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.
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.
51 #include <linux/kthread.h>
52 #include <linux/wait.h>
53 #include <linux/sched.h>
54 #include <linux/completion.h>
55 #include <linux/dma-resv.h>
56 #include <uapi/linux/sched/types.h>
58 #include <drm/drm_print.h>
59 #include <drm/drm_gem.h>
60 #include <drm/drm_syncobj.h>
61 #include <drm/gpu_scheduler.h>
62 #include <drm/spsc_queue.h>
64 #define CREATE_TRACE_POINTS
65 #include "gpu_scheduler_trace.h"
67 #define to_drm_sched_job(sched_job) \
68 container_of((sched_job), struct drm_sched_job, queue_node)
70 int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
73 * DOC: sched_policy (int)
74 * Used to override default entities scheduling policy in a run queue.
76 MODULE_PARM_DESC(sched_policy, "Specify the scheduling policy for entities on a run-queue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin, " __stringify(DRM_SCHED_POLICY_FIFO) " = FIFO (default).");
77 module_param_named(sched_policy, drm_sched_policy, int, 0444);
79 static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
80 const struct rb_node *b)
82 struct drm_sched_entity *ent_a = rb_entry((a), struct drm_sched_entity, rb_tree_node);
83 struct drm_sched_entity *ent_b = rb_entry((b), struct drm_sched_entity, rb_tree_node);
85 return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
88 static inline void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity)
90 struct drm_sched_rq *rq = entity->rq;
92 if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
93 rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
94 RB_CLEAR_NODE(&entity->rb_tree_node);
98 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts)
101 * Both locks need to be grabbed, one to protect from entity->rq change
102 * for entity from within concurrent drm_sched_entity_select_rq and the
103 * other to update the rb tree structure.
105 spin_lock(&entity->rq_lock);
106 spin_lock(&entity->rq->lock);
108 drm_sched_rq_remove_fifo_locked(entity);
110 entity->oldest_job_waiting = ts;
112 rb_add_cached(&entity->rb_tree_node, &entity->rq->rb_tree_root,
113 drm_sched_entity_compare_before);
115 spin_unlock(&entity->rq->lock);
116 spin_unlock(&entity->rq_lock);
120 * drm_sched_rq_init - initialize a given run queue struct
122 * @sched: scheduler instance to associate with this run queue
123 * @rq: scheduler run queue
125 * Initializes a scheduler runqueue.
127 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
128 struct drm_sched_rq *rq)
130 spin_lock_init(&rq->lock);
131 INIT_LIST_HEAD(&rq->entities);
132 rq->rb_tree_root = RB_ROOT_CACHED;
133 rq->current_entity = NULL;
138 * drm_sched_rq_add_entity - add an entity
140 * @rq: scheduler run queue
141 * @entity: scheduler entity
143 * Adds a scheduler entity to the run queue.
145 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
146 struct drm_sched_entity *entity)
148 if (!list_empty(&entity->list))
151 spin_lock(&rq->lock);
153 atomic_inc(rq->sched->score);
154 list_add_tail(&entity->list, &rq->entities);
156 spin_unlock(&rq->lock);
160 * drm_sched_rq_remove_entity - remove an entity
162 * @rq: scheduler run queue
163 * @entity: scheduler entity
165 * Removes a scheduler entity from the run queue.
167 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
168 struct drm_sched_entity *entity)
170 if (list_empty(&entity->list))
173 spin_lock(&rq->lock);
175 atomic_dec(rq->sched->score);
176 list_del_init(&entity->list);
178 if (rq->current_entity == entity)
179 rq->current_entity = NULL;
181 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
182 drm_sched_rq_remove_fifo_locked(entity);
184 spin_unlock(&rq->lock);
188 * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
190 * @rq: scheduler run queue to check.
192 * Try to find a ready entity, returns NULL if none found.
194 static struct drm_sched_entity *
195 drm_sched_rq_select_entity_rr(struct drm_sched_rq *rq)
197 struct drm_sched_entity *entity;
199 spin_lock(&rq->lock);
201 entity = rq->current_entity;
203 list_for_each_entry_continue(entity, &rq->entities, list) {
204 if (drm_sched_entity_is_ready(entity)) {
205 rq->current_entity = entity;
206 reinit_completion(&entity->entity_idle);
207 spin_unlock(&rq->lock);
213 list_for_each_entry(entity, &rq->entities, list) {
215 if (drm_sched_entity_is_ready(entity)) {
216 rq->current_entity = entity;
217 reinit_completion(&entity->entity_idle);
218 spin_unlock(&rq->lock);
222 if (entity == rq->current_entity)
226 spin_unlock(&rq->lock);
232 * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
234 * @rq: scheduler run queue to check.
236 * Find oldest waiting ready entity, returns NULL if none found.
238 static struct drm_sched_entity *
239 drm_sched_rq_select_entity_fifo(struct drm_sched_rq *rq)
243 spin_lock(&rq->lock);
244 for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
245 struct drm_sched_entity *entity;
247 entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
248 if (drm_sched_entity_is_ready(entity)) {
249 rq->current_entity = entity;
250 reinit_completion(&entity->entity_idle);
254 spin_unlock(&rq->lock);
256 return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
260 * drm_sched_job_done - complete a job
261 * @s_job: pointer to the job which is done
263 * Finish the job's fence and wake up the worker thread.
265 static void drm_sched_job_done(struct drm_sched_job *s_job, int result)
267 struct drm_sched_fence *s_fence = s_job->s_fence;
268 struct drm_gpu_scheduler *sched = s_fence->sched;
270 atomic_dec(&sched->hw_rq_count);
271 atomic_dec(sched->score);
273 trace_drm_sched_process_job(s_fence);
275 dma_fence_get(&s_fence->finished);
276 drm_sched_fence_finished(s_fence, result);
277 dma_fence_put(&s_fence->finished);
278 wake_up_interruptible(&sched->wake_up_worker);
282 * drm_sched_job_done_cb - the callback for a done job
284 * @cb: fence callbacks
286 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
288 struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
290 drm_sched_job_done(s_job, f->error);
294 * drm_sched_start_timeout - start timeout for reset worker
296 * @sched: scheduler instance to start the worker for
298 * Start the timeout for the given scheduler.
300 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
302 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
303 !list_empty(&sched->pending_list))
304 queue_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
308 * drm_sched_fault - immediately start timeout handler
310 * @sched: scheduler where the timeout handling should be started.
312 * Start timeout handling immediately when the driver detects a hardware fault.
314 void drm_sched_fault(struct drm_gpu_scheduler *sched)
316 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
318 EXPORT_SYMBOL(drm_sched_fault);
321 * drm_sched_suspend_timeout - Suspend scheduler job timeout
323 * @sched: scheduler instance for which to suspend the timeout
325 * Suspend the delayed work timeout for the scheduler. This is done by
326 * modifying the delayed work timeout to an arbitrary large value,
327 * MAX_SCHEDULE_TIMEOUT in this case.
329 * Returns the timeout remaining
332 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
334 unsigned long sched_timeout, now = jiffies;
336 sched_timeout = sched->work_tdr.timer.expires;
339 * Modify the timeout to an arbitrarily large value. This also prevents
340 * the timeout to be restarted when new submissions arrive
342 if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
343 && time_after(sched_timeout, now))
344 return sched_timeout - now;
346 return sched->timeout;
348 EXPORT_SYMBOL(drm_sched_suspend_timeout);
351 * drm_sched_resume_timeout - Resume scheduler job timeout
353 * @sched: scheduler instance for which to resume the timeout
354 * @remaining: remaining timeout
356 * Resume the delayed work timeout for the scheduler.
358 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
359 unsigned long remaining)
361 spin_lock(&sched->job_list_lock);
363 if (list_empty(&sched->pending_list))
364 cancel_delayed_work(&sched->work_tdr);
366 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
368 spin_unlock(&sched->job_list_lock);
370 EXPORT_SYMBOL(drm_sched_resume_timeout);
372 static void drm_sched_job_begin(struct drm_sched_job *s_job)
374 struct drm_gpu_scheduler *sched = s_job->sched;
376 spin_lock(&sched->job_list_lock);
377 list_add_tail(&s_job->list, &sched->pending_list);
378 drm_sched_start_timeout(sched);
379 spin_unlock(&sched->job_list_lock);
382 static void drm_sched_job_timedout(struct work_struct *work)
384 struct drm_gpu_scheduler *sched;
385 struct drm_sched_job *job;
386 enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
388 sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
390 /* Protects against concurrent deletion in drm_sched_get_cleanup_job */
391 spin_lock(&sched->job_list_lock);
392 job = list_first_entry_or_null(&sched->pending_list,
393 struct drm_sched_job, list);
397 * Remove the bad job so it cannot be freed by concurrent
398 * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
399 * is parked at which point it's safe.
401 list_del_init(&job->list);
402 spin_unlock(&sched->job_list_lock);
404 status = job->sched->ops->timedout_job(job);
407 * Guilty job did complete and hence needs to be manually removed
408 * See drm_sched_stop doc.
410 if (sched->free_guilty) {
411 job->sched->ops->free_job(job);
412 sched->free_guilty = false;
415 spin_unlock(&sched->job_list_lock);
418 if (status != DRM_GPU_SCHED_STAT_ENODEV) {
419 spin_lock(&sched->job_list_lock);
420 drm_sched_start_timeout(sched);
421 spin_unlock(&sched->job_list_lock);
426 * drm_sched_stop - stop the scheduler
428 * @sched: scheduler instance
429 * @bad: job which caused the time out
431 * Stop the scheduler and also removes and frees all completed jobs.
432 * Note: bad job will not be freed as it might be used later and so it's
433 * callers responsibility to release it manually if it's not part of the
434 * pending list any more.
437 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
439 struct drm_sched_job *s_job, *tmp;
441 kthread_park(sched->thread);
444 * Reinsert back the bad job here - now it's safe as
445 * drm_sched_get_cleanup_job cannot race against us and release the
446 * bad job at this point - we parked (waited for) any in progress
447 * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
448 * now until the scheduler thread is unparked.
450 if (bad && bad->sched == sched)
452 * Add at the head of the queue to reflect it was the earliest
455 list_add(&bad->list, &sched->pending_list);
458 * Iterate the job list from later to earlier one and either deactive
459 * their HW callbacks or remove them from pending list if they already
461 * This iteration is thread safe as sched thread is stopped.
463 list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
465 if (s_job->s_fence->parent &&
466 dma_fence_remove_callback(s_job->s_fence->parent,
468 dma_fence_put(s_job->s_fence->parent);
469 s_job->s_fence->parent = NULL;
470 atomic_dec(&sched->hw_rq_count);
473 * remove job from pending_list.
474 * Locking here is for concurrent resume timeout
476 spin_lock(&sched->job_list_lock);
477 list_del_init(&s_job->list);
478 spin_unlock(&sched->job_list_lock);
481 * Wait for job's HW fence callback to finish using s_job
482 * before releasing it.
484 * Job is still alive so fence refcount at least 1
486 dma_fence_wait(&s_job->s_fence->finished, false);
489 * We must keep bad job alive for later use during
490 * recovery by some of the drivers but leave a hint
491 * that the guilty job must be released.
494 sched->ops->free_job(s_job);
496 sched->free_guilty = true;
501 * Stop pending timer in flight as we rearm it in drm_sched_start. This
502 * avoids the pending timeout work in progress to fire right away after
503 * this TDR finished and before the newly restarted jobs had a
504 * chance to complete.
506 cancel_delayed_work(&sched->work_tdr);
509 EXPORT_SYMBOL(drm_sched_stop);
512 * drm_sched_start - recover jobs after a reset
514 * @sched: scheduler instance
515 * @full_recovery: proceed with complete sched restart
518 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
520 struct drm_sched_job *s_job, *tmp;
524 * Locking the list is not required here as the sched thread is parked
525 * so no new jobs are being inserted or removed. Also concurrent
526 * GPU recovers can't run in parallel.
528 list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
529 struct dma_fence *fence = s_job->s_fence->parent;
531 atomic_inc(&sched->hw_rq_count);
537 r = dma_fence_add_callback(fence, &s_job->cb,
538 drm_sched_job_done_cb);
540 drm_sched_job_done(s_job, fence->error);
542 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
545 drm_sched_job_done(s_job, -ECANCELED);
549 spin_lock(&sched->job_list_lock);
550 drm_sched_start_timeout(sched);
551 spin_unlock(&sched->job_list_lock);
554 kthread_unpark(sched->thread);
556 EXPORT_SYMBOL(drm_sched_start);
559 * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
561 * @sched: scheduler instance
563 * Re-submitting jobs was a concept AMD came up as cheap way to implement
564 * recovery after a job timeout.
566 * This turned out to be not working very well. First of all there are many
567 * problem with the dma_fence implementation and requirements. Either the
568 * implementation is risking deadlocks with core memory management or violating
569 * documented implementation details of the dma_fence object.
571 * Drivers can still save and restore their state for recovery operations, but
572 * we shouldn't make this a general scheduler feature around the dma_fence
575 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
577 struct drm_sched_job *s_job, *tmp;
578 uint64_t guilty_context;
579 bool found_guilty = false;
580 struct dma_fence *fence;
582 list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
583 struct drm_sched_fence *s_fence = s_job->s_fence;
585 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
587 guilty_context = s_job->s_fence->scheduled.context;
590 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
591 dma_fence_set_error(&s_fence->finished, -ECANCELED);
593 fence = sched->ops->run_job(s_job);
595 if (IS_ERR_OR_NULL(fence)) {
597 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
599 s_job->s_fence->parent = NULL;
602 s_job->s_fence->parent = dma_fence_get(fence);
604 /* Drop for orignal kref_init */
605 dma_fence_put(fence);
609 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
612 * drm_sched_job_init - init a scheduler job
613 * @job: scheduler job to init
614 * @entity: scheduler entity to use
615 * @owner: job owner for debugging
617 * Refer to drm_sched_entity_push_job() documentation
618 * for locking considerations.
620 * Drivers must make sure drm_sched_job_cleanup() if this function returns
621 * successfully, even when @job is aborted before drm_sched_job_arm() is called.
623 * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
624 * has died, which can mean that there's no valid runqueue for a @entity.
625 * This function returns -ENOENT in this case (which probably should be -EIO as
626 * a more meanigful return value).
628 * Returns 0 for success, negative error code otherwise.
630 int drm_sched_job_init(struct drm_sched_job *job,
631 struct drm_sched_entity *entity,
637 job->entity = entity;
638 job->s_fence = drm_sched_fence_alloc(entity, owner);
642 INIT_LIST_HEAD(&job->list);
644 xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
648 EXPORT_SYMBOL(drm_sched_job_init);
651 * drm_sched_job_arm - arm a scheduler job for execution
652 * @job: scheduler job to arm
654 * This arms a scheduler job for execution. Specifically it initializes the
655 * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
656 * or other places that need to track the completion of this job.
658 * Refer to drm_sched_entity_push_job() documentation for locking
661 * This can only be called if drm_sched_job_init() succeeded.
663 void drm_sched_job_arm(struct drm_sched_job *job)
665 struct drm_gpu_scheduler *sched;
666 struct drm_sched_entity *entity = job->entity;
669 drm_sched_entity_select_rq(entity);
670 sched = entity->rq->sched;
673 job->s_priority = entity->rq - sched->sched_rq;
674 job->id = atomic64_inc_return(&sched->job_id_count);
676 drm_sched_fence_init(job->s_fence, job->entity);
678 EXPORT_SYMBOL(drm_sched_job_arm);
681 * drm_sched_job_add_dependency - adds the fence as a job dependency
682 * @job: scheduler job to add the dependencies to
683 * @fence: the dma_fence to add to the list of dependencies.
685 * Note that @fence is consumed in both the success and error cases.
688 * 0 on success, or an error on failing to expand the array.
690 int drm_sched_job_add_dependency(struct drm_sched_job *job,
691 struct dma_fence *fence)
693 struct dma_fence *entry;
701 /* Deduplicate if we already depend on a fence from the same context.
702 * This lets the size of the array of deps scale with the number of
703 * engines involved, rather than the number of BOs.
705 xa_for_each(&job->dependencies, index, entry) {
706 if (entry->context != fence->context)
709 if (dma_fence_is_later(fence, entry)) {
710 dma_fence_put(entry);
711 xa_store(&job->dependencies, index, fence, GFP_KERNEL);
713 dma_fence_put(fence);
718 ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
720 dma_fence_put(fence);
724 EXPORT_SYMBOL(drm_sched_job_add_dependency);
727 * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
728 * @job: scheduler job to add the dependencies to
729 * @file: drm file private pointer
730 * @handle: syncobj handle to lookup
731 * @point: timeline point
733 * This adds the fence matching the given syncobj to @job.
736 * 0 on success, or an error on failing to expand the array.
738 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
739 struct drm_file *file,
743 struct dma_fence *fence;
746 ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
750 return drm_sched_job_add_dependency(job, fence);
752 EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
755 * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
756 * @job: scheduler job to add the dependencies to
757 * @resv: the dma_resv object to get the fences from
758 * @usage: the dma_resv_usage to use to filter the fences
760 * This adds all fences matching the given usage from @resv to @job.
761 * Must be called with the @resv lock held.
764 * 0 on success, or an error on failing to expand the array.
766 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
767 struct dma_resv *resv,
768 enum dma_resv_usage usage)
770 struct dma_resv_iter cursor;
771 struct dma_fence *fence;
774 dma_resv_assert_held(resv);
776 dma_resv_for_each_fence(&cursor, resv, usage, fence) {
777 /* Make sure to grab an additional ref on the added fence */
778 dma_fence_get(fence);
779 ret = drm_sched_job_add_dependency(job, fence);
781 dma_fence_put(fence);
787 EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
790 * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
792 * @job: scheduler job to add the dependencies to
793 * @obj: the gem object to add new dependencies from.
794 * @write: whether the job might write the object (so we need to depend on
795 * shared fences in the reservation object).
797 * This should be called after drm_gem_lock_reservations() on your array of
798 * GEM objects used in the job but before updating the reservations with your
802 * 0 on success, or an error on failing to expand the array.
804 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
805 struct drm_gem_object *obj,
808 return drm_sched_job_add_resv_dependencies(job, obj->resv,
809 dma_resv_usage_rw(write));
811 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
814 * drm_sched_job_cleanup - clean up scheduler job resources
815 * @job: scheduler job to clean up
817 * Cleans up the resources allocated with drm_sched_job_init().
819 * Drivers should call this from their error unwind code if @job is aborted
820 * before drm_sched_job_arm() is called.
822 * After that point of no return @job is committed to be executed by the
823 * scheduler, and this function should be called from the
824 * &drm_sched_backend_ops.free_job callback.
826 void drm_sched_job_cleanup(struct drm_sched_job *job)
828 struct dma_fence *fence;
831 if (kref_read(&job->s_fence->finished.refcount)) {
832 /* drm_sched_job_arm() has been called */
833 dma_fence_put(&job->s_fence->finished);
835 /* aborted job before committing to run it */
836 drm_sched_fence_free(job->s_fence);
841 xa_for_each(&job->dependencies, index, fence) {
842 dma_fence_put(fence);
844 xa_destroy(&job->dependencies);
847 EXPORT_SYMBOL(drm_sched_job_cleanup);
850 * drm_sched_ready - is the scheduler ready
852 * @sched: scheduler instance
854 * Return true if we can push more jobs to the hw, otherwise false.
856 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
858 return atomic_read(&sched->hw_rq_count) <
859 sched->hw_submission_limit;
863 * drm_sched_wakeup - Wake up the scheduler when it is ready
865 * @sched: scheduler instance
868 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
870 if (drm_sched_ready(sched))
871 wake_up_interruptible(&sched->wake_up_worker);
875 * drm_sched_select_entity - Select next entity to process
877 * @sched: scheduler instance
879 * Returns the entity to process or NULL if none are found.
881 static struct drm_sched_entity *
882 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
884 struct drm_sched_entity *entity;
887 if (!drm_sched_ready(sched))
890 /* Kernel run queue has higher priority than normal run queue*/
891 for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
892 entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
893 drm_sched_rq_select_entity_fifo(&sched->sched_rq[i]) :
894 drm_sched_rq_select_entity_rr(&sched->sched_rq[i]);
903 * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
905 * @sched: scheduler instance
907 * Returns the next finished job from the pending list (if there is one)
908 * ready for it to be destroyed.
910 static struct drm_sched_job *
911 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
913 struct drm_sched_job *job, *next;
915 spin_lock(&sched->job_list_lock);
917 job = list_first_entry_or_null(&sched->pending_list,
918 struct drm_sched_job, list);
920 if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
921 /* remove job from pending_list */
922 list_del_init(&job->list);
924 /* cancel this job's TO timer */
925 cancel_delayed_work(&sched->work_tdr);
926 /* make the scheduled timestamp more accurate */
927 next = list_first_entry_or_null(&sched->pending_list,
928 typeof(*next), list);
931 next->s_fence->scheduled.timestamp =
932 job->s_fence->finished.timestamp;
933 /* start TO timer for next job */
934 drm_sched_start_timeout(sched);
940 spin_unlock(&sched->job_list_lock);
943 job->entity->elapsed_ns += ktime_to_ns(
944 ktime_sub(job->s_fence->finished.timestamp,
945 job->s_fence->scheduled.timestamp));
952 * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
953 * @sched_list: list of drm_gpu_schedulers
954 * @num_sched_list: number of drm_gpu_schedulers in the sched_list
956 * Returns pointer of the sched with the least load or NULL if none of the
957 * drm_gpu_schedulers are ready
959 struct drm_gpu_scheduler *
960 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
961 unsigned int num_sched_list)
963 struct drm_gpu_scheduler *sched, *picked_sched = NULL;
965 unsigned int min_score = UINT_MAX, num_score;
967 for (i = 0; i < num_sched_list; ++i) {
968 sched = sched_list[i];
971 DRM_WARN("scheduler %s is not ready, skipping",
976 num_score = atomic_read(sched->score);
977 if (num_score < min_score) {
978 min_score = num_score;
979 picked_sched = sched;
985 EXPORT_SYMBOL(drm_sched_pick_best);
988 * drm_sched_blocked - check if the scheduler is blocked
990 * @sched: scheduler instance
992 * Returns true if blocked, otherwise false.
994 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
996 if (kthread_should_park()) {
1005 * drm_sched_main - main scheduler thread
1007 * @param: scheduler instance
1011 static int drm_sched_main(void *param)
1013 struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
1016 sched_set_fifo_low(current);
1018 while (!kthread_should_stop()) {
1019 struct drm_sched_entity *entity = NULL;
1020 struct drm_sched_fence *s_fence;
1021 struct drm_sched_job *sched_job;
1022 struct dma_fence *fence;
1023 struct drm_sched_job *cleanup_job = NULL;
1025 wait_event_interruptible(sched->wake_up_worker,
1026 (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
1027 (!drm_sched_blocked(sched) &&
1028 (entity = drm_sched_select_entity(sched))) ||
1029 kthread_should_stop());
1032 sched->ops->free_job(cleanup_job);
1037 sched_job = drm_sched_entity_pop_job(entity);
1040 complete_all(&entity->entity_idle);
1044 s_fence = sched_job->s_fence;
1046 atomic_inc(&sched->hw_rq_count);
1047 drm_sched_job_begin(sched_job);
1049 trace_drm_run_job(sched_job, entity);
1050 fence = sched->ops->run_job(sched_job);
1051 complete_all(&entity->entity_idle);
1052 drm_sched_fence_scheduled(s_fence);
1054 if (!IS_ERR_OR_NULL(fence)) {
1055 drm_sched_fence_set_parent(s_fence, fence);
1056 /* Drop for original kref_init of the fence */
1057 dma_fence_put(fence);
1059 r = dma_fence_add_callback(fence, &sched_job->cb,
1060 drm_sched_job_done_cb);
1062 drm_sched_job_done(sched_job, fence->error);
1064 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
1067 drm_sched_job_done(sched_job, IS_ERR(fence) ?
1068 PTR_ERR(fence) : 0);
1071 wake_up(&sched->job_scheduled);
1077 * drm_sched_init - Init a gpu scheduler instance
1079 * @sched: scheduler instance
1080 * @ops: backend operations for this scheduler
1081 * @hw_submission: number of hw submissions that can be in flight
1082 * @hang_limit: number of times to allow a job to hang before dropping it
1083 * @timeout: timeout value in jiffies for the scheduler
1084 * @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
1086 * @score: optional score atomic shared with other schedulers
1087 * @name: name used for debugging
1088 * @dev: target &struct device
1090 * Return 0 on success, otherwise error code.
1092 int drm_sched_init(struct drm_gpu_scheduler *sched,
1093 const struct drm_sched_backend_ops *ops,
1094 unsigned hw_submission, unsigned hang_limit,
1095 long timeout, struct workqueue_struct *timeout_wq,
1096 atomic_t *score, const char *name, struct device *dev)
1100 sched->hw_submission_limit = hw_submission;
1102 sched->timeout = timeout;
1103 sched->timeout_wq = timeout_wq ? : system_wq;
1104 sched->hang_limit = hang_limit;
1105 sched->score = score ? score : &sched->_score;
1107 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
1108 drm_sched_rq_init(sched, &sched->sched_rq[i]);
1110 init_waitqueue_head(&sched->wake_up_worker);
1111 init_waitqueue_head(&sched->job_scheduled);
1112 INIT_LIST_HEAD(&sched->pending_list);
1113 spin_lock_init(&sched->job_list_lock);
1114 atomic_set(&sched->hw_rq_count, 0);
1115 INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1116 atomic_set(&sched->_score, 0);
1117 atomic64_set(&sched->job_id_count, 0);
1119 /* Each scheduler will run on a seperate kernel thread */
1120 sched->thread = kthread_run(drm_sched_main, sched, sched->name);
1121 if (IS_ERR(sched->thread)) {
1122 ret = PTR_ERR(sched->thread);
1123 sched->thread = NULL;
1124 DRM_DEV_ERROR(sched->dev, "Failed to create scheduler for %s.\n", name);
1128 sched->ready = true;
1131 EXPORT_SYMBOL(drm_sched_init);
1134 * drm_sched_fini - Destroy a gpu scheduler
1136 * @sched: scheduler instance
1138 * Tears down and cleans up the scheduler.
1140 void drm_sched_fini(struct drm_gpu_scheduler *sched)
1142 struct drm_sched_entity *s_entity;
1146 kthread_stop(sched->thread);
1148 for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
1149 struct drm_sched_rq *rq = &sched->sched_rq[i];
1154 spin_lock(&rq->lock);
1155 list_for_each_entry(s_entity, &rq->entities, list)
1157 * Prevents reinsertion and marks job_queue as idle,
1158 * it will removed from rq in drm_sched_entity_fini
1161 s_entity->stopped = true;
1162 spin_unlock(&rq->lock);
1166 /* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1167 wake_up_all(&sched->job_scheduled);
1169 /* Confirm no work left behind accessing device structures */
1170 cancel_delayed_work_sync(&sched->work_tdr);
1172 sched->ready = false;
1174 EXPORT_SYMBOL(drm_sched_fini);
1177 * drm_sched_increase_karma - Update sched_entity guilty flag
1179 * @bad: The job guilty of time out
1181 * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1182 * limit of the scheduler then the respective sched entity is marked guilty and
1183 * jobs from it will not be scheduled further
1185 void drm_sched_increase_karma(struct drm_sched_job *bad)
1188 struct drm_sched_entity *tmp;
1189 struct drm_sched_entity *entity;
1190 struct drm_gpu_scheduler *sched = bad->sched;
1192 /* don't change @bad's karma if it's from KERNEL RQ,
1193 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1194 * corrupt but keep in mind that kernel jobs always considered good.
1196 if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1197 atomic_inc(&bad->karma);
1199 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
1201 struct drm_sched_rq *rq = &sched->sched_rq[i];
1203 spin_lock(&rq->lock);
1204 list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1205 if (bad->s_fence->scheduled.context ==
1206 entity->fence_context) {
1208 atomic_set(entity->guilty, 1);
1212 spin_unlock(&rq->lock);
1213 if (&entity->list != &rq->entities)
1218 EXPORT_SYMBOL(drm_sched_increase_karma);