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 <linux/completion.h>
51 #include <linux/dma-resv.h>
52 #include <uapi/linux/sched/types.h>
54 #include <drm/drm_print.h>
55 #include <drm/drm_gem.h>
56 #include <drm/gpu_scheduler.h>
57 #include <drm/spsc_queue.h>
59 #define CREATE_TRACE_POINTS
60 #include "gpu_scheduler_trace.h"
62 #define to_drm_sched_job(sched_job) \
63 container_of((sched_job), struct drm_sched_job, queue_node)
66 * drm_sched_rq_init - initialize a given run queue struct
68 * @sched: scheduler instance to associate with this run queue
69 * @rq: scheduler run queue
71 * Initializes a scheduler runqueue.
73 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
74 struct drm_sched_rq *rq)
76 spin_lock_init(&rq->lock);
77 INIT_LIST_HEAD(&rq->entities);
78 rq->current_entity = NULL;
83 * drm_sched_rq_add_entity - add an entity
85 * @rq: scheduler run queue
86 * @entity: scheduler entity
88 * Adds a scheduler entity to the run queue.
90 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
91 struct drm_sched_entity *entity)
93 if (!list_empty(&entity->list))
96 atomic_inc(rq->sched->score);
97 list_add_tail(&entity->list, &rq->entities);
98 spin_unlock(&rq->lock);
102 * drm_sched_rq_remove_entity - remove an entity
104 * @rq: scheduler run queue
105 * @entity: scheduler entity
107 * Removes a scheduler entity from the run queue.
109 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
110 struct drm_sched_entity *entity)
112 if (list_empty(&entity->list))
114 spin_lock(&rq->lock);
115 atomic_dec(rq->sched->score);
116 list_del_init(&entity->list);
117 if (rq->current_entity == entity)
118 rq->current_entity = NULL;
119 spin_unlock(&rq->lock);
123 * drm_sched_rq_select_entity - Select an entity which could provide a job to run
125 * @rq: scheduler run queue to check.
127 * Try to find a ready entity, returns NULL if none found.
129 static struct drm_sched_entity *
130 drm_sched_rq_select_entity(struct drm_sched_rq *rq)
132 struct drm_sched_entity *entity;
134 spin_lock(&rq->lock);
136 entity = rq->current_entity;
138 list_for_each_entry_continue(entity, &rq->entities, list) {
139 if (drm_sched_entity_is_ready(entity)) {
140 rq->current_entity = entity;
141 reinit_completion(&entity->entity_idle);
142 spin_unlock(&rq->lock);
148 list_for_each_entry(entity, &rq->entities, list) {
150 if (drm_sched_entity_is_ready(entity)) {
151 rq->current_entity = entity;
152 reinit_completion(&entity->entity_idle);
153 spin_unlock(&rq->lock);
157 if (entity == rq->current_entity)
161 spin_unlock(&rq->lock);
167 * drm_sched_job_done - complete a job
168 * @s_job: pointer to the job which is done
170 * Finish the job's fence and wake up the worker thread.
172 static void drm_sched_job_done(struct drm_sched_job *s_job)
174 struct drm_sched_fence *s_fence = s_job->s_fence;
175 struct drm_gpu_scheduler *sched = s_fence->sched;
177 atomic_dec(&sched->hw_rq_count);
178 atomic_dec(sched->score);
180 trace_drm_sched_process_job(s_fence);
182 dma_fence_get(&s_fence->finished);
183 drm_sched_fence_finished(s_fence);
184 dma_fence_put(&s_fence->finished);
185 wake_up_interruptible(&sched->wake_up_worker);
189 * drm_sched_job_done_cb - the callback for a done job
191 * @cb: fence callbacks
193 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
195 struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
197 drm_sched_job_done(s_job);
201 * drm_sched_dependency_optimized
203 * @fence: the dependency fence
204 * @entity: the entity which depends on the above fence
206 * Returns true if the dependency can be optimized and false otherwise
208 bool drm_sched_dependency_optimized(struct dma_fence* fence,
209 struct drm_sched_entity *entity)
211 struct drm_gpu_scheduler *sched = entity->rq->sched;
212 struct drm_sched_fence *s_fence;
214 if (!fence || dma_fence_is_signaled(fence))
216 if (fence->context == entity->fence_context)
218 s_fence = to_drm_sched_fence(fence);
219 if (s_fence && s_fence->sched == sched)
224 EXPORT_SYMBOL(drm_sched_dependency_optimized);
227 * drm_sched_start_timeout - start timeout for reset worker
229 * @sched: scheduler instance to start the worker for
231 * Start the timeout for the given scheduler.
233 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
235 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
236 !list_empty(&sched->pending_list))
237 queue_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
241 * drm_sched_fault - immediately start timeout handler
243 * @sched: scheduler where the timeout handling should be started.
245 * Start timeout handling immediately when the driver detects a hardware fault.
247 void drm_sched_fault(struct drm_gpu_scheduler *sched)
249 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
251 EXPORT_SYMBOL(drm_sched_fault);
254 * drm_sched_suspend_timeout - Suspend scheduler job timeout
256 * @sched: scheduler instance for which to suspend the timeout
258 * Suspend the delayed work timeout for the scheduler. This is done by
259 * modifying the delayed work timeout to an arbitrary large value,
260 * MAX_SCHEDULE_TIMEOUT in this case.
262 * Returns the timeout remaining
265 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
267 unsigned long sched_timeout, now = jiffies;
269 sched_timeout = sched->work_tdr.timer.expires;
272 * Modify the timeout to an arbitrarily large value. This also prevents
273 * the timeout to be restarted when new submissions arrive
275 if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
276 && time_after(sched_timeout, now))
277 return sched_timeout - now;
279 return sched->timeout;
281 EXPORT_SYMBOL(drm_sched_suspend_timeout);
284 * drm_sched_resume_timeout - Resume scheduler job timeout
286 * @sched: scheduler instance for which to resume the timeout
287 * @remaining: remaining timeout
289 * Resume the delayed work timeout for the scheduler.
291 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
292 unsigned long remaining)
294 spin_lock(&sched->job_list_lock);
296 if (list_empty(&sched->pending_list))
297 cancel_delayed_work(&sched->work_tdr);
299 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
301 spin_unlock(&sched->job_list_lock);
303 EXPORT_SYMBOL(drm_sched_resume_timeout);
305 static void drm_sched_job_begin(struct drm_sched_job *s_job)
307 struct drm_gpu_scheduler *sched = s_job->sched;
309 spin_lock(&sched->job_list_lock);
310 list_add_tail(&s_job->list, &sched->pending_list);
311 drm_sched_start_timeout(sched);
312 spin_unlock(&sched->job_list_lock);
315 static void drm_sched_job_timedout(struct work_struct *work)
317 struct drm_gpu_scheduler *sched;
318 struct drm_sched_job *job;
319 enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
321 sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
323 /* Protects against concurrent deletion in drm_sched_get_cleanup_job */
324 spin_lock(&sched->job_list_lock);
325 job = list_first_entry_or_null(&sched->pending_list,
326 struct drm_sched_job, list);
330 * Remove the bad job so it cannot be freed by concurrent
331 * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
332 * is parked at which point it's safe.
334 list_del_init(&job->list);
335 spin_unlock(&sched->job_list_lock);
337 status = job->sched->ops->timedout_job(job);
340 * Guilty job did complete and hence needs to be manually removed
341 * See drm_sched_stop doc.
343 if (sched->free_guilty) {
344 job->sched->ops->free_job(job);
345 sched->free_guilty = false;
348 spin_unlock(&sched->job_list_lock);
351 if (status != DRM_GPU_SCHED_STAT_ENODEV) {
352 spin_lock(&sched->job_list_lock);
353 drm_sched_start_timeout(sched);
354 spin_unlock(&sched->job_list_lock);
359 * drm_sched_increase_karma - Update sched_entity guilty flag
361 * @bad: The job guilty of time out
363 * Increment on every hang caused by the 'bad' job. If this exceeds the hang
364 * limit of the scheduler then the respective sched entity is marked guilty and
365 * jobs from it will not be scheduled further
367 void drm_sched_increase_karma(struct drm_sched_job *bad)
369 drm_sched_increase_karma_ext(bad, 1);
371 EXPORT_SYMBOL(drm_sched_increase_karma);
373 void drm_sched_reset_karma(struct drm_sched_job *bad)
375 drm_sched_increase_karma_ext(bad, 0);
377 EXPORT_SYMBOL(drm_sched_reset_karma);
380 * drm_sched_stop - stop the scheduler
382 * @sched: scheduler instance
383 * @bad: job which caused the time out
385 * Stop the scheduler and also removes and frees all completed jobs.
386 * Note: bad job will not be freed as it might be used later and so it's
387 * callers responsibility to release it manually if it's not part of the
388 * pending list any more.
391 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
393 struct drm_sched_job *s_job, *tmp;
395 kthread_park(sched->thread);
398 * Reinsert back the bad job here - now it's safe as
399 * drm_sched_get_cleanup_job cannot race against us and release the
400 * bad job at this point - we parked (waited for) any in progress
401 * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
402 * now until the scheduler thread is unparked.
404 if (bad && bad->sched == sched)
406 * Add at the head of the queue to reflect it was the earliest
409 list_add(&bad->list, &sched->pending_list);
412 * Iterate the job list from later to earlier one and either deactive
413 * their HW callbacks or remove them from pending list if they already
415 * This iteration is thread safe as sched thread is stopped.
417 list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
419 if (s_job->s_fence->parent &&
420 dma_fence_remove_callback(s_job->s_fence->parent,
422 atomic_dec(&sched->hw_rq_count);
425 * remove job from pending_list.
426 * Locking here is for concurrent resume timeout
428 spin_lock(&sched->job_list_lock);
429 list_del_init(&s_job->list);
430 spin_unlock(&sched->job_list_lock);
433 * Wait for job's HW fence callback to finish using s_job
434 * before releasing it.
436 * Job is still alive so fence refcount at least 1
438 dma_fence_wait(&s_job->s_fence->finished, false);
441 * We must keep bad job alive for later use during
442 * recovery by some of the drivers but leave a hint
443 * that the guilty job must be released.
446 sched->ops->free_job(s_job);
448 sched->free_guilty = true;
453 * Stop pending timer in flight as we rearm it in drm_sched_start. This
454 * avoids the pending timeout work in progress to fire right away after
455 * this TDR finished and before the newly restarted jobs had a
456 * chance to complete.
458 cancel_delayed_work(&sched->work_tdr);
461 EXPORT_SYMBOL(drm_sched_stop);
464 * drm_sched_start - recover jobs after a reset
466 * @sched: scheduler instance
467 * @full_recovery: proceed with complete sched restart
470 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
472 struct drm_sched_job *s_job, *tmp;
476 * Locking the list is not required here as the sched thread is parked
477 * so no new jobs are being inserted or removed. Also concurrent
478 * GPU recovers can't run in parallel.
480 list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
481 struct dma_fence *fence = s_job->s_fence->parent;
483 atomic_inc(&sched->hw_rq_count);
489 r = dma_fence_add_callback(fence, &s_job->cb,
490 drm_sched_job_done_cb);
492 drm_sched_job_done(s_job);
494 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
497 drm_sched_job_done(s_job);
501 spin_lock(&sched->job_list_lock);
502 drm_sched_start_timeout(sched);
503 spin_unlock(&sched->job_list_lock);
506 kthread_unpark(sched->thread);
508 EXPORT_SYMBOL(drm_sched_start);
511 * drm_sched_resubmit_jobs - helper to relaunch jobs from the pending list
513 * @sched: scheduler instance
516 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
518 drm_sched_resubmit_jobs_ext(sched, INT_MAX);
520 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
523 * drm_sched_resubmit_jobs_ext - helper to relunch certain number of jobs from mirror ring list
525 * @sched: scheduler instance
526 * @max: job numbers to relaunch
529 void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)
531 struct drm_sched_job *s_job, *tmp;
532 uint64_t guilty_context;
533 bool found_guilty = false;
534 struct dma_fence *fence;
537 list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
538 struct drm_sched_fence *s_fence = s_job->s_fence;
543 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
545 guilty_context = s_job->s_fence->scheduled.context;
548 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
549 dma_fence_set_error(&s_fence->finished, -ECANCELED);
551 dma_fence_put(s_job->s_fence->parent);
552 fence = sched->ops->run_job(s_job);
555 if (IS_ERR_OR_NULL(fence)) {
557 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
559 s_job->s_fence->parent = NULL;
561 s_job->s_fence->parent = fence;
565 EXPORT_SYMBOL(drm_sched_resubmit_jobs_ext);
568 * drm_sched_job_init - init a scheduler job
569 * @job: scheduler job to init
570 * @entity: scheduler entity to use
571 * @owner: job owner for debugging
573 * Refer to drm_sched_entity_push_job() documentation
574 * for locking considerations.
576 * Drivers must make sure drm_sched_job_cleanup() if this function returns
577 * successfully, even when @job is aborted before drm_sched_job_arm() is called.
579 * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
580 * has died, which can mean that there's no valid runqueue for a @entity.
581 * This function returns -ENOENT in this case (which probably should be -EIO as
582 * a more meanigful return value).
584 * Returns 0 for success, negative error code otherwise.
586 int drm_sched_job_init(struct drm_sched_job *job,
587 struct drm_sched_entity *entity,
590 drm_sched_entity_select_rq(entity);
594 job->entity = entity;
595 job->s_fence = drm_sched_fence_alloc(entity, owner);
599 INIT_LIST_HEAD(&job->list);
601 xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
605 EXPORT_SYMBOL(drm_sched_job_init);
608 * drm_sched_job_arm - arm a scheduler job for execution
609 * @job: scheduler job to arm
611 * This arms a scheduler job for execution. Specifically it initializes the
612 * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
613 * or other places that need to track the completion of this job.
615 * Refer to drm_sched_entity_push_job() documentation for locking
618 * This can only be called if drm_sched_job_init() succeeded.
620 void drm_sched_job_arm(struct drm_sched_job *job)
622 struct drm_gpu_scheduler *sched;
623 struct drm_sched_entity *entity = job->entity;
627 sched = entity->rq->sched;
630 job->s_priority = entity->rq - sched->sched_rq;
631 job->id = atomic64_inc_return(&sched->job_id_count);
633 drm_sched_fence_init(job->s_fence, job->entity);
635 EXPORT_SYMBOL(drm_sched_job_arm);
638 * drm_sched_job_add_dependency - adds the fence as a job dependency
639 * @job: scheduler job to add the dependencies to
640 * @fence: the dma_fence to add to the list of dependencies.
642 * Note that @fence is consumed in both the success and error cases.
645 * 0 on success, or an error on failing to expand the array.
647 int drm_sched_job_add_dependency(struct drm_sched_job *job,
648 struct dma_fence *fence)
650 struct dma_fence *entry;
658 /* Deduplicate if we already depend on a fence from the same context.
659 * This lets the size of the array of deps scale with the number of
660 * engines involved, rather than the number of BOs.
662 xa_for_each(&job->dependencies, index, entry) {
663 if (entry->context != fence->context)
666 if (dma_fence_is_later(fence, entry)) {
667 dma_fence_put(entry);
668 xa_store(&job->dependencies, index, fence, GFP_KERNEL);
670 dma_fence_put(fence);
675 ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
677 dma_fence_put(fence);
681 EXPORT_SYMBOL(drm_sched_job_add_dependency);
684 * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
686 * @job: scheduler job to add the dependencies to
687 * @obj: the gem object to add new dependencies from.
688 * @write: whether the job might write the object (so we need to depend on
689 * shared fences in the reservation object).
691 * This should be called after drm_gem_lock_reservations() on your array of
692 * GEM objects used in the job but before updating the reservations with your
696 * 0 on success, or an error on failing to expand the array.
698 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
699 struct drm_gem_object *obj,
702 struct dma_resv_iter cursor;
703 struct dma_fence *fence;
706 dma_resv_assert_held(obj->resv);
708 dma_resv_for_each_fence(&cursor, obj->resv, dma_resv_usage_rw(write),
710 /* Make sure to grab an additional ref on the added fence */
711 dma_fence_get(fence);
712 ret = drm_sched_job_add_dependency(job, fence);
714 dma_fence_put(fence);
720 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
724 * drm_sched_job_cleanup - clean up scheduler job resources
725 * @job: scheduler job to clean up
727 * Cleans up the resources allocated with drm_sched_job_init().
729 * Drivers should call this from their error unwind code if @job is aborted
730 * before drm_sched_job_arm() is called.
732 * After that point of no return @job is committed to be executed by the
733 * scheduler, and this function should be called from the
734 * &drm_sched_backend_ops.free_job callback.
736 void drm_sched_job_cleanup(struct drm_sched_job *job)
738 struct dma_fence *fence;
741 if (kref_read(&job->s_fence->finished.refcount)) {
742 /* drm_sched_job_arm() has been called */
743 dma_fence_put(&job->s_fence->finished);
745 /* aborted job before committing to run it */
746 drm_sched_fence_free(job->s_fence);
751 xa_for_each(&job->dependencies, index, fence) {
752 dma_fence_put(fence);
754 xa_destroy(&job->dependencies);
757 EXPORT_SYMBOL(drm_sched_job_cleanup);
760 * drm_sched_ready - is the scheduler ready
762 * @sched: scheduler instance
764 * Return true if we can push more jobs to the hw, otherwise false.
766 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
768 return atomic_read(&sched->hw_rq_count) <
769 sched->hw_submission_limit;
773 * drm_sched_wakeup - Wake up the scheduler when it is ready
775 * @sched: scheduler instance
778 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
780 if (drm_sched_ready(sched))
781 wake_up_interruptible(&sched->wake_up_worker);
785 * drm_sched_select_entity - Select next entity to process
787 * @sched: scheduler instance
789 * Returns the entity to process or NULL if none are found.
791 static struct drm_sched_entity *
792 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
794 struct drm_sched_entity *entity;
797 if (!drm_sched_ready(sched))
800 /* Kernel run queue has higher priority than normal run queue*/
801 for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
802 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
811 * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
813 * @sched: scheduler instance
815 * Returns the next finished job from the pending list (if there is one)
816 * ready for it to be destroyed.
818 static struct drm_sched_job *
819 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
821 struct drm_sched_job *job, *next;
823 spin_lock(&sched->job_list_lock);
825 job = list_first_entry_or_null(&sched->pending_list,
826 struct drm_sched_job, list);
828 if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
829 /* remove job from pending_list */
830 list_del_init(&job->list);
832 /* cancel this job's TO timer */
833 cancel_delayed_work(&sched->work_tdr);
834 /* make the scheduled timestamp more accurate */
835 next = list_first_entry_or_null(&sched->pending_list,
836 typeof(*next), list);
839 next->s_fence->scheduled.timestamp =
840 job->s_fence->finished.timestamp;
841 /* start TO timer for next job */
842 drm_sched_start_timeout(sched);
848 spin_unlock(&sched->job_list_lock);
854 * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
855 * @sched_list: list of drm_gpu_schedulers
856 * @num_sched_list: number of drm_gpu_schedulers in the sched_list
858 * Returns pointer of the sched with the least load or NULL if none of the
859 * drm_gpu_schedulers are ready
861 struct drm_gpu_scheduler *
862 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
863 unsigned int num_sched_list)
865 struct drm_gpu_scheduler *sched, *picked_sched = NULL;
867 unsigned int min_score = UINT_MAX, num_score;
869 for (i = 0; i < num_sched_list; ++i) {
870 sched = sched_list[i];
873 DRM_WARN("scheduler %s is not ready, skipping",
878 num_score = atomic_read(sched->score);
879 if (num_score < min_score) {
880 min_score = num_score;
881 picked_sched = sched;
887 EXPORT_SYMBOL(drm_sched_pick_best);
890 * drm_sched_blocked - check if the scheduler is blocked
892 * @sched: scheduler instance
894 * Returns true if blocked, otherwise false.
896 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
898 if (kthread_should_park()) {
907 * drm_sched_main - main scheduler thread
909 * @param: scheduler instance
913 static int drm_sched_main(void *param)
915 struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
918 sched_set_fifo_low(current);
920 while (!kthread_should_stop()) {
921 struct drm_sched_entity *entity = NULL;
922 struct drm_sched_fence *s_fence;
923 struct drm_sched_job *sched_job;
924 struct dma_fence *fence;
925 struct drm_sched_job *cleanup_job = NULL;
927 wait_event_interruptible(sched->wake_up_worker,
928 (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
929 (!drm_sched_blocked(sched) &&
930 (entity = drm_sched_select_entity(sched))) ||
931 kthread_should_stop());
934 sched->ops->free_job(cleanup_job);
939 sched_job = drm_sched_entity_pop_job(entity);
942 complete(&entity->entity_idle);
946 s_fence = sched_job->s_fence;
948 atomic_inc(&sched->hw_rq_count);
949 drm_sched_job_begin(sched_job);
951 trace_drm_run_job(sched_job, entity);
952 fence = sched->ops->run_job(sched_job);
953 complete(&entity->entity_idle);
954 drm_sched_fence_scheduled(s_fence);
956 if (!IS_ERR_OR_NULL(fence)) {
957 s_fence->parent = dma_fence_get(fence);
958 r = dma_fence_add_callback(fence, &sched_job->cb,
959 drm_sched_job_done_cb);
961 drm_sched_job_done(sched_job);
963 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
965 dma_fence_put(fence);
968 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
970 drm_sched_job_done(sched_job);
973 wake_up(&sched->job_scheduled);
979 * drm_sched_init - Init a gpu scheduler instance
981 * @sched: scheduler instance
982 * @ops: backend operations for this scheduler
983 * @hw_submission: number of hw submissions that can be in flight
984 * @hang_limit: number of times to allow a job to hang before dropping it
985 * @timeout: timeout value in jiffies for the scheduler
986 * @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
988 * @score: optional score atomic shared with other schedulers
989 * @name: name used for debugging
991 * Return 0 on success, otherwise error code.
993 int drm_sched_init(struct drm_gpu_scheduler *sched,
994 const struct drm_sched_backend_ops *ops,
995 unsigned hw_submission, unsigned hang_limit,
996 long timeout, struct workqueue_struct *timeout_wq,
997 atomic_t *score, const char *name, struct device *dev)
1001 sched->hw_submission_limit = hw_submission;
1003 sched->timeout = timeout;
1004 sched->timeout_wq = timeout_wq ? : system_wq;
1005 sched->hang_limit = hang_limit;
1006 sched->score = score ? score : &sched->_score;
1008 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
1009 drm_sched_rq_init(sched, &sched->sched_rq[i]);
1011 init_waitqueue_head(&sched->wake_up_worker);
1012 init_waitqueue_head(&sched->job_scheduled);
1013 INIT_LIST_HEAD(&sched->pending_list);
1014 spin_lock_init(&sched->job_list_lock);
1015 atomic_set(&sched->hw_rq_count, 0);
1016 INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1017 atomic_set(&sched->_score, 0);
1018 atomic64_set(&sched->job_id_count, 0);
1020 /* Each scheduler will run on a seperate kernel thread */
1021 sched->thread = kthread_run(drm_sched_main, sched, sched->name);
1022 if (IS_ERR(sched->thread)) {
1023 ret = PTR_ERR(sched->thread);
1024 sched->thread = NULL;
1025 DRM_DEV_ERROR(sched->dev, "Failed to create scheduler for %s.\n", name);
1029 sched->ready = true;
1032 EXPORT_SYMBOL(drm_sched_init);
1035 * drm_sched_fini - Destroy a gpu scheduler
1037 * @sched: scheduler instance
1039 * Tears down and cleans up the scheduler.
1041 void drm_sched_fini(struct drm_gpu_scheduler *sched)
1043 struct drm_sched_entity *s_entity;
1047 kthread_stop(sched->thread);
1049 for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
1050 struct drm_sched_rq *rq = &sched->sched_rq[i];
1055 spin_lock(&rq->lock);
1056 list_for_each_entry(s_entity, &rq->entities, list)
1058 * Prevents reinsertion and marks job_queue as idle,
1059 * it will removed from rq in drm_sched_entity_fini
1062 s_entity->stopped = true;
1063 spin_unlock(&rq->lock);
1067 /* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1068 wake_up_all(&sched->job_scheduled);
1070 /* Confirm no work left behind accessing device structures */
1071 cancel_delayed_work_sync(&sched->work_tdr);
1073 sched->ready = false;
1075 EXPORT_SYMBOL(drm_sched_fini);
1078 * drm_sched_increase_karma_ext - Update sched_entity guilty flag
1080 * @bad: The job guilty of time out
1081 * @type: type for increase/reset karma
1084 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type)
1087 struct drm_sched_entity *tmp;
1088 struct drm_sched_entity *entity;
1089 struct drm_gpu_scheduler *sched = bad->sched;
1091 /* don't change @bad's karma if it's from KERNEL RQ,
1092 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1093 * corrupt but keep in mind that kernel jobs always considered good.
1095 if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1097 atomic_set(&bad->karma, 0);
1099 atomic_inc(&bad->karma);
1101 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
1103 struct drm_sched_rq *rq = &sched->sched_rq[i];
1105 spin_lock(&rq->lock);
1106 list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1107 if (bad->s_fence->scheduled.context ==
1108 entity->fence_context) {
1110 atomic_set(entity->guilty, type);
1114 spin_unlock(&rq->lock);
1115 if (&entity->list != &rq->entities)
1120 EXPORT_SYMBOL(drm_sched_increase_karma_ext);