2 * Copyright 2015 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
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.
35 * The organisation of the scheduler is the following:
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
44 * The jobs in a entity are always scheduled in the order that they were pushed.
47 #include <linux/kthread.h>
48 #include <linux/wait.h>
49 #include <linux/sched.h>
50 #include <uapi/linux/sched/types.h>
52 #include <drm/gpu_scheduler.h>
53 #include <drm/spsc_queue.h>
55 #define CREATE_TRACE_POINTS
56 #include "gpu_scheduler_trace.h"
58 #define to_drm_sched_job(sched_job) \
59 container_of((sched_job), struct drm_sched_job, queue_node)
61 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
64 * drm_sched_rq_init - initialize a given run queue struct
66 * @rq: scheduler run queue
68 * Initializes a scheduler runqueue.
70 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
71 struct drm_sched_rq *rq)
73 spin_lock_init(&rq->lock);
74 INIT_LIST_HEAD(&rq->entities);
75 rq->current_entity = NULL;
80 * drm_sched_rq_add_entity - add an entity
82 * @rq: scheduler run queue
83 * @entity: scheduler entity
85 * Adds a scheduler entity to the run queue.
87 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
88 struct drm_sched_entity *entity)
90 if (!list_empty(&entity->list))
93 list_add_tail(&entity->list, &rq->entities);
94 spin_unlock(&rq->lock);
98 * drm_sched_rq_remove_entity - remove an entity
100 * @rq: scheduler run queue
101 * @entity: scheduler entity
103 * Removes a scheduler entity from the run queue.
105 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
106 struct drm_sched_entity *entity)
108 if (list_empty(&entity->list))
110 spin_lock(&rq->lock);
111 list_del_init(&entity->list);
112 if (rq->current_entity == entity)
113 rq->current_entity = NULL;
114 spin_unlock(&rq->lock);
118 * drm_sched_rq_select_entity - Select an entity which could provide a job to run
120 * @rq: scheduler run queue to check.
122 * Try to find a ready entity, returns NULL if none found.
124 static struct drm_sched_entity *
125 drm_sched_rq_select_entity(struct drm_sched_rq *rq)
127 struct drm_sched_entity *entity;
129 spin_lock(&rq->lock);
131 entity = rq->current_entity;
133 list_for_each_entry_continue(entity, &rq->entities, list) {
134 if (drm_sched_entity_is_ready(entity)) {
135 rq->current_entity = entity;
136 spin_unlock(&rq->lock);
142 list_for_each_entry(entity, &rq->entities, list) {
144 if (drm_sched_entity_is_ready(entity)) {
145 rq->current_entity = entity;
146 spin_unlock(&rq->lock);
150 if (entity == rq->current_entity)
154 spin_unlock(&rq->lock);
160 * drm_sched_dependency_optimized
162 * @fence: the dependency fence
163 * @entity: the entity which depends on the above fence
165 * Returns true if the dependency can be optimized and false otherwise
167 bool drm_sched_dependency_optimized(struct dma_fence* fence,
168 struct drm_sched_entity *entity)
170 struct drm_gpu_scheduler *sched = entity->rq->sched;
171 struct drm_sched_fence *s_fence;
173 if (!fence || dma_fence_is_signaled(fence))
175 if (fence->context == entity->fence_context)
177 s_fence = to_drm_sched_fence(fence);
178 if (s_fence && s_fence->sched == sched)
183 EXPORT_SYMBOL(drm_sched_dependency_optimized);
186 * drm_sched_start_timeout - start timeout for reset worker
188 * @sched: scheduler instance to start the worker for
190 * Start the timeout for the given scheduler.
192 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
194 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
195 !list_empty(&sched->ring_mirror_list))
196 schedule_delayed_work(&sched->work_tdr, sched->timeout);
200 * drm_sched_fault - immediately start timeout handler
202 * @sched: scheduler where the timeout handling should be started.
204 * Start timeout handling immediately when the driver detects a hardware fault.
206 void drm_sched_fault(struct drm_gpu_scheduler *sched)
208 mod_delayed_work(system_wq, &sched->work_tdr, 0);
210 EXPORT_SYMBOL(drm_sched_fault);
213 * drm_sched_suspend_timeout - Suspend scheduler job timeout
215 * @sched: scheduler instance for which to suspend the timeout
217 * Suspend the delayed work timeout for the scheduler. This is done by
218 * modifying the delayed work timeout to an arbitrary large value,
219 * MAX_SCHEDULE_TIMEOUT in this case. Note that this function can be
220 * called from an IRQ context.
222 * Returns the timeout remaining
225 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
227 unsigned long sched_timeout, now = jiffies;
229 sched_timeout = sched->work_tdr.timer.expires;
232 * Modify the timeout to an arbitrarily large value. This also prevents
233 * the timeout to be restarted when new submissions arrive
235 if (mod_delayed_work(system_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
236 && time_after(sched_timeout, now))
237 return sched_timeout - now;
239 return sched->timeout;
241 EXPORT_SYMBOL(drm_sched_suspend_timeout);
244 * drm_sched_resume_timeout - Resume scheduler job timeout
246 * @sched: scheduler instance for which to resume the timeout
247 * @remaining: remaining timeout
249 * Resume the delayed work timeout for the scheduler. Note that
250 * this function can be called from an IRQ context.
252 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
253 unsigned long remaining)
257 spin_lock_irqsave(&sched->job_list_lock, flags);
259 if (list_empty(&sched->ring_mirror_list))
260 cancel_delayed_work(&sched->work_tdr);
262 mod_delayed_work(system_wq, &sched->work_tdr, remaining);
264 spin_unlock_irqrestore(&sched->job_list_lock, flags);
266 EXPORT_SYMBOL(drm_sched_resume_timeout);
268 static void drm_sched_job_begin(struct drm_sched_job *s_job)
270 struct drm_gpu_scheduler *sched = s_job->sched;
273 spin_lock_irqsave(&sched->job_list_lock, flags);
274 list_add_tail(&s_job->node, &sched->ring_mirror_list);
275 drm_sched_start_timeout(sched);
276 spin_unlock_irqrestore(&sched->job_list_lock, flags);
279 static void drm_sched_job_timedout(struct work_struct *work)
281 struct drm_gpu_scheduler *sched;
282 struct drm_sched_job *job;
285 sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
286 job = list_first_entry_or_null(&sched->ring_mirror_list,
287 struct drm_sched_job, node);
290 job->sched->ops->timedout_job(job);
293 * Guilty job did complete and hence needs to be manually removed
294 * See drm_sched_stop doc.
296 if (list_empty(&job->node))
297 job->sched->ops->free_job(job);
299 spin_lock_irqsave(&sched->job_list_lock, flags);
300 drm_sched_start_timeout(sched);
301 spin_unlock_irqrestore(&sched->job_list_lock, flags);
305 * drm_sched_increase_karma - Update sched_entity guilty flag
307 * @bad: The job guilty of time out
309 * Increment on every hang caused by the 'bad' job. If this exceeds the hang
310 * limit of the scheduler then the respective sched entity is marked guilty and
311 * jobs from it will not be scheduled further
313 void drm_sched_increase_karma(struct drm_sched_job *bad)
316 struct drm_sched_entity *tmp;
317 struct drm_sched_entity *entity;
318 struct drm_gpu_scheduler *sched = bad->sched;
320 /* don't increase @bad's karma if it's from KERNEL RQ,
321 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
322 * corrupt but keep in mind that kernel jobs always considered good.
324 if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
325 atomic_inc(&bad->karma);
326 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
328 struct drm_sched_rq *rq = &sched->sched_rq[i];
330 spin_lock(&rq->lock);
331 list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
332 if (bad->s_fence->scheduled.context ==
333 entity->fence_context) {
334 if (atomic_read(&bad->karma) >
335 bad->sched->hang_limit)
337 atomic_set(entity->guilty, 1);
341 spin_unlock(&rq->lock);
342 if (&entity->list != &rq->entities)
347 EXPORT_SYMBOL(drm_sched_increase_karma);
350 * drm_sched_hw_job_reset - stop the scheduler if it contains the bad job
352 * @sched: scheduler instance
353 * @bad: bad scheduler job
355 * Stop the scheduler and also removes and frees all completed jobs.
356 * Note: bad job will not be freed as it might be used later and so it's
357 * callers responsibility to release it manually if it's not part of the
358 * mirror list any more.
361 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
363 struct drm_sched_job *s_job, *tmp;
366 kthread_park(sched->thread);
369 * Iterate the job list from later to earlier one and either deactive
370 * their HW callbacks or remove them from mirror list if they already
372 * This iteration is thread safe as sched thread is stopped.
374 list_for_each_entry_safe_reverse(s_job, tmp, &sched->ring_mirror_list, node) {
375 if (s_job->s_fence->parent &&
376 dma_fence_remove_callback(s_job->s_fence->parent,
378 atomic_dec(&sched->hw_rq_count);
381 * remove job from ring_mirror_list.
382 * Locking here is for concurrent resume timeout
384 spin_lock_irqsave(&sched->job_list_lock, flags);
385 list_del_init(&s_job->node);
386 spin_unlock_irqrestore(&sched->job_list_lock, flags);
389 * Wait for job's HW fence callback to finish using s_job
390 * before releasing it.
392 * Job is still alive so fence refcount at least 1
394 dma_fence_wait(&s_job->s_fence->finished, false);
397 * We must keep bad job alive for later use during
398 * recovery by some of the drivers
401 sched->ops->free_job(s_job);
406 * Stop pending timer in flight as we rearm it in drm_sched_start. This
407 * avoids the pending timeout work in progress to fire right away after
408 * this TDR finished and before the newly restarted jobs had a
409 * chance to complete.
411 cancel_delayed_work(&sched->work_tdr);
414 EXPORT_SYMBOL(drm_sched_stop);
417 * drm_sched_job_recovery - recover jobs after a reset
419 * @sched: scheduler instance
422 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
424 struct drm_sched_job *s_job, *tmp;
429 * Locking the list is not required here as the sched thread is parked
430 * so no new jobs are being inserted or removed. Also concurrent
431 * GPU recovers can't run in parallel.
433 list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
434 struct dma_fence *fence = s_job->s_fence->parent;
436 atomic_inc(&sched->hw_rq_count);
442 r = dma_fence_add_callback(fence, &s_job->cb,
443 drm_sched_process_job);
445 drm_sched_process_job(fence, &s_job->cb);
447 DRM_ERROR("fence add callback failed (%d)\n",
450 drm_sched_process_job(NULL, &s_job->cb);
454 spin_lock_irqsave(&sched->job_list_lock, flags);
455 drm_sched_start_timeout(sched);
456 spin_unlock_irqrestore(&sched->job_list_lock, flags);
459 kthread_unpark(sched->thread);
461 EXPORT_SYMBOL(drm_sched_start);
464 * drm_sched_resubmit_jobs - helper to relunch job from mirror ring list
466 * @sched: scheduler instance
469 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
471 struct drm_sched_job *s_job, *tmp;
472 uint64_t guilty_context;
473 bool found_guilty = false;
475 list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
476 struct drm_sched_fence *s_fence = s_job->s_fence;
478 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
480 guilty_context = s_job->s_fence->scheduled.context;
483 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
484 dma_fence_set_error(&s_fence->finished, -ECANCELED);
486 dma_fence_put(s_job->s_fence->parent);
487 s_job->s_fence->parent = sched->ops->run_job(s_job);
490 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
493 * drm_sched_job_init - init a scheduler job
495 * @job: scheduler job to init
496 * @entity: scheduler entity to use
497 * @owner: job owner for debugging
499 * Refer to drm_sched_entity_push_job() documentation
500 * for locking considerations.
502 * Returns 0 for success, negative error code otherwise.
504 int drm_sched_job_init(struct drm_sched_job *job,
505 struct drm_sched_entity *entity,
508 struct drm_gpu_scheduler *sched;
510 drm_sched_entity_select_rq(entity);
514 sched = entity->rq->sched;
517 job->entity = entity;
518 job->s_priority = entity->rq - sched->sched_rq;
519 job->s_fence = drm_sched_fence_create(entity, owner);
522 job->id = atomic64_inc_return(&sched->job_id_count);
524 INIT_LIST_HEAD(&job->node);
528 EXPORT_SYMBOL(drm_sched_job_init);
531 * drm_sched_job_cleanup - clean up scheduler job resources
533 * @job: scheduler job to clean up
535 void drm_sched_job_cleanup(struct drm_sched_job *job)
537 dma_fence_put(&job->s_fence->finished);
540 EXPORT_SYMBOL(drm_sched_job_cleanup);
543 * drm_sched_ready - is the scheduler ready
545 * @sched: scheduler instance
547 * Return true if we can push more jobs to the hw, otherwise false.
549 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
551 return atomic_read(&sched->hw_rq_count) <
552 sched->hw_submission_limit;
556 * drm_sched_wakeup - Wake up the scheduler when it is ready
558 * @sched: scheduler instance
561 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
563 if (drm_sched_ready(sched))
564 wake_up_interruptible(&sched->wake_up_worker);
568 * drm_sched_select_entity - Select next entity to process
570 * @sched: scheduler instance
572 * Returns the entity to process or NULL if none are found.
574 static struct drm_sched_entity *
575 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
577 struct drm_sched_entity *entity;
580 if (!drm_sched_ready(sched))
583 /* Kernel run queue has higher priority than normal run queue*/
584 for (i = DRM_SCHED_PRIORITY_MAX - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
585 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
594 * drm_sched_process_job - process a job
597 * @cb: fence callbacks
599 * Called after job has finished execution.
601 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
603 struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
604 struct drm_sched_fence *s_fence = s_job->s_fence;
605 struct drm_gpu_scheduler *sched = s_fence->sched;
607 atomic_dec(&sched->hw_rq_count);
608 atomic_dec(&sched->num_jobs);
610 trace_drm_sched_process_job(s_fence);
612 drm_sched_fence_finished(s_fence);
613 wake_up_interruptible(&sched->wake_up_worker);
617 * drm_sched_cleanup_jobs - destroy finished jobs
619 * @sched: scheduler instance
621 * Remove all finished jobs from the mirror list and destroy them.
623 static void drm_sched_cleanup_jobs(struct drm_gpu_scheduler *sched)
627 /* Don't destroy jobs while the timeout worker is running */
628 if (!cancel_delayed_work(&sched->work_tdr))
632 while (!list_empty(&sched->ring_mirror_list)) {
633 struct drm_sched_job *job;
635 job = list_first_entry(&sched->ring_mirror_list,
636 struct drm_sched_job, node);
637 if (!dma_fence_is_signaled(&job->s_fence->finished))
640 spin_lock_irqsave(&sched->job_list_lock, flags);
641 /* remove job from ring_mirror_list */
642 list_del_init(&job->node);
643 spin_unlock_irqrestore(&sched->job_list_lock, flags);
645 sched->ops->free_job(job);
648 /* queue timeout for next job */
649 spin_lock_irqsave(&sched->job_list_lock, flags);
650 drm_sched_start_timeout(sched);
651 spin_unlock_irqrestore(&sched->job_list_lock, flags);
656 * drm_sched_blocked - check if the scheduler is blocked
658 * @sched: scheduler instance
660 * Returns true if blocked, otherwise false.
662 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
664 if (kthread_should_park()) {
673 * drm_sched_main - main scheduler thread
675 * @param: scheduler instance
679 static int drm_sched_main(void *param)
681 struct sched_param sparam = {.sched_priority = 1};
682 struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
685 sched_setscheduler(current, SCHED_FIFO, &sparam);
687 while (!kthread_should_stop()) {
688 struct drm_sched_entity *entity = NULL;
689 struct drm_sched_fence *s_fence;
690 struct drm_sched_job *sched_job;
691 struct dma_fence *fence;
693 wait_event_interruptible(sched->wake_up_worker,
694 (drm_sched_cleanup_jobs(sched),
695 (!drm_sched_blocked(sched) &&
696 (entity = drm_sched_select_entity(sched))) ||
697 kthread_should_stop()));
702 sched_job = drm_sched_entity_pop_job(entity);
706 s_fence = sched_job->s_fence;
708 atomic_inc(&sched->hw_rq_count);
709 drm_sched_job_begin(sched_job);
711 fence = sched->ops->run_job(sched_job);
712 drm_sched_fence_scheduled(s_fence);
715 s_fence->parent = dma_fence_get(fence);
716 r = dma_fence_add_callback(fence, &sched_job->cb,
717 drm_sched_process_job);
719 drm_sched_process_job(fence, &sched_job->cb);
721 DRM_ERROR("fence add callback failed (%d)\n",
723 dma_fence_put(fence);
725 drm_sched_process_job(NULL, &sched_job->cb);
727 wake_up(&sched->job_scheduled);
733 * drm_sched_init - Init a gpu scheduler instance
735 * @sched: scheduler instance
736 * @ops: backend operations for this scheduler
737 * @hw_submission: number of hw submissions that can be in flight
738 * @hang_limit: number of times to allow a job to hang before dropping it
739 * @timeout: timeout value in jiffies for the scheduler
740 * @name: name used for debugging
742 * Return 0 on success, otherwise error code.
744 int drm_sched_init(struct drm_gpu_scheduler *sched,
745 const struct drm_sched_backend_ops *ops,
746 unsigned hw_submission,
753 sched->hw_submission_limit = hw_submission;
755 sched->timeout = timeout;
756 sched->hang_limit = hang_limit;
757 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++)
758 drm_sched_rq_init(sched, &sched->sched_rq[i]);
760 init_waitqueue_head(&sched->wake_up_worker);
761 init_waitqueue_head(&sched->job_scheduled);
762 INIT_LIST_HEAD(&sched->ring_mirror_list);
763 spin_lock_init(&sched->job_list_lock);
764 atomic_set(&sched->hw_rq_count, 0);
765 INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
766 atomic_set(&sched->num_jobs, 0);
767 atomic64_set(&sched->job_id_count, 0);
769 /* Each scheduler will run on a seperate kernel thread */
770 sched->thread = kthread_run(drm_sched_main, sched, sched->name);
771 if (IS_ERR(sched->thread)) {
772 ret = PTR_ERR(sched->thread);
773 sched->thread = NULL;
774 DRM_ERROR("Failed to create scheduler for %s.\n", name);
781 EXPORT_SYMBOL(drm_sched_init);
784 * drm_sched_fini - Destroy a gpu scheduler
786 * @sched: scheduler instance
788 * Tears down and cleans up the scheduler.
790 void drm_sched_fini(struct drm_gpu_scheduler *sched)
793 kthread_stop(sched->thread);
795 sched->ready = false;
797 EXPORT_SYMBOL(drm_sched_fini);