other to update the rb tree structure.
179 */
180 lockdep_assert_held(&entity->lock);
181 lockdep_assert_held(&rq->lock);
182
183 drm_sched_rq_remove_fifo_locked(entity, rq);
184
185 entity->oldest_job_waiting = ts;
186
187 rb_add_cached(&entity->rb_tree_node, &rq->rb_tree_root,
188 drm_sched_entity_compare_before);
189}
190
191/**
192 * drm_sched_rq_init - initialize a given run queue struct
193 *
194 * @sched: scheduler instance to associate with this run queue
195 * @rq: scheduler run queue
196 *
197 * Initializes a scheduler runqueue.
198 */
199static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
200 struct drm_sched_rq *rq)
201{
202 spin_lock_init(&rq->lock);
203 INIT_LIST_HEAD(&rq->entities);
204 rq->rb_tree_root = RB_ROOT_CACHED;
205 rq->current_entity = NULL;
206 rq->sched = sched;
207}
208
209/**
210 * drm_sched_rq_add_entity - add an entity
211 *
212 * @rq: scheduler run queue
213 * @entity: scheduler entity
214 *
215 * Adds a scheduler entity to the run queue.
216 */
217void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
218 struct drm_sched_entity *entity)
219{
220 lockdep_assert_held(&entity->lock);
221 lockdep_assert_held(&rq->lock);
222
223 if (!list_empty(&entity->list))
224 return;
225
226 atomic_inc(rq->sched->score);
227 list_add_tail(&entity->list, &rq->entities);
228}
229
230/**
231 * drm_sched_rq_remove_entity - remove an entity
232 *
233 * @rq: scheduler run queue
234 * @entity: scheduler entity
235 *
236 * Removes a scheduler entity from the run queue.
237 */
238void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
239 struct drm_sched_entity *entity)
240{
241 lockdep_assert_held(&entity->lock);
242
243 if (list_empty(&entity->list))
244 return;
245
246 spin_lock(&rq->lock);
247
248 atomic_dec(rq->sched->score);
249 list_del_init(&entity->list);
250
251 if (rq->current_entity == entity)
252 rq->current_entity = NULL;
253
254 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
255 drm_sched_rq_remove_fifo_locked(entity, rq);
256
257 spin_unlock(&rq->lock);
258}
259
260/**
261 * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
262 *
263 * @sched: the gpu scheduler
264 * @rq: scheduler run queue to check.
265 *
266 * Try to find the next ready entity.
267 *
268 * Return an entity if one is found; return an error-pointer (!NULL) if an
269 * entity was ready, but the scheduler had insufficient credits to accommodate
270 * its job; return NULL, if no ready entity was found.
271 */
272static struct drm_sched_entity *
273drm_sched_rq_select_entity_rr(struct drm_gpu_scheduler *sched,
274 struct drm_sched_rq *rq)
275{
276 struct drm_sched_entity *entity;
277
278 spin_lock(&rq->lock);
279
280 entity = rq->current_entity;
281 if (entity) {
282 list_for_each_entry_continue(entity, &rq->entities, list) {
283 if (drm_sched_entity_is_ready(entity)) {
284 /* If we can't queue yet, preserve the current
285 * entity in terms of fairness.
286 */
287 if (!drm_sched_can_queue(sched, entity)) {
288 spin_unlock(&rq->lock);
289 return ERR_PTR(-ENOSPC);
290 }
291
292 rq->current_entity = entity;
293 reinit_completion(&entity->entity_idle);
294 spin_unlock(&rq->lock);
295 return entity;
296 }
297 }
298 }
299
300 list_for_each_entry(entity, &rq->entities, list) {
301 if (drm_sched_entity_is_ready(entity)) {
302 /* If we can't queue yet, preserve the current entity in
303 * terms of fairness.
304 */
305 if (!drm_sched_can_queue(sched, entity)) {
306 spin_unlock(&rq->lock);
307 return ERR_PTR(-ENOSPC);
308 }
309
310 rq->current_entity = entity;
311 reinit_completion(&entity->entity_idle);
312 spin_unlock(&rq->lock);
313 return entity;
314 }
315
316 if (entity == rq->current_entity)
317 break;
318 }
319
320 spin_unlock(&rq->lock);
321
322 return NULL;
323}
324
325/**
326 * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
327 *
328 * @sched: the gpu scheduler
329 * @rq: scheduler run queue to check.
330 *
331 * Find oldest waiting ready entity.
332 *
333 * Return an entity if one is found; return an error-pointer (!NULL) if an
334 * entity was ready, but the scheduler had insufficient credits to accommodate
335 * its job; return NULL, if no ready entity was found.
336 */
337static struct drm_sched_entity *
338drm_sched_rq_select_entity_fifo(struct drm_gpu_scheduler *sched,
339 struct drm_sched_rq *rq)
340{
341 struct rb_node *rb;
342
343 spin_lock(&rq->lock);
344 for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
345 struct drm_sched_entity *entity;
346
347 entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
348 if (drm_sched_entity_is_ready(entity)) {
349 /* If we can't queue yet, preserve the current entity in
350 * terms of fairness.
351 */
352 if (!drm_sched_can_queue(sched, entity)) {
353 spin_unlock(&rq->lock);
354 return ERR_PTR(-ENOSPC);
355 }
356
357 reinit_completion(&entity->entity_idle);
358 break;
359 }
360 }
361 spin_unlock(&rq->lock);
362
363 return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
364}
365
366/**
367 * drm_sched_run_job_queue - enqueue run-job work
368 * @sched: scheduler instance
369 */
370static void drm_sched_run_job_queue(struct drm_gpu_scheduler *sched)
371{
372 if (!READ_ONCE(sched->pause_submit))
373 queue_work(sched->submit_wq, &sched->work_run_job);
374}
375
376/**
377 * __drm_sched_run_free_queue - enqueue free-job work
378 * @sched: scheduler instance
379 */
380static void __drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
381{
382 if (!READ_ONCE(sched->pause_submit))
383 queue_work(sched->submit_wq, &sched->work_free_job);
384}
385
386/**
387 * drm_sched_run_free_queue - enqueue free-job work if ready
388 * @sched: scheduler instance
389 */
390static void drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
391{
392 struct drm_sched_job *job;
393
394 spin_lock(&sched->job_list_lock);
395 job = list_first_entry_or_null(&sched->pending_list,
396 struct drm_sched_job, list);
397 if (job && dma_fence_is_signaled(&job->s_fence->finished))
398 __drm_sched_run_free_queue(sched);
399 spin_unlock(&sched->job_list_lock);
400}
401
402/**
403 * drm_sched_job_done - complete a job
404 * @s_job: pointer to the job which is done
405 *
406 * Finish the job's fence and wake up the worker thread.
407 */
408static void drm_sched_job_done(struct drm_sched_job *s_job, int result)
409{
410 struct drm_sched_fence *s_fence = s_job->s_fence;
411 struct drm_gpu_scheduler *sched = s_fence->sched;
412
413 atomic_sub(s_job->credits, &sched->credit_count);
414 atomic_dec(sched->score);
415
416 trace_drm_sched_process_job(s_fence);
417
418 dma_fence_get(&s_fence->finished);
419 drm_sched_fence_finished(s_fence, result);
420 dma_fence_put(&s_fence->finished);
421 __drm_sched_run_free_queue(sched);
422}
423
424/**
425 * drm_sched_job_done_cb - the callback for a done job
426 * @f: fence
427 * @cb: fence callbacks
428 */
429static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
430{
431 struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
432
433 drm_sched_job_done(s_job, f->error);
434}
435
436/**
437 * drm_sched_start_timeout - start timeout for reset worker
438 *
439 * @sched: scheduler instance to start the worker for
440 *
441 * Start the timeout for the given scheduler.
442 */
443static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
444{
445 lockdep_assert_held(&sched->job_list_lock);
446
447 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
448 !list_empty(&sched->pending_list))
449 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
450}
451
452static void drm_sched_start_timeout_unlocked(struct drm_gpu_scheduler *sched)
453{
454 spin_lock(&sched->job_list_lock);
455 drm_sched_start_timeout(sched);
456 spin_unlock(&sched->job_list_lock);
457}
458
459/**
460 * drm_sched_tdr_queue_imm: - immediately start job timeout handler
461 *
462 * @sched: scheduler for which the timeout handling should be started.
463 *
464 * Start timeout handling immediately for the named scheduler.
465 */
466void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched)
467{
468 spin_lock(&sched->job_list_lock);
469 sched->timeout = 0;
470 drm_sched_start_timeout(sched);
471 spin_unlock(&sched->job_list_lock);
472}
473EXPORT_SYMBOL(drm_sched_tdr_queue_imm);
474
475/**
476 * drm_sched_fault - immediately start timeout handler
477 *
478 * @sched: scheduler where the timeout handling should be started.
479 *
480 * Start timeout handling immediately when the driver detects a hardware fault.
481 */
482void drm_sched_fault(struct drm_gpu_scheduler *sched)
483{
484 if (sched->timeout_wq)
485 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
486}
487EXPORT_SYMBOL(drm_sched_fault);
488
489/**
490 * drm_sched_suspend_timeout - Suspend scheduler job timeout
491 *
492 * @sched: scheduler instance for which to suspend the timeout
493 *
494 * Suspend the delayed work timeout for the scheduler. This is done by
495 * modifying the delayed work timeout to an arbitrary large value,
496 * MAX_SCHEDULE_TIMEOUT in this case.
497 *
498 * Returns the timeout remaining
499 *
500 */
501unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
502{
503 unsigned long sched_timeout, now = jiffies;
504
505 sched_timeout = sched->work_tdr.timer.expires;
506
507 /*
508 * Modify the timeout to an arbitrarily large value. This also prevents
509 * the timeout to be restarted when new submissions arrive
510 */
511 if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
512 && time_after(sched_timeout, now))
513 return sched_timeout - now;
514 else
515 return sched->timeout;
516}
517EXPORT_SYMBOL(drm_sched_suspend_timeout);
518
519/**
520 * drm_sched_resume_timeout - Resume scheduler job timeout
521 *
522 * @sched: scheduler instance for which to resume the timeout
523 * @remaining: remaining timeout
524 *
525 * Resume the delayed work timeout for the scheduler.
526 */
527void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
528 unsigned long remaining)
529{
530 spin_lock(&sched->job_list_lock);
531
532 if (list_empty(&sched->pending_list))
533 cancel_delayed_work(&sched->work_tdr);
534 else
535 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
536
537 spin_unlock(&sched->job_list_lock);
538}
539EXPORT_SYMBOL(drm_sched_resume_timeout);
540
541static void drm_sched_job_begin(struct drm_sched_job *s_job)
542{
543 struct drm_gpu_scheduler *sched = s_job->sched;
544
545 spin_lock(&sched->job_list_lock);
546 list_add_tail(&s_job->list, &sched->pending_list);
547 drm_sched_start_timeout(sched);
548 spin_unlock(&sched->job_list_lock);
549}
550
551static void drm_sched_job_timedout(struct work_struct *work)
552{
553 struct drm_gpu_scheduler *sched;
554 struct drm_sched_job *job;
555 enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
556
557 sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
558
559 /* Protects against concurrent deletion in drm_sched_get_finished_job */
560 spin_lock(&sched->job_list_lock);
561 job = list_first_entry_or_null(&sched->pending_list,
562 struct drm_sched_job, list);
563
564 if (job) {
565 /*
566 * Remove the bad job so it cannot be freed by concurrent
567 * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
568 * is parked at which point it's safe.
569 */
570 list_del_init(&job->list);
571 spin_unlock(&sched->job_list_lock);
572
573 status = job->sched->ops->timedout_job(job);
574
575 /*
576 * Guilty job did complete and hence needs to be manually removed
577 * See drm_sched_stop doc.
578 */
579 if (sched->free_guilty) {
580 job->sched->ops->free_job(job);
581 sched->free_guilty = false;
582 }
583 } else {
584 spin_unlock(&sched->job_list_lock);
585 }
586
587 if (status != DRM_GPU_SCHED_STAT_ENODEV)
588 drm_sched_start_timeout_unlocked(sched);
589}
590
591/**
592 * drm_sched_stop - stop the scheduler
593 *
594 * @sched: scheduler instance
595 * @bad: job which caused the time out
596 *
597 * Stop the scheduler and also removes and frees all completed jobs.
598 * Note: bad job will not be freed as it might be used later and so it's
599 * callers responsibility to release it manually if it's not part of the
600 * pending list any more.
601 *
602 * This function is typically used for reset recovery (see the docu of
603 * drm_sched_backend_ops.timedout_job() for details). Do not call it for
604 * scheduler teardown, i.e., before calling drm_sched_fini().
605 */
606void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
607{
608 struct drm_sched_job *s_job, *tmp;
609
610 drm_sched_wqueue_stop(sched);
611
612 /*
613 * Reinsert back the bad job here - now it's safe as
614 * drm_sched_get_finished_job cannot race against us and release the
615 * bad job at this point - we parked (waited for) any in progress
616 * (earlier) cleanups and drm_sched_get_finished_job will not be called
617 * now until the scheduler thread is unparked.
618 */
619 if (bad && bad->sched == sched)
620 /*
621 * Add at the head of the queue to reflect it was the earliest
622 * job extracted.
623 */
624 list_add(&bad->list, &sched->pending_list);
625
626 /*
627 * Iterate the job list from later to earlier one and either deactive
628 * their HW callbacks or remove them from pending list if they already
629 * signaled.
630 * This iteration is thread safe as sched thread is stopped.
631 */
632 list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
633 list) {
634 if (s_job->s_fence->parent &&
635 dma_fence_remove_callback(s_job->s_fence->parent,
636 &s_job->cb)) {
637 dma_fence_put(s_job->s_fence->parent);
638 s_job->s_fence->parent = NULL;
639 atomic_sub(s_job->credits, &sched->credit_count);
640 } else {
641 /*
642 * remove job from pending_list.
643 * Locking here is for concurrent resume timeout
644 */
645 spin_lock(&sched->job_list_lock);
646 list_del_init(&s_job->list);
647 spin_unlock(&sched->job_list_lock);
648
649 /*
650 * Wait for job's HW fence callback to finish using s_job
651 * before releasing it.
652 *
653 * Job is still alive so fence refcount at least 1
654 */
655 dma_fence_wait(&s_job->s_fence->finished, false);
656
657 /*
658 * We must keep bad job alive for later use during
659 * recovery by some of the drivers but leave a hint
660 * that the guilty job must be released.
661 */
662 if (bad != s_job)
663 sched->ops->free_job(s_job);
664 else
665 sched->free_guilty = true;
666 }
667 }
668
669 /*
670 * Stop pending timer in flight as we rearm it in drm_sched_start. This
671 * avoids the pending timeout work in progress to fire right away after
672 * this TDR finished and before the newly restarted jobs had a
673 * chance to complete.
674 */
675 cancel_delayed_work(&sched->work_tdr);
676}
677EXPORT_SYMBOL(drm_sched_stop);
678
679/**
680 * drm_sched_start - recover jobs after a reset
681 *
682 * @sched: scheduler instance
683 * @errno: error to set on the pending fences
684 *
685 * This function is typically used for reset recovery (see the docu of
686 * drm_sched_backend_ops.timedout_job() for details). Do not call it for
687 * scheduler startup. The scheduler itself is fully operational after
688 * drm_sched_init() succeeded.
689 */
690void drm_sched_start(struct drm_gpu_scheduler *sched, int errno)
691{
692 struct drm_sched_job *s_job, *tmp;
693
694 /*
695 * Locking the list is not required here as the sched thread is parked
696 * so no new jobs are being inserted or removed. Also concurrent
697 * GPU recovers can't run in parallel.
698 */
699 list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
700 struct dma_fence *fence = s_job->s_fence->parent;
701
702 atomic_add(s_job->credits, &sched->credit_count);
703
704 if (!fence) {
705 drm_sched_job_done(s_job, errno ?: -ECANCELED);
706 continue;
707 }
708
709 if (dma_fence_add_callback(fence, &s_job->cb,
710 drm_sched_job_done_cb))
711 drm_sched_job_done(s_job, fence->error ?: errno);
712 }
713
714 drm_sched_start_timeout_unlocked(sched);
715 drm_sched_wqueue_start(sched);
716}
717EXPORT_SYMBOL(drm_sched_start);
718
719/**
720 * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
721 *
722 * @sched: scheduler instance
723 *
724 * Re-submitting jobs was a concept AMD came up as cheap way to implement
725 * recovery after a job timeout.
726 *
727 * This turned out to be not working very well. First of all there are many
728 * problem with the dma_fence implementation and requirements. Either the
729 * implementation is risking deadlocks with core memory management or violating
730 * documented implementation details of the dma_fence object.
731 *
732 * Drivers can still save and restore their state for recovery operations, but
733 * we shouldn't make this a general scheduler feature around the dma_fence
734 * interface.
735 */
736void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
737{
738 struct drm_sched_job *s_job, *tmp;
739 uint64_t guilty_context;
740 bool found_guilty = false;
741 struct dma_fence *fence;
742
743 list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
744 struct drm_sched_fence *s_fence = s_job->s_fence;
745
746 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
747 found_guilty = true;
748 guilty_context = s_job->s_fence->scheduled.context;
749 }
750
751 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
752 dma_fence_set_error(&s_fence->finished, -ECANCELED);
753
754 fence = sched->ops->run_job(s_job);
755
756 if (IS_ERR_OR_NULL(fence)) {
757 if (IS_ERR(fence))
758 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
759
760 s_job->s_fence->parent = NULL;
761 } else {
762
763 s_job->s_fence->parent = dma_fence_get(fence);
764
765 /* Drop for orignal kref_init */
766 dma_fence_put(fence);
767 }
768 }
769}
770EXPORT_SYMBOL(drm_sched_resubmit_jobs);
771
772/**
773 * drm_sched_job_init - init a scheduler job
774 * @job: scheduler job to init
775 * @entity: scheduler entity to use
776 * @credits: the number of credits this job contributes to the schedulers
777 * credit limit
778 * @owner: job owner for debugging
779 *
780 * Refer to drm_sched_entity_push_job() documentation
781 * for locking considerations.
782 *
783 * Drivers must make sure drm_sched_job_cleanup() if this function returns
784 * successfully, even when @job is aborted before drm_sched_job_arm() is called.
785 *
786 * Note that this function does not assign a valid value to each struct member
787 * of struct drm_sched_job. Take a look at that struct's documentation to see
788 * who sets which struct member with what lifetime.
789 *
790 * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
791 * has died, which can mean that there's no valid runqueue for a @entity.
792 * This function returns -ENOENT in this case (which probably should be -EIO as
793 * a more meanigful return value).
794 *
795 * Returns 0 for success, negative error code otherwise.
796 */
797int drm_sched_job_init(struct drm_sched_job *job,
798 struct drm_sched_entity *entity,
799 u32 credits, void *owner)
800{
801 if (!entity->rq) {
802 /* This will most likely be followed by missing frames
803 * or worse--a blank screen--leave a trail in the
804 * logs, so this can be debugged easier.
805 */
806 drm_err(job->sched, "%s: entity has no rq!\n", __func__);
807 return -ENOENT;
808 }
809
810 if (unlikely(!credits)) {
811 pr_err("*ERROR* %s: credits cannot be 0!\n", __func__);
812 return -EINVAL;
813 }
814
815 /*
816 * We don't know for sure how the user has allocated. Thus, zero the
817 * struct so that unallowed (i.e., too early) usage of pointers that
818 * this function does not set is guaranteed to lead to a NULL pointer
819 * exception instead of UB.
820 */
821 memset(job, 0, sizeof(*job));
822
823 job->entity = entity;
824 job->credits = credits;
825 job->s_fence = drm_sched_fence_alloc(entity, owner);
826 if (!job->s_fence)
827 return -ENOMEM;
828
829 INIT_LIST_HEAD(&job->list);
830
831 xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
832
833 return 0;
834}
835EXPORT_SYMBOL(drm_sched_job_init);
836
837/**
838 * drm_sched_job_arm - arm a scheduler job for execution
839 * @job: scheduler job to arm
840 *
841 * This arms a scheduler job for execution. Specifically it initializes the
842 * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
843 * or other places that need to track the completion of this job.
844 *
845 * Refer to drm_sched_entity_push_job() documentation for locking
846 * considerations.
847 *
848 * This can only be called if drm_sched_job_init() succeeded.
849 */
850void drm_sched_job_arm(struct drm_sched_job *job)
851{
852 struct drm_gpu_scheduler *sched;
853 struct drm_sched_entity *entity = job->entity;
854
855 BUG_ON(!entity);
856 drm_sched_entity_select_rq(entity);
857 sched = entity->rq->sched;
858
859 job->sched = sched;
860 job->s_priority = entity->priority;
861 job->id = atomic64_inc_return(&sched->job_id_count);
862
863 drm_sched_fence_init(job->s_fence, job->entity);
864}
865EXPORT_SYMBOL(drm_sched_job_arm);
866
867/**
868 * drm_sched_job_add_dependency - adds the fence as a job dependency
869 * @job: scheduler job to add the dependencies to
870 * @fence: the dma_fence to add to the list of dependencies.
871 *
872 * Note that @fence is consumed in both the success and error cases.
873 *
874 * Returns:
875 * 0 on success, or an error on failing to expand the array.
876 */
877int drm_sched_job_add_dependency(struct drm_sched_job *job,
878 struct dma_fence *fence)
879{
880 struct dma_fence *entry;
881 unsigned long index;
882 u32 id = 0;
883 int ret;
884
885 if (!fence)
886 return 0;
887
888 /* Deduplicate if we already depend on a fence from the same context.
889 * This lets the size of the array of deps scale with the number of
890 * engines involved, rather than the number of BOs.
891 */
892 xa_for_each(&job->dependencies, index, entry) {
893 if (entry->context != fence->context)
894 continue;
895
896 if (dma_fence_is_later(fence, entry)) {
897 dma_fence_put(entry);
898 xa_store(&job->dependencies, index, fence, GFP_KERNEL);
899 } else {
900 dma_fence_put(fence);
901 }
902 return 0;
903 }
904
905 ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
906 if (ret != 0)
907 dma_fence_put(fence);
908
909 return ret;
910}
911EXPORT_SYMBOL(drm_sched_job_add_dependency);
912
913/**
914 * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
915 * @job: scheduler job to add the dependencies to
916 * @file: drm file private pointer
917 * @handle: syncobj handle to lookup
918 * @point: timeline point
919 *
920 * This adds the fence matching the given syncobj to @job.
921 *
922 * Returns:
923 * 0 on success, or an error on failing to expand the array.
924 */
925int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
926 struct drm_file *file,
927 u32 handle,
928 u32 point)
929{
930 struct dma_fence *fence;
931 int ret;
932
933 ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
934 if (ret)
935 return ret;
936
937 return drm_sched_job_add_dependency(job, fence);
938}
939EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
940
941/**
942 * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
943 * @job: scheduler job to add the dependencies to
944 * @resv: the dma_resv object to get the fences from
945 * @usage: the dma_resv_usage to use to filter the fences
946 *
947 * This adds all fences matching the given usage from @resv to @job.
948 * Must be called with the @resv lock held.
949 *
950 * Returns:
951 * 0 on success, or an error on failing to expand the array.
952 */
953int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
954 struct dma_resv *resv,
955 enum dma_resv_usage usage)
956{
957 struct dma_resv_iter cursor;
958 struct dma_fence *fence;
959 int ret;
960
961 dma_resv_assert_held(resv);
962
963 dma_resv_for_each_fence(&cursor, resv, usage, fence) {
964 /* Make sure to grab an additional ref on the added fence */
965 dma_fence_get(fence);
966 ret = drm_sched_job_add_dependency(job, fence);
967 if (ret) {
968 dma_fence_put(fence);
969 return ret;
970 }
971 }
972 return 0;
973}
974EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
975
976/**
977 * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
978 * dependencies
979 * @job: scheduler job to add the dependencies to
980 * @obj: the gem object to add new dependencies from.
981 * @write: whether the job might write the object (so we need to depend on
982 * shared fences in the reservation object).
983 *
984 * This should be called after drm_gem_lock_reservations() on your array of
985 * GEM objects used in the job but before updating the reservations with your
986 * own fences.
987 *
988 * Returns:
989 * 0 on success, or an error on failing to expand the array.
990 */
991int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
992 struct drm_gem_object *obj,
993 bool write)
994{
995 return drm_sched_job_add_resv_dependencies(job, obj->resv,
996 dma_resv_usage_rw(write));
997}
998EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
999
1000/**
1001 * drm_sched_job_cleanup - clean up scheduler job resources
1002 * @job: scheduler job to clean up
1003 *
1004 * Cleans up the resources allocated with drm_sched_job_init().
1005 *
1006 * Drivers should call this from their error unwind code if @job is aborted
1007 * before drm_sched_job_arm() is called.
1008 *
1009 * After that point of no return @job is committed to be executed by the
1010 * scheduler, and this function should be called from the
1011 * &drm_sched_backend_ops.free_job callback.
1012 */
1013void drm_sched_job_cleanup(struct drm_sched_job *job)
1014{
1015 struct dma_fence *fence;
1016 unsigned long index;
1017
1018 if (kref_read(&job->s_fence->finished.refcount)) {
1019 /* drm_sched_job_arm() has been called */
1020 dma_fence_put(&job->s_fence->finished);
1021 } else {
1022 /* aborted job before committing to run it */
1023 drm_sched_fence_free(job->s_fence);
1024 }
1025
1026 job->s_fence = NULL;
1027
1028 xa_for_each(&job->dependencies, index, fence) {
1029 dma_fence_put(fence);
1030 }
1031 xa_destroy(&job->dependencies);
1032
1033}
1034EXPORT_SYMBOL(drm_sched_job_cleanup);
1035
1036/**
1037 * drm_sched_wakeup - Wake up the scheduler if it is ready to queue
1038 * @sched: scheduler instance
1039 *
1040 * Wake up the scheduler if we can queue jobs.
1041 */
1042void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
1043{
1044 drm_sched_run_job_queue(sched);
1045}
1046
1047/**
1048 * drm_sched_select_entity - Select next entity to process
1049 *
1050 * @sched: scheduler instance
1051 *
1052 * Return an entity to process or NULL if none are found.
1053 *
1054 * Note, that we break out of the for-loop when "entity" is non-null, which can
1055 * also be an error-pointer--this assures we don't process lower priority
1056 * run-queues. See comments in the respectively called functions.
1057 */
1058static struct drm_sched_entity *
1059drm_sched_select_entity(struct drm_gpu_scheduler *sched)
1060{
1061 struct drm_sched_entity *entity;
1062 int i;
1063
1064 /* Start with the highest priority.
1065 */
1066 for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1067 entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
1068 drm_sched_rq_select_entity_fifo(sched, sched->sched_rq[i]) :
1069 drm_sched_rq_select_entity_rr(sched, sched->sched_rq[i]);
1070 if (entity)
1071 break;
1072 }
1073
1074 return IS_ERR(entity) ? NULL : entity;
1075}
1076
1077/**
1078 * drm_sched_get_finished_job - fetch the next finished job to be destroyed
1079 *
1080 * @sched: scheduler instance
1081 *
1082 * Returns the next finished job from the pending list (if there is one)
1083 * ready for it to be destroyed.
1084 */
1085static struct drm_sched_job *
1086drm_sched_get_finished_job(struct drm_gpu_scheduler *sched)
1087{
1088 struct drm_sched_job *job, *next;
1089
1090 spin_lock(&sched->job_list_lock);
1091
1092 job = list_first_entry_or_null(&sched->pending_list,
1093 struct drm_sched_job, list);
1094
1095 if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
1096 /* remove job from pending_list */
1097 list_del_init(&job->list);
1098
1099 /* cancel this job's TO timer */
1100 cancel_delayed_work(&sched->work_tdr);
1101 /* make the scheduled timestamp more accurate */
1102 next = list_first_entry_or_null(&sched->pending_list,
1103 typeof(*next), list);
1104
1105 if (next) {
1106 if (test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT,
1107 &next->s_fence->scheduled.flags))
1108 next->s_fence->scheduled.timestamp =
1109 dma_fence_timestamp(&job->s_fence->finished);
1110 /* start TO timer for next job */
1111 drm_sched_start_timeout(sched);
1112 }
1113 } else {
1114 job = NULL;
1115 }
1116
1117 spin_unlock(&sched->job_list_lock);
1118
1119 return job;
1120}
1121
1122/**
1123 * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
1124 * @sched_list: list of drm_gpu_schedulers
1125 * @num_sched_list: number of drm_gpu_schedulers in the sched_list
1126 *
1127 * Returns pointer of the sched with the least load or NULL if none of the
1128 * drm_gpu_schedulers are ready
1129 */
1130struct drm_gpu_scheduler *
1131drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
1132 unsigned int num_sched_list)
1133{
1134 struct drm_gpu_scheduler *sched, *picked_sched = NULL;
1135 int i;
1136 unsigned int min_score = UINT_MAX, num_score;
1137
1138 for (i = 0; i < num_sched_list; ++i) {
1139 sched = sched_list[i];
1140
1141 if (!sched->ready) {
1142 DRM_WARN("scheduler %s is not ready, skipping",
1143 sched->name);
1144 continue;
1145 }
1146
1147 num_score = atomic_read(sched->score);
1148 if (num_score < min_score) {
1149 min_score = num_score;
1150 picked_sched = sched;
1151 }
1152 }
1153
1154 return picked_sched;
1155}
1156EXPORT_SYMBOL(drm_sched_pick_best);
1157
1158/**
1159 * drm_sched_free_job_work - worker to call free_job
1160 *
1161 * @w: free job work
1162 */
1163static void drm_sched_free_job_work(struct work_struct *w)
1164{
1165 struct drm_gpu_scheduler *sched =
1166 container_of(w, struct drm_gpu_scheduler, work_free_job);
1167 struct drm_sched_job *job;
1168
1169 if (READ_ONCE(sched->pause_submit))
1170 return;
1171
1172 job = drm_sched_get_finished_job(sched);
1173 if (job)
1174 sched->ops->free_job(job);
1175
1176 drm_sched_run_free_queue(sched);
1177 drm_sched_run_job_queue(sched);
1178}
1179
1180/**
1181 * drm_sched_run_job_work - worker to call run_job
1182 *
1183 * @w: run job work
1184 */
1185static void drm_sched_run_job_work(struct work_struct *w)
1186{
1187 struct drm_gpu_scheduler *sched =
1188 container_of(w, struct drm_gpu_scheduler, work_run_job);
1189 struct drm_sched_entity *entity;
1190 struct dma_fence *fence;
1191 struct drm_sched_fence *s_fence;
1192 struct drm_sched_job *sched_job;
1193 int r;
1194
1195 if (READ_ONCE(sched->pause_submit))
1196 return;
1197
1198 /* Find entity with a ready job */
1199 entity = drm_sched_select_entity(sched);
1200 if (!entity)
1201 return; /* No more work */
1202
1203 sched_job = drm_sched_entity_pop_job(entity);
1204 if (!sched_job) {
1205 complete_all(&entity->entity_idle);
1206 drm_sched_run_job_queue(sched);
1207 return;
1208 }
1209
1210 s_fence = sched_job->s_fence;
1211
1212 atomic_add(sched_job->credits, &sched->credit_count);
1213 drm_sched_job_begin(sched_job);
1214
1215 trace_drm_run_job(sched_job, entity);
1216 fence = sched->ops->run_job(sched_job);
1217 complete_all(&entity->entity_idle);
1218 drm_sched_fence_scheduled(s_fence, fence);
1219
1220 if (!IS_ERR_OR_NULL(fence)) {
1221 /* Drop for original kref_init of the fence */
1222 dma_fence_put(fence);
1223
1224 r = dma_fence_add_callback(fence, &sched_job->cb,
1225 drm_sched_job_done_cb);
1226 if (r == -ENOENT)
1227 drm_sched_job_done(sched_job, fence->error);
1228 else if (r)
1229 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
1230 } else {
1231 drm_sched_job_done(sched_job, IS_ERR(fence) ?
1232 PTR_ERR(fence) : 0);
1233 }
1234
1235 wake_up(&sched->job_scheduled);
1236 drm_sched_run_job_queue(sched);
1237}
1238
1239/**
1240 * drm_sched_init - Init a gpu scheduler instance
1241 *
1242 * @sched: scheduler instance
1243 * @ops: backend operations for this scheduler
1244 * @submit_wq: workqueue to use for submission. If NULL, an ordered wq is
1245 * allocated and used
1246 * @num_rqs: number of runqueues, one for each priority, up to DRM_SCHED_PRIORITY_COUNT
1247 * @credit_limit: the number of credits this scheduler can hold from all jobs
1248 * @hang_limit: number of times to allow a job to hang before dropping it
1249 * @timeout: timeout value in jiffies for the scheduler
1250 * @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
1251 * used
1252 * @score: optional score atomic shared with other schedulers
1253 * @name: name used for debugging
1254 * @dev: target &struct device
1255 *
1256 * Return 0 on success, otherwise error code.
1257 */
1258int drm_sched_init(struct drm_gpu_scheduler *sched,
1259 const struct drm_sched_backend_ops *ops,
1260 struct workqueue_struct *submit_wq,
1261 u32 num_rqs, u32 credit_limit, unsigned int hang_limit,
1262 long timeout, struct workqueue_struct *timeout_wq,
1263 atomic_t *score, const char *name, struct device *dev)
1264{
1265 int i;
1266
1267 sched->ops = ops;
1268 sched->credit_limit = credit_limit;
1269 sched->name = name;
1270 sched->timeout = timeout;
1271 sched->timeout_wq = timeout_wq ? : system_wq;
1272 sched->hang_limit = hang_limit;
1273 sched->score = score ? score : &sched->_score;
1274 sched->dev = dev;
1275
1276 if (num_rqs > DRM_SCHED_PRIORITY_COUNT) {
1277 /* This is a gross violation--tell drivers what the problem is.
1278 */
1279 drm_err(sched, "%s: num_rqs cannot be greater than DRM_SCHED_PRIORITY_COUNT\n",
1280 __func__);
1281 return -EINVAL;
1282 } else if (sched->sched_rq) {
1283 /* Not an error, but warn anyway so drivers can
1284 * fine-tune their DRM calling order, and return all
1285 * is good.
1286 */
1287 drm_warn(sched, "%s: scheduler already initialized!\n", __func__);
1288 return 0;
1289 }
1290
1291 if (submit_wq) {
1292 sched->submit_wq = submit_wq;
1293 sched->own_submit_wq = false;
1294 } else {
1295#ifdef CONFIG_LOCKDEP
1296 sched->submit_wq = alloc_ordered_workqueue_lockdep_map(name,
1297 WQ_MEM_RECLAIM,
1298 &drm_sched_lockdep_map);
1299#else
1300 sched->submit_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
1301#endif
1302 if (!sched->submit_wq)
1303 return -ENOMEM;
1304
1305 sched->own_submit_wq = true;
1306 }
1307
1308 sched->sched_rq = kmalloc_array(num_rqs, sizeof(*sched->sched_rq),
1309 GFP_KERNEL | __GFP_ZERO);
1310 if (!sched->sched_rq)
1311 goto Out_check_own;
1312 sched->num_rqs = num_rqs;
1313 for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1314 sched->sched_rq[i] = kzalloc(sizeof(*sched->sched_rq[i]), GFP_KERNEL);
1315 if (!sched->sched_rq[i])
1316 goto Out_unroll;
1317 drm_sched_rq_init(sched, sched->sched_rq[i]);
1318 }
1319
1320 init_waitqueue_head(&sched->job_scheduled);
1321 INIT_LIST_HEAD(&sched->pending_list);
1322 spin_lock_init(&sched->job_list_lock);
1323 atomic_set(&sched->credit_count, 0);
1324 INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1325 INIT_WORK(&sched->work_run_job, drm_sched_run_job_work);
1326 INIT_WORK(&sched->work_free_job, drm_sched_free_job_work);
1327 atomic_set(&sched->_score, 0);
1328 atomic64_set(&sched->job_id_count, 0);
1329 sched->pause_submit = false;
1330
1331 sched->ready = true;
1332 return 0;
1333Out_unroll:
1334 for (--i ; i >= DRM_SCHED_PRIORITY_KERNEL; i--)
1335 kfree(sched->sched_rq[i]);
1336
1337 kfree(sched->sched_rq);
1338 sched->sched_rq = NULL;
1339Out_check_own:
1340 if (sched->own_submit_wq)
1341 destroy_workqueue(sched->submit_wq);
1342 drm_err(sched, "%s: Failed to setup GPU scheduler--out of memory\n", __func__);
1343 return -ENOMEM;
1344}
1345EXPORT_SYMBOL(drm_sched_init);
1346
1347/**
1348 * drm_sched_fini - Destroy a gpu scheduler
1349 *
1350 * @sched: scheduler instance
1351 *
1352 * Tears down and cleans up the scheduler.
1353 *
1354 * This stops submission of new jobs to the hardware through
1355 * drm_sched_backend_ops.run_job(). Consequently, drm_sched_backend_ops.free_job()
1356 * will not be called for all jobs still in drm_gpu_scheduler.pending_list.
1357 * There is no solution for this currently. Thus, it is up to the driver to make
1358 * sure that:
1359 *
1360 * a) drm_sched_fini() is only called after for all submitted jobs
1361 * drm_sched_backend_ops.free_job() has been called or that
1362 * b) the jobs for which drm_sched_backend_ops.free_job() has not been called
1363 * after drm_sched_fini() ran are freed manually.
1364 *
1365 * FIXME: Take care of the above problem and prevent this function from leaking
1366 * the jobs in drm_gpu_scheduler.pending_list under any circumstances.
1367 */
1368void drm_sched_fini(struct drm_gpu_scheduler *sched)
1369{
1370 struct drm_sched_entity *s_entity;
1371 int i;
1372
1373 drm_sched_wqueue_stop(sched);
1374
1375 for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1376 struct drm_sched_rq *rq = sched->sched_rq[i];
1377
1378 spin_lock(&rq->lock);
1379 list_for_each_entry(s_entity, &rq->entities, list)
1380 /*
1381 * Prevents reinsertion and marks job_queue as idle,
1382 * it will be removed from the rq in drm_sched_entity_fini()
1383 * eventually
1384 */
1385 s_entity->stopped = true;
1386 spin_unlock(&rq->lock);
1387 kfree(sched->sched_rq[i]);
1388 }
1389
1390 /* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1391 wake_up_all(&sched->job_scheduled);
1392
1393 /* Confirm no work left behind accessing device structures */
1394 cancel_delayed_work_sync(&sched->work_tdr);
1395
1396 if (sched->own_submit_wq)
1397 destroy_workqueue(sched->submit_wq);
1398 sched->ready = false;
1399 kfree(sched->sched_rq);
1400 sched->sched_rq = NULL;
1401}
1402EXPORT_SYMBOL(drm_sched_fini);
1403
1404/**
1405 * drm_sched_increase_karma - Update sched_entity guilty flag
1406 *
1407 * @bad: The job guilty of time out
1408 *
1409 * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1410 * limit of the scheduler then the respective sched entity is marked guilty and
1411 * jobs from it will not be scheduled further
1412 */
1413void drm_sched_increase_karma(struct drm_sched_job *bad)
1414{
1415 int i;
1416 struct drm_sched_entity *tmp;
1417 struct drm_sched_entity *entity;
1418 struct drm_gpu_scheduler *sched = bad->sched;
1419
1420 /* don't change @bad's karma if it's from KERNEL RQ,
1421 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1422 * corrupt but keep in mind that kernel jobs always considered good.
1423 */
1424 if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1425 atomic_inc(&bad->karma);
1426
1427 for (i = DRM_SCHED_PRIORITY_HIGH; i < sched->num_rqs; i++) {
1428 struct drm_sched_rq *rq = sched->sched_rq[i];
1429
1430 spin_lock(&rq->lock);
1431 list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1432 if (bad->s_fence->scheduled.context ==
1433 entity->fence_context) {
1434 if (entity->guilty)
1435 atomic_set(entity->guilty, 1);
1436 break;
1437 }
1438 }
1439 spin_unlock(&rq->lock);
1440 if (&entity->list != &rq->entities)
1441 break;
1442 }
1443 }
1444}
1445EXPORT_SYMBOL(drm_sched_increase_karma);
1446
1447/**
1448 * drm_sched_wqueue_ready - Is the scheduler ready for submission
1449 *
1450 * @sched: scheduler instance
1451 *
1452 * Returns true if submission is ready
1453 */
1454bool drm_sched_wqueue_ready(struct drm_gpu_scheduler *sched)
1455{
1456 return sched->ready;
1457}
1458EXPORT_SYMBOL(drm_sched_wqueue_ready);
1459
1460/**
1461 * drm_sched_wqueue_stop - stop scheduler submission
1462 * @sched: scheduler instance
1463 *
1464 * Stops the scheduler from pulling new jobs from entities. It also stops
1465 * freeing jobs automatically through drm_sched_backend_ops.free_job().
1466 */
1467void drm_sched_wqueue_stop(struct drm_gpu_scheduler *sched)
1468{
1469 WRITE_ONCE(sched->pause_submit, true);
1470 cancel_work_sync(&sched->work_run_job);
1471 cancel_work_sync(&sched->work_free_job);
1472}
1473EXPORT_SYMBOL(drm_sched_wqueue_stop);
1474
1475/**
1476 * drm_sched_wqueue_start - start scheduler submission
1477 * @sched: scheduler instance
1478 *
1479 * Restarts the scheduler after drm_sched_wqueue_stop() has stopped it.
1480 *
1481 * This function is not necessary for 'conventional' startup. The scheduler is
1482 * fully operational after drm_sched_init() succeeded.
1483 */
1484void drm_sched_wqueue_start(struct drm_gpu_scheduler *sched)
1485{
1486 WRITE_ONCE(sched->pause_submit, false);
1487 queue_work(sched->submit_wq, &sched->work_run_job);
1488 queue_work(sched->submit_wq, &sched->work_free_job);
1489}
1490EXPORT_SYMBOL(drm_sched_wqueue_start);
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
47#include <linux/kthread.h>
48#include <linux/wait.h>
49#include <linux/sched.h>
50#include <linux/completion.h>
51#include <uapi/linux/sched/types.h>
52
53#include <drm/drm_print.h>
54#include <drm/gpu_scheduler.h>
55#include <drm/spsc_queue.h>
56
57#define CREATE_TRACE_POINTS
58#include "gpu_scheduler_trace.h"
59
60#define to_drm_sched_job(sched_job) \
61 container_of((sched_job), struct drm_sched_job, queue_node)
62
63static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
64
65/**
66 * drm_sched_rq_init - initialize a given run queue struct
67 *
68 * @rq: scheduler run queue
69 *
70 * Initializes a scheduler runqueue.
71 */
72static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
73 struct drm_sched_rq *rq)
74{
75 spin_lock_init(&rq->lock);
76 INIT_LIST_HEAD(&rq->entities);
77 rq->current_entity = NULL;
78 rq->sched = sched;
79}
80
81/**
82 * drm_sched_rq_add_entity - add an entity
83 *
84 * @rq: scheduler run queue
85 * @entity: scheduler entity
86 *
87 * Adds a scheduler entity to the run queue.
88 */
89void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
90 struct drm_sched_entity *entity)
91{
92 if (!list_empty(&entity->list))
93 return;
94 spin_lock(&rq->lock);
95 atomic_inc(&rq->sched->score);
96 list_add_tail(&entity->list, &rq->entities);
97 spin_unlock(&rq->lock);
98}
99
100/**
101 * drm_sched_rq_remove_entity - remove an entity
102 *
103 * @rq: scheduler run queue
104 * @entity: scheduler entity
105 *
106 * Removes a scheduler entity from the run queue.
107 */
108void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
109 struct drm_sched_entity *entity)
110{
111 if (list_empty(&entity->list))
112 return;
113 spin_lock(&rq->lock);
114 atomic_dec(&rq->sched->score);
115 list_del_init(&entity->list);
116 if (rq->current_entity == entity)
117 rq->current_entity = NULL;
118 spin_unlock(&rq->lock);
119}
120
121/**
122 * drm_sched_rq_select_entity - Select an entity which could provide a job to run
123 *
124 * @rq: scheduler run queue to check.
125 *
126 * Try to find a ready entity, returns NULL if none found.
127 */
128static struct drm_sched_entity *
129drm_sched_rq_select_entity(struct drm_sched_rq *rq)
130{
131 struct drm_sched_entity *entity;
132
133 spin_lock(&rq->lock);
134
135 entity = rq->current_entity;
136 if (entity) {
137 list_for_each_entry_continue(entity, &rq->entities, list) {
138 if (drm_sched_entity_is_ready(entity)) {
139 rq->current_entity = entity;
140 reinit_completion(&entity->entity_idle);
141 spin_unlock(&rq->lock);
142 return entity;
143 }
144 }
145 }
146
147 list_for_each_entry(entity, &rq->entities, list) {
148
149 if (drm_sched_entity_is_ready(entity)) {
150 rq->current_entity = entity;
151 reinit_completion(&entity->entity_idle);
152 spin_unlock(&rq->lock);
153 return entity;
154 }
155
156 if (entity == rq->current_entity)
157 break;
158 }
159
160 spin_unlock(&rq->lock);
161
162 return NULL;
163}
164
165/**
166 * drm_sched_dependency_optimized
167 *
168 * @fence: the dependency fence
169 * @entity: the entity which depends on the above fence
170 *
171 * Returns true if the dependency can be optimized and false otherwise
172 */
173bool drm_sched_dependency_optimized(struct dma_fence* fence,
174 struct drm_sched_entity *entity)
175{
176 struct drm_gpu_scheduler *sched = entity->rq->sched;
177 struct drm_sched_fence *s_fence;
178
179 if (!fence || dma_fence_is_signaled(fence))
180 return false;
181 if (fence->context == entity->fence_context)
182 return true;
183 s_fence = to_drm_sched_fence(fence);
184 if (s_fence && s_fence->sched == sched)
185 return true;
186
187 return false;
188}
189EXPORT_SYMBOL(drm_sched_dependency_optimized);
190
191/**
192 * drm_sched_start_timeout - start timeout for reset worker
193 *
194 * @sched: scheduler instance to start the worker for
195 *
196 * Start the timeout for the given scheduler.
197 */
198static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
199{
200 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
201 !list_empty(&sched->ring_mirror_list))
202 schedule_delayed_work(&sched->work_tdr, sched->timeout);
203}
204
205/**
206 * drm_sched_fault - immediately start timeout handler
207 *
208 * @sched: scheduler where the timeout handling should be started.
209 *
210 * Start timeout handling immediately when the driver detects a hardware fault.
211 */
212void drm_sched_fault(struct drm_gpu_scheduler *sched)
213{
214 mod_delayed_work(system_wq, &sched->work_tdr, 0);
215}
216EXPORT_SYMBOL(drm_sched_fault);
217
218/**
219 * drm_sched_suspend_timeout - Suspend scheduler job timeout
220 *
221 * @sched: scheduler instance for which to suspend the timeout
222 *
223 * Suspend the delayed work timeout for the scheduler. This is done by
224 * modifying the delayed work timeout to an arbitrary large value,
225 * MAX_SCHEDULE_TIMEOUT in this case.
226 *
227 * Returns the timeout remaining
228 *
229 */
230unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
231{
232 unsigned long sched_timeout, now = jiffies;
233
234 sched_timeout = sched->work_tdr.timer.expires;
235
236 /*
237 * Modify the timeout to an arbitrarily large value. This also prevents
238 * the timeout to be restarted when new submissions arrive
239 */
240 if (mod_delayed_work(system_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
241 && time_after(sched_timeout, now))
242 return sched_timeout - now;
243 else
244 return sched->timeout;
245}
246EXPORT_SYMBOL(drm_sched_suspend_timeout);
247
248/**
249 * drm_sched_resume_timeout - Resume scheduler job timeout
250 *
251 * @sched: scheduler instance for which to resume the timeout
252 * @remaining: remaining timeout
253 *
254 * Resume the delayed work timeout for the scheduler.
255 */
256void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
257 unsigned long remaining)
258{
259 spin_lock(&sched->job_list_lock);
260
261 if (list_empty(&sched->ring_mirror_list))
262 cancel_delayed_work(&sched->work_tdr);
263 else
264 mod_delayed_work(system_wq, &sched->work_tdr, remaining);
265
266 spin_unlock(&sched->job_list_lock);
267}
268EXPORT_SYMBOL(drm_sched_resume_timeout);
269
270static void drm_sched_job_begin(struct drm_sched_job *s_job)
271{
272 struct drm_gpu_scheduler *sched = s_job->sched;
273
274 spin_lock(&sched->job_list_lock);
275 list_add_tail(&s_job->node, &sched->ring_mirror_list);
276 drm_sched_start_timeout(sched);
277 spin_unlock(&sched->job_list_lock);
278}
279
280static void drm_sched_job_timedout(struct work_struct *work)
281{
282 struct drm_gpu_scheduler *sched;
283 struct drm_sched_job *job;
284
285 sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
286
287 /* Protects against concurrent deletion in drm_sched_get_cleanup_job */
288 spin_lock(&sched->job_list_lock);
289 job = list_first_entry_or_null(&sched->ring_mirror_list,
290 struct drm_sched_job, node);
291
292 if (job) {
293 /*
294 * Remove the bad job so it cannot be freed by concurrent
295 * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
296 * is parked at which point it's safe.
297 */
298 list_del_init(&job->node);
299 spin_unlock(&sched->job_list_lock);
300
301 job->sched->ops->timedout_job(job);
302
303 /*
304 * Guilty job did complete and hence needs to be manually removed
305 * See drm_sched_stop doc.
306 */
307 if (sched->free_guilty) {
308 job->sched->ops->free_job(job);
309 sched->free_guilty = false;
310 }
311 } else {
312 spin_unlock(&sched->job_list_lock);
313 }
314
315 spin_lock(&sched->job_list_lock);
316 drm_sched_start_timeout(sched);
317 spin_unlock(&sched->job_list_lock);
318}
319
320 /**
321 * drm_sched_increase_karma - Update sched_entity guilty flag
322 *
323 * @bad: The job guilty of time out
324 *
325 * Increment on every hang caused by the 'bad' job. If this exceeds the hang
326 * limit of the scheduler then the respective sched entity is marked guilty and
327 * jobs from it will not be scheduled further
328 */
329void drm_sched_increase_karma(struct drm_sched_job *bad)
330{
331 int i;
332 struct drm_sched_entity *tmp;
333 struct drm_sched_entity *entity;
334 struct drm_gpu_scheduler *sched = bad->sched;
335
336 /* don't increase @bad's karma if it's from KERNEL RQ,
337 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
338 * corrupt but keep in mind that kernel jobs always considered good.
339 */
340 if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
341 atomic_inc(&bad->karma);
342 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
343 i++) {
344 struct drm_sched_rq *rq = &sched->sched_rq[i];
345
346 spin_lock(&rq->lock);
347 list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
348 if (bad->s_fence->scheduled.context ==
349 entity->fence_context) {
350 if (atomic_read(&bad->karma) >
351 bad->sched->hang_limit)
352 if (entity->guilty)
353 atomic_set(entity->guilty, 1);
354 break;
355 }
356 }
357 spin_unlock(&rq->lock);
358 if (&entity->list != &rq->entities)
359 break;
360 }
361 }
362}
363EXPORT_SYMBOL(drm_sched_increase_karma);
364
365/**
366 * drm_sched_stop - stop the scheduler
367 *
368 * @sched: scheduler instance
369 * @bad: job which caused the time out
370 *
371 * Stop the scheduler and also removes and frees all completed jobs.
372 * Note: bad job will not be freed as it might be used later and so it's
373 * callers responsibility to release it manually if it's not part of the
374 * mirror list any more.
375 *
376 */
377void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
378{
379 struct drm_sched_job *s_job, *tmp;
380
381 kthread_park(sched->thread);
382
383 /*
384 * Reinsert back the bad job here - now it's safe as
385 * drm_sched_get_cleanup_job cannot race against us and release the
386 * bad job at this point - we parked (waited for) any in progress
387 * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
388 * now until the scheduler thread is unparked.
389 */
390 if (bad && bad->sched == sched)
391 /*
392 * Add at the head of the queue to reflect it was the earliest
393 * job extracted.
394 */
395 list_add(&bad->node, &sched->ring_mirror_list);
396
397 /*
398 * Iterate the job list from later to earlier one and either deactive
399 * their HW callbacks or remove them from mirror list if they already
400 * signaled.
401 * This iteration is thread safe as sched thread is stopped.
402 */
403 list_for_each_entry_safe_reverse(s_job, tmp, &sched->ring_mirror_list, node) {
404 if (s_job->s_fence->parent &&
405 dma_fence_remove_callback(s_job->s_fence->parent,
406 &s_job->cb)) {
407 atomic_dec(&sched->hw_rq_count);
408 } else {
409 /*
410 * remove job from ring_mirror_list.
411 * Locking here is for concurrent resume timeout
412 */
413 spin_lock(&sched->job_list_lock);
414 list_del_init(&s_job->node);
415 spin_unlock(&sched->job_list_lock);
416
417 /*
418 * Wait for job's HW fence callback to finish using s_job
419 * before releasing it.
420 *
421 * Job is still alive so fence refcount at least 1
422 */
423 dma_fence_wait(&s_job->s_fence->finished, false);
424
425 /*
426 * We must keep bad job alive for later use during
427 * recovery by some of the drivers but leave a hint
428 * that the guilty job must be released.
429 */
430 if (bad != s_job)
431 sched->ops->free_job(s_job);
432 else
433 sched->free_guilty = true;
434 }
435 }
436
437 /*
438 * Stop pending timer in flight as we rearm it in drm_sched_start. This
439 * avoids the pending timeout work in progress to fire right away after
440 * this TDR finished and before the newly restarted jobs had a
441 * chance to complete.
442 */
443 cancel_delayed_work(&sched->work_tdr);
444}
445
446EXPORT_SYMBOL(drm_sched_stop);
447
448/**
449 * drm_sched_job_recovery - recover jobs after a reset
450 *
451 * @sched: scheduler instance
452 * @full_recovery: proceed with complete sched restart
453 *
454 */
455void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
456{
457 struct drm_sched_job *s_job, *tmp;
458 int r;
459
460 /*
461 * Locking the list is not required here as the sched thread is parked
462 * so no new jobs are being inserted or removed. Also concurrent
463 * GPU recovers can't run in parallel.
464 */
465 list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
466 struct dma_fence *fence = s_job->s_fence->parent;
467
468 atomic_inc(&sched->hw_rq_count);
469
470 if (!full_recovery)
471 continue;
472
473 if (fence) {
474 r = dma_fence_add_callback(fence, &s_job->cb,
475 drm_sched_process_job);
476 if (r == -ENOENT)
477 drm_sched_process_job(fence, &s_job->cb);
478 else if (r)
479 DRM_ERROR("fence add callback failed (%d)\n",
480 r);
481 } else
482 drm_sched_process_job(NULL, &s_job->cb);
483 }
484
485 if (full_recovery) {
486 spin_lock(&sched->job_list_lock);
487 drm_sched_start_timeout(sched);
488 spin_unlock(&sched->job_list_lock);
489 }
490
491 kthread_unpark(sched->thread);
492}
493EXPORT_SYMBOL(drm_sched_start);
494
495/**
496 * drm_sched_resubmit_jobs - helper to relunch job from mirror ring list
497 *
498 * @sched: scheduler instance
499 *
500 */
501void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
502{
503 struct drm_sched_job *s_job, *tmp;
504 uint64_t guilty_context;
505 bool found_guilty = false;
506 struct dma_fence *fence;
507
508 list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
509 struct drm_sched_fence *s_fence = s_job->s_fence;
510
511 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
512 found_guilty = true;
513 guilty_context = s_job->s_fence->scheduled.context;
514 }
515
516 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
517 dma_fence_set_error(&s_fence->finished, -ECANCELED);
518
519 dma_fence_put(s_job->s_fence->parent);
520 fence = sched->ops->run_job(s_job);
521
522 if (IS_ERR_OR_NULL(fence)) {
523 if (IS_ERR(fence))
524 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
525
526 s_job->s_fence->parent = NULL;
527 } else {
528 s_job->s_fence->parent = fence;
529 }
530
531
532 }
533}
534EXPORT_SYMBOL(drm_sched_resubmit_jobs);
535
536/**
537 * drm_sched_job_init - init a scheduler job
538 *
539 * @job: scheduler job to init
540 * @entity: scheduler entity to use
541 * @owner: job owner for debugging
542 *
543 * Refer to drm_sched_entity_push_job() documentation
544 * for locking considerations.
545 *
546 * Returns 0 for success, negative error code otherwise.
547 */
548int drm_sched_job_init(struct drm_sched_job *job,
549 struct drm_sched_entity *entity,
550 void *owner)
551{
552 struct drm_gpu_scheduler *sched;
553
554 drm_sched_entity_select_rq(entity);
555 if (!entity->rq)
556 return -ENOENT;
557
558 sched = entity->rq->sched;
559
560 job->sched = sched;
561 job->entity = entity;
562 job->s_priority = entity->rq - sched->sched_rq;
563 job->s_fence = drm_sched_fence_create(entity, owner);
564 if (!job->s_fence)
565 return -ENOMEM;
566 job->id = atomic64_inc_return(&sched->job_id_count);
567
568 INIT_LIST_HEAD(&job->node);
569
570 return 0;
571}
572EXPORT_SYMBOL(drm_sched_job_init);
573
574/**
575 * drm_sched_job_cleanup - clean up scheduler job resources
576 *
577 * @job: scheduler job to clean up
578 */
579void drm_sched_job_cleanup(struct drm_sched_job *job)
580{
581 dma_fence_put(&job->s_fence->finished);
582 job->s_fence = NULL;
583}
584EXPORT_SYMBOL(drm_sched_job_cleanup);
585
586/**
587 * drm_sched_ready - is the scheduler ready
588 *
589 * @sched: scheduler instance
590 *
591 * Return true if we can push more jobs to the hw, otherwise false.
592 */
593static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
594{
595 return atomic_read(&sched->hw_rq_count) <
596 sched->hw_submission_limit;
597}
598
599/**
600 * drm_sched_wakeup - Wake up the scheduler when it is ready
601 *
602 * @sched: scheduler instance
603 *
604 */
605void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
606{
607 if (drm_sched_ready(sched))
608 wake_up_interruptible(&sched->wake_up_worker);
609}
610
611/**
612 * drm_sched_select_entity - Select next entity to process
613 *
614 * @sched: scheduler instance
615 *
616 * Returns the entity to process or NULL if none are found.
617 */
618static struct drm_sched_entity *
619drm_sched_select_entity(struct drm_gpu_scheduler *sched)
620{
621 struct drm_sched_entity *entity;
622 int i;
623
624 if (!drm_sched_ready(sched))
625 return NULL;
626
627 /* Kernel run queue has higher priority than normal run queue*/
628 for (i = DRM_SCHED_PRIORITY_MAX - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
629 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
630 if (entity)
631 break;
632 }
633
634 return entity;
635}
636
637/**
638 * drm_sched_process_job - process a job
639 *
640 * @f: fence
641 * @cb: fence callbacks
642 *
643 * Called after job has finished execution.
644 */
645static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
646{
647 struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
648 struct drm_sched_fence *s_fence = s_job->s_fence;
649 struct drm_gpu_scheduler *sched = s_fence->sched;
650
651 atomic_dec(&sched->hw_rq_count);
652 atomic_dec(&sched->score);
653
654 trace_drm_sched_process_job(s_fence);
655
656 dma_fence_get(&s_fence->finished);
657 drm_sched_fence_finished(s_fence);
658 dma_fence_put(&s_fence->finished);
659 wake_up_interruptible(&sched->wake_up_worker);
660}
661
662/**
663 * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
664 *
665 * @sched: scheduler instance
666 *
667 * Returns the next finished job from the mirror list (if there is one)
668 * ready for it to be destroyed.
669 */
670static struct drm_sched_job *
671drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
672{
673 struct drm_sched_job *job;
674
675 /*
676 * Don't destroy jobs while the timeout worker is running OR thread
677 * is being parked and hence assumed to not touch ring_mirror_list
678 */
679 if ((sched->timeout != MAX_SCHEDULE_TIMEOUT &&
680 !cancel_delayed_work(&sched->work_tdr)) ||
681 kthread_should_park())
682 return NULL;
683
684 spin_lock(&sched->job_list_lock);
685
686 job = list_first_entry_or_null(&sched->ring_mirror_list,
687 struct drm_sched_job, node);
688
689 if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
690 /* remove job from ring_mirror_list */
691 list_del_init(&job->node);
692 } else {
693 job = NULL;
694 /* queue timeout for next job */
695 drm_sched_start_timeout(sched);
696 }
697
698 spin_unlock(&sched->job_list_lock);
699
700 return job;
701}
702
703/**
704 * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
705 * @sched_list: list of drm_gpu_schedulers
706 * @num_sched_list: number of drm_gpu_schedulers in the sched_list
707 *
708 * Returns pointer of the sched with the least load or NULL if none of the
709 * drm_gpu_schedulers are ready
710 */
711struct drm_gpu_scheduler *
712drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
713 unsigned int num_sched_list)
714{
715 struct drm_gpu_scheduler *sched, *picked_sched = NULL;
716 int i;
717 unsigned int min_score = UINT_MAX, num_score;
718
719 for (i = 0; i < num_sched_list; ++i) {
720 sched = sched_list[i];
721
722 if (!sched->ready) {
723 DRM_WARN("scheduler %s is not ready, skipping",
724 sched->name);
725 continue;
726 }
727
728 num_score = atomic_read(&sched->score);
729 if (num_score < min_score) {
730 min_score = num_score;
731 picked_sched = sched;
732 }
733 }
734
735 return picked_sched;
736}
737EXPORT_SYMBOL(drm_sched_pick_best);
738
739/**
740 * drm_sched_blocked - check if the scheduler is blocked
741 *
742 * @sched: scheduler instance
743 *
744 * Returns true if blocked, otherwise false.
745 */
746static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
747{
748 if (kthread_should_park()) {
749 kthread_parkme();
750 return true;
751 }
752
753 return false;
754}
755
756/**
757 * drm_sched_main - main scheduler thread
758 *
759 * @param: scheduler instance
760 *
761 * Returns 0.
762 */
763static int drm_sched_main(void *param)
764{
765 struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
766 int r;
767
768 sched_set_fifo_low(current);
769
770 while (!kthread_should_stop()) {
771 struct drm_sched_entity *entity = NULL;
772 struct drm_sched_fence *s_fence;
773 struct drm_sched_job *sched_job;
774 struct dma_fence *fence;
775 struct drm_sched_job *cleanup_job = NULL;
776
777 wait_event_interruptible(sched->wake_up_worker,
778 (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
779 (!drm_sched_blocked(sched) &&
780 (entity = drm_sched_select_entity(sched))) ||
781 kthread_should_stop());
782
783 if (cleanup_job) {
784 sched->ops->free_job(cleanup_job);
785 /* queue timeout for next job */
786 drm_sched_start_timeout(sched);
787 }
788
789 if (!entity)
790 continue;
791
792 sched_job = drm_sched_entity_pop_job(entity);
793
794 complete(&entity->entity_idle);
795
796 if (!sched_job)
797 continue;
798
799 s_fence = sched_job->s_fence;
800
801 atomic_inc(&sched->hw_rq_count);
802 drm_sched_job_begin(sched_job);
803
804 trace_drm_run_job(sched_job, entity);
805 fence = sched->ops->run_job(sched_job);
806 drm_sched_fence_scheduled(s_fence);
807
808 if (!IS_ERR_OR_NULL(fence)) {
809 s_fence->parent = dma_fence_get(fence);
810 r = dma_fence_add_callback(fence, &sched_job->cb,
811 drm_sched_process_job);
812 if (r == -ENOENT)
813 drm_sched_process_job(fence, &sched_job->cb);
814 else if (r)
815 DRM_ERROR("fence add callback failed (%d)\n",
816 r);
817 dma_fence_put(fence);
818 } else {
819 if (IS_ERR(fence))
820 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
821
822 drm_sched_process_job(NULL, &sched_job->cb);
823 }
824
825 wake_up(&sched->job_scheduled);
826 }
827 return 0;
828}
829
830/**
831 * drm_sched_init - Init a gpu scheduler instance
832 *
833 * @sched: scheduler instance
834 * @ops: backend operations for this scheduler
835 * @hw_submission: number of hw submissions that can be in flight
836 * @hang_limit: number of times to allow a job to hang before dropping it
837 * @timeout: timeout value in jiffies for the scheduler
838 * @name: name used for debugging
839 *
840 * Return 0 on success, otherwise error code.
841 */
842int drm_sched_init(struct drm_gpu_scheduler *sched,
843 const struct drm_sched_backend_ops *ops,
844 unsigned hw_submission,
845 unsigned hang_limit,
846 long timeout,
847 const char *name)
848{
849 int i, ret;
850 sched->ops = ops;
851 sched->hw_submission_limit = hw_submission;
852 sched->name = name;
853 sched->timeout = timeout;
854 sched->hang_limit = hang_limit;
855 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++)
856 drm_sched_rq_init(sched, &sched->sched_rq[i]);
857
858 init_waitqueue_head(&sched->wake_up_worker);
859 init_waitqueue_head(&sched->job_scheduled);
860 INIT_LIST_HEAD(&sched->ring_mirror_list);
861 spin_lock_init(&sched->job_list_lock);
862 atomic_set(&sched->hw_rq_count, 0);
863 INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
864 atomic_set(&sched->score, 0);
865 atomic64_set(&sched->job_id_count, 0);
866
867 /* Each scheduler will run on a seperate kernel thread */
868 sched->thread = kthread_run(drm_sched_main, sched, sched->name);
869 if (IS_ERR(sched->thread)) {
870 ret = PTR_ERR(sched->thread);
871 sched->thread = NULL;
872 DRM_ERROR("Failed to create scheduler for %s.\n", name);
873 return ret;
874 }
875
876 sched->ready = true;
877 return 0;
878}
879EXPORT_SYMBOL(drm_sched_init);
880
881/**
882 * drm_sched_fini - Destroy a gpu scheduler
883 *
884 * @sched: scheduler instance
885 *
886 * Tears down and cleans up the scheduler.
887 */
888void drm_sched_fini(struct drm_gpu_scheduler *sched)
889{
890 if (sched->thread)
891 kthread_stop(sched->thread);
892
893 sched->ready = false;
894}
895EXPORT_SYMBOL(drm_sched_fini);