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.
24 #include <linux/kthread.h>
25 #include <linux/wait.h>
26 #include <linux/sched.h>
28 #include "gpu_scheduler.h"
30 #define CREATE_TRACE_POINTS
31 #include "gpu_sched_trace.h"
33 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity);
34 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched);
36 struct kmem_cache *sched_fence_slab;
37 atomic_t sched_fence_slab_ref = ATOMIC_INIT(0);
39 /* Initialize a given run queue struct */
40 static void amd_sched_rq_init(struct amd_sched_rq *rq)
42 spin_lock_init(&rq->lock);
43 INIT_LIST_HEAD(&rq->entities);
44 rq->current_entity = NULL;
47 static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
48 struct amd_sched_entity *entity)
50 if (!list_empty(&entity->list))
53 list_add_tail(&entity->list, &rq->entities);
54 spin_unlock(&rq->lock);
57 static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
58 struct amd_sched_entity *entity)
60 if (list_empty(&entity->list))
63 list_del_init(&entity->list);
64 if (rq->current_entity == entity)
65 rq->current_entity = NULL;
66 spin_unlock(&rq->lock);
70 * Select an entity which could provide a job to run
72 * @rq The run queue to check.
74 * Try to find a ready entity, returns NULL if none found.
76 static struct amd_sched_entity *
77 amd_sched_rq_select_entity(struct amd_sched_rq *rq)
79 struct amd_sched_entity *entity;
83 entity = rq->current_entity;
85 list_for_each_entry_continue(entity, &rq->entities, list) {
86 if (amd_sched_entity_is_ready(entity)) {
87 rq->current_entity = entity;
88 spin_unlock(&rq->lock);
94 list_for_each_entry(entity, &rq->entities, list) {
96 if (amd_sched_entity_is_ready(entity)) {
97 rq->current_entity = entity;
98 spin_unlock(&rq->lock);
102 if (entity == rq->current_entity)
106 spin_unlock(&rq->lock);
112 * Init a context entity used by scheduler when submit to HW ring.
114 * @sched The pointer to the scheduler
115 * @entity The pointer to a valid amd_sched_entity
116 * @rq The run queue this entity belongs
117 * @kernel If this is an entity for the kernel
118 * @jobs The max number of jobs in the job queue
120 * return 0 if succeed. negative error code on failure
122 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
123 struct amd_sched_entity *entity,
124 struct amd_sched_rq *rq,
129 if (!(sched && entity && rq))
132 memset(entity, 0, sizeof(struct amd_sched_entity));
133 INIT_LIST_HEAD(&entity->list);
135 entity->sched = sched;
137 spin_lock_init(&entity->queue_lock);
138 r = kfifo_alloc(&entity->job_queue, jobs * sizeof(void *), GFP_KERNEL);
142 atomic_set(&entity->fence_seq, 0);
143 entity->fence_context = fence_context_alloc(2);
149 * Query if entity is initialized
151 * @sched Pointer to scheduler instance
152 * @entity The pointer to a valid scheduler entity
154 * return true if entity is initialized, false otherwise
156 static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
157 struct amd_sched_entity *entity)
159 return entity->sched == sched &&
164 * Check if entity is idle
166 * @entity The pointer to a valid scheduler entity
168 * Return true if entity don't has any unscheduled jobs.
170 static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
173 if (kfifo_is_empty(&entity->job_queue))
180 * Check if entity is ready
182 * @entity The pointer to a valid scheduler entity
184 * Return true if entity could provide a job.
186 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity)
188 if (kfifo_is_empty(&entity->job_queue))
191 if (ACCESS_ONCE(entity->dependency))
198 * Destroy a context entity
200 * @sched Pointer to scheduler instance
201 * @entity The pointer to a valid scheduler entity
203 * Cleanup and free the allocated resources.
205 void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
206 struct amd_sched_entity *entity)
208 struct amd_sched_rq *rq = entity->rq;
210 if (!amd_sched_entity_is_initialized(sched, entity))
214 * The client will not queue more IBs during this fini, consume existing
217 wait_event(sched->job_scheduled, amd_sched_entity_is_idle(entity));
219 amd_sched_rq_remove_entity(rq, entity);
220 kfifo_free(&entity->job_queue);
223 static void amd_sched_entity_wakeup(struct fence *f, struct fence_cb *cb)
225 struct amd_sched_entity *entity =
226 container_of(cb, struct amd_sched_entity, cb);
227 entity->dependency = NULL;
229 amd_sched_wakeup(entity->sched);
232 static void amd_sched_entity_clear_dep(struct fence *f, struct fence_cb *cb)
234 struct amd_sched_entity *entity =
235 container_of(cb, struct amd_sched_entity, cb);
236 entity->dependency = NULL;
240 static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
242 struct amd_gpu_scheduler *sched = entity->sched;
243 struct fence * fence = entity->dependency;
244 struct amd_sched_fence *s_fence;
246 if (fence->context == entity->fence_context) {
247 /* We can ignore fences from ourself */
248 fence_put(entity->dependency);
252 s_fence = to_amd_sched_fence(fence);
253 if (s_fence && s_fence->sched == sched) {
256 * Fence is from the same scheduler, only need to wait for
259 fence = fence_get(&s_fence->scheduled);
260 fence_put(entity->dependency);
261 entity->dependency = fence;
262 if (!fence_add_callback(fence, &entity->cb,
263 amd_sched_entity_clear_dep))
266 /* Ignore it when it is already scheduled */
271 if (!fence_add_callback(entity->dependency, &entity->cb,
272 amd_sched_entity_wakeup))
275 fence_put(entity->dependency);
279 static struct amd_sched_job *
280 amd_sched_entity_pop_job(struct amd_sched_entity *entity)
282 struct amd_gpu_scheduler *sched = entity->sched;
283 struct amd_sched_job *sched_job;
285 if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job)))
288 while ((entity->dependency = sched->ops->dependency(sched_job)))
289 if (amd_sched_entity_add_dependency_cb(entity))
296 * Helper to submit a job to the job queue
298 * @sched_job The pointer to job required to submit
300 * Returns true if we could submit the job.
302 static bool amd_sched_entity_in(struct amd_sched_job *sched_job)
304 struct amd_gpu_scheduler *sched = sched_job->sched;
305 struct amd_sched_entity *entity = sched_job->s_entity;
306 bool added, first = false;
308 spin_lock(&entity->queue_lock);
309 added = kfifo_in(&entity->job_queue, &sched_job,
310 sizeof(sched_job)) == sizeof(sched_job);
312 if (added && kfifo_len(&entity->job_queue) == sizeof(sched_job))
315 spin_unlock(&entity->queue_lock);
317 /* first job wakes up scheduler */
319 /* Add the entity to the run queue */
320 amd_sched_rq_add_entity(entity->rq, entity);
321 amd_sched_wakeup(sched);
326 /* job_finish is called after hw fence signaled, and
327 * the job had already been deleted from ring_mirror_list
329 static void amd_sched_job_finish(struct work_struct *work)
331 struct amd_sched_job *s_job = container_of(work, struct amd_sched_job,
333 struct amd_gpu_scheduler *sched = s_job->sched;
336 /* remove job from ring_mirror_list */
337 spin_lock_irqsave(&sched->job_list_lock, flags);
338 list_del_init(&s_job->node);
339 if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
340 struct amd_sched_job *next;
342 spin_unlock_irqrestore(&sched->job_list_lock, flags);
343 cancel_delayed_work_sync(&s_job->work_tdr);
344 spin_lock_irqsave(&sched->job_list_lock, flags);
346 /* queue TDR for next job */
347 next = list_first_entry_or_null(&sched->ring_mirror_list,
348 struct amd_sched_job, node);
351 schedule_delayed_work(&next->work_tdr, sched->timeout);
353 spin_unlock_irqrestore(&sched->job_list_lock, flags);
354 sched->ops->free_job(s_job);
357 static void amd_sched_job_finish_cb(struct fence *f, struct fence_cb *cb)
359 struct amd_sched_job *job = container_of(cb, struct amd_sched_job,
361 schedule_work(&job->finish_work);
364 static void amd_sched_job_begin(struct amd_sched_job *s_job)
366 struct amd_gpu_scheduler *sched = s_job->sched;
369 spin_lock_irqsave(&sched->job_list_lock, flags);
370 list_add_tail(&s_job->node, &sched->ring_mirror_list);
371 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
372 list_first_entry_or_null(&sched->ring_mirror_list,
373 struct amd_sched_job, node) == s_job)
374 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
375 spin_unlock_irqrestore(&sched->job_list_lock, flags);
378 static void amd_sched_job_timedout(struct work_struct *work)
380 struct amd_sched_job *job = container_of(work, struct amd_sched_job,
383 job->sched->ops->timedout_job(job);
387 * Submit a job to the job queue
389 * @sched_job The pointer to job required to submit
391 * Returns 0 for success, negative error code otherwise.
393 void amd_sched_entity_push_job(struct amd_sched_job *sched_job)
395 struct amd_sched_entity *entity = sched_job->s_entity;
397 trace_amd_sched_job(sched_job);
398 fence_add_callback(&sched_job->s_fence->finished, &sched_job->finish_cb,
399 amd_sched_job_finish_cb);
400 wait_event(entity->sched->job_scheduled,
401 amd_sched_entity_in(sched_job));
404 /* init a sched_job with basic field */
405 int amd_sched_job_init(struct amd_sched_job *job,
406 struct amd_gpu_scheduler *sched,
407 struct amd_sched_entity *entity,
408 void *owner, struct fence **fence)
411 job->s_entity = entity;
412 job->s_fence = amd_sched_fence_create(entity, owner);
416 INIT_WORK(&job->finish_work, amd_sched_job_finish);
417 INIT_LIST_HEAD(&job->node);
418 INIT_DELAYED_WORK(&job->work_tdr, amd_sched_job_timedout);
421 *fence = &job->s_fence->finished;
426 * Return ture if we can push more jobs to the hw.
428 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
430 return atomic_read(&sched->hw_rq_count) <
431 sched->hw_submission_limit;
435 * Wake up the scheduler when it is ready
437 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched)
439 if (amd_sched_ready(sched))
440 wake_up_interruptible(&sched->wake_up_worker);
444 * Select next entity to process
446 static struct amd_sched_entity *
447 amd_sched_select_entity(struct amd_gpu_scheduler *sched)
449 struct amd_sched_entity *entity;
452 if (!amd_sched_ready(sched))
455 /* Kernel run queue has higher priority than normal run queue*/
456 for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++) {
457 entity = amd_sched_rq_select_entity(&sched->sched_rq[i]);
465 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
467 struct amd_sched_fence *s_fence =
468 container_of(cb, struct amd_sched_fence, cb);
469 struct amd_gpu_scheduler *sched = s_fence->sched;
471 atomic_dec(&sched->hw_rq_count);
472 amd_sched_fence_finished(s_fence);
474 trace_amd_sched_process_job(s_fence);
475 fence_put(&s_fence->finished);
476 wake_up_interruptible(&sched->wake_up_worker);
479 static int amd_sched_main(void *param)
481 struct sched_param sparam = {.sched_priority = 1};
482 struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
485 sched_setscheduler(current, SCHED_FIFO, &sparam);
487 while (!kthread_should_stop()) {
488 struct amd_sched_entity *entity;
489 struct amd_sched_fence *s_fence;
490 struct amd_sched_job *sched_job;
493 wait_event_interruptible(sched->wake_up_worker,
494 (entity = amd_sched_select_entity(sched)) ||
495 kthread_should_stop());
500 sched_job = amd_sched_entity_pop_job(entity);
504 s_fence = sched_job->s_fence;
506 atomic_inc(&sched->hw_rq_count);
507 amd_sched_job_begin(sched_job);
509 fence = sched->ops->run_job(sched_job);
510 amd_sched_fence_scheduled(s_fence);
512 r = fence_add_callback(fence, &s_fence->cb,
513 amd_sched_process_job);
515 amd_sched_process_job(fence, &s_fence->cb);
517 DRM_ERROR("fence add callback failed (%d)\n",
521 DRM_ERROR("Failed to run job!\n");
522 amd_sched_process_job(NULL, &s_fence->cb);
525 count = kfifo_out(&entity->job_queue, &sched_job,
527 WARN_ON(count != sizeof(sched_job));
528 wake_up(&sched->job_scheduled);
534 * Init a gpu scheduler instance
536 * @sched The pointer to the scheduler
537 * @ops The backend operations for this scheduler.
538 * @hw_submissions Number of hw submissions to do.
539 * @name Name used for debugging
541 * Return 0 on success, otherwise error code.
543 int amd_sched_init(struct amd_gpu_scheduler *sched,
544 const struct amd_sched_backend_ops *ops,
545 unsigned hw_submission, long timeout, const char *name)
549 sched->hw_submission_limit = hw_submission;
551 sched->timeout = timeout;
552 for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++)
553 amd_sched_rq_init(&sched->sched_rq[i]);
555 init_waitqueue_head(&sched->wake_up_worker);
556 init_waitqueue_head(&sched->job_scheduled);
557 INIT_LIST_HEAD(&sched->ring_mirror_list);
558 spin_lock_init(&sched->job_list_lock);
559 atomic_set(&sched->hw_rq_count, 0);
560 if (atomic_inc_return(&sched_fence_slab_ref) == 1) {
561 sched_fence_slab = kmem_cache_create(
562 "amd_sched_fence", sizeof(struct amd_sched_fence), 0,
563 SLAB_HWCACHE_ALIGN, NULL);
564 if (!sched_fence_slab)
568 /* Each scheduler will run on a seperate kernel thread */
569 sched->thread = kthread_run(amd_sched_main, sched, sched->name);
570 if (IS_ERR(sched->thread)) {
571 DRM_ERROR("Failed to create scheduler for %s.\n", name);
572 return PTR_ERR(sched->thread);
579 * Destroy a gpu scheduler
581 * @sched The pointer to the scheduler
583 void amd_sched_fini(struct amd_gpu_scheduler *sched)
586 kthread_stop(sched->thread);
587 if (atomic_dec_and_test(&sched_fence_slab_ref))
588 kmem_cache_destroy(sched_fence_slab);