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(1);
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) {
254 /* Fence is from the same scheduler */
255 if (test_bit(AMD_SCHED_FENCE_SCHEDULED_BIT, &fence->flags)) {
256 /* Ignore it when it is already scheduled */
257 fence_put(entity->dependency);
261 /* Wait for fence to be scheduled */
262 entity->cb.func = amd_sched_entity_clear_dep;
263 list_add_tail(&entity->cb.node, &s_fence->scheduled_cb);
267 if (!fence_add_callback(entity->dependency, &entity->cb,
268 amd_sched_entity_wakeup))
271 fence_put(entity->dependency);
275 static struct amd_sched_job *
276 amd_sched_entity_pop_job(struct amd_sched_entity *entity)
278 struct amd_gpu_scheduler *sched = entity->sched;
279 struct amd_sched_job *sched_job;
281 if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job)))
284 while ((entity->dependency = sched->ops->dependency(sched_job)))
285 if (amd_sched_entity_add_dependency_cb(entity))
292 * Helper to submit a job to the job queue
294 * @sched_job The pointer to job required to submit
296 * Returns true if we could submit the job.
298 static bool amd_sched_entity_in(struct amd_sched_job *sched_job)
300 struct amd_gpu_scheduler *sched = sched_job->sched;
301 struct amd_sched_entity *entity = sched_job->s_entity;
302 bool added, first = false;
304 spin_lock(&entity->queue_lock);
305 added = kfifo_in(&entity->job_queue, &sched_job,
306 sizeof(sched_job)) == sizeof(sched_job);
308 if (added && kfifo_len(&entity->job_queue) == sizeof(sched_job))
311 spin_unlock(&entity->queue_lock);
313 /* first job wakes up scheduler */
315 /* Add the entity to the run queue */
316 amd_sched_rq_add_entity(entity->rq, entity);
317 amd_sched_wakeup(sched);
322 static void amd_sched_free_job(struct fence *f, struct fence_cb *cb) {
323 struct amd_sched_job *job = container_of(cb, struct amd_sched_job, cb_free_job);
324 schedule_work(&job->work_free_job);
327 /* job_finish is called after hw fence signaled, and
328 * the job had already been deleted from ring_mirror_list
330 void amd_sched_job_finish(struct amd_sched_job *s_job)
332 struct amd_sched_job *next;
333 struct amd_gpu_scheduler *sched = s_job->sched;
335 if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
336 if (cancel_delayed_work(&s_job->work_tdr))
337 amd_sched_job_put(s_job);
339 /* queue TDR for next job */
340 next = list_first_entry_or_null(&sched->ring_mirror_list,
341 struct amd_sched_job, node);
344 INIT_DELAYED_WORK(&next->work_tdr, s_job->timeout_callback);
345 amd_sched_job_get(next);
346 schedule_delayed_work(&next->work_tdr, sched->timeout);
351 void amd_sched_job_begin(struct amd_sched_job *s_job)
353 struct amd_gpu_scheduler *sched = s_job->sched;
355 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
356 list_first_entry_or_null(&sched->ring_mirror_list, struct amd_sched_job, node) == s_job)
358 INIT_DELAYED_WORK(&s_job->work_tdr, s_job->timeout_callback);
359 amd_sched_job_get(s_job);
360 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
365 * Submit a job to the job queue
367 * @sched_job The pointer to job required to submit
369 * Returns 0 for success, negative error code otherwise.
371 void amd_sched_entity_push_job(struct amd_sched_job *sched_job)
373 struct amd_sched_entity *entity = sched_job->s_entity;
375 sched_job->use_sched = 1;
376 fence_add_callback(&sched_job->s_fence->base,
377 &sched_job->cb_free_job, amd_sched_free_job);
378 trace_amd_sched_job(sched_job);
379 wait_event(entity->sched->job_scheduled,
380 amd_sched_entity_in(sched_job));
383 /* init a sched_job with basic field */
384 int amd_sched_job_init(struct amd_sched_job *job,
385 struct amd_gpu_scheduler *sched,
386 struct amd_sched_entity *entity,
387 void (*timeout_cb)(struct work_struct *work),
388 void (*free_cb)(struct kref *refcount),
389 void *owner, struct fence **fence)
391 INIT_LIST_HEAD(&job->node);
392 kref_init(&job->refcount);
394 job->s_entity = entity;
395 job->s_fence = amd_sched_fence_create(entity, owner);
399 job->s_fence->s_job = job;
400 job->timeout_callback = timeout_cb;
401 job->free_callback = free_cb;
404 *fence = &job->s_fence->base;
409 * Return ture if we can push more jobs to the hw.
411 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
413 return atomic_read(&sched->hw_rq_count) <
414 sched->hw_submission_limit;
418 * Wake up the scheduler when it is ready
420 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched)
422 if (amd_sched_ready(sched))
423 wake_up_interruptible(&sched->wake_up_worker);
427 * Select next entity to process
429 static struct amd_sched_entity *
430 amd_sched_select_entity(struct amd_gpu_scheduler *sched)
432 struct amd_sched_entity *entity;
435 if (!amd_sched_ready(sched))
438 /* Kernel run queue has higher priority than normal run queue*/
439 for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++) {
440 entity = amd_sched_rq_select_entity(&sched->sched_rq[i]);
448 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
450 struct amd_sched_fence *s_fence =
451 container_of(cb, struct amd_sched_fence, cb);
452 struct amd_gpu_scheduler *sched = s_fence->sched;
455 atomic_dec(&sched->hw_rq_count);
457 /* remove job from ring_mirror_list */
458 spin_lock_irqsave(&sched->job_list_lock, flags);
459 list_del_init(&s_fence->s_job->node);
460 sched->ops->finish_job(s_fence->s_job);
461 spin_unlock_irqrestore(&sched->job_list_lock, flags);
463 amd_sched_fence_signal(s_fence);
465 trace_amd_sched_process_job(s_fence);
466 fence_put(&s_fence->base);
467 wake_up_interruptible(&sched->wake_up_worker);
470 static int amd_sched_main(void *param)
472 struct sched_param sparam = {.sched_priority = 1};
473 struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
476 sched_setscheduler(current, SCHED_FIFO, &sparam);
478 while (!kthread_should_stop()) {
479 struct amd_sched_entity *entity;
480 struct amd_sched_fence *s_fence;
481 struct amd_sched_job *sched_job;
484 wait_event_interruptible(sched->wake_up_worker,
485 (entity = amd_sched_select_entity(sched)) ||
486 kthread_should_stop());
491 sched_job = amd_sched_entity_pop_job(entity);
495 s_fence = sched_job->s_fence;
497 atomic_inc(&sched->hw_rq_count);
498 amd_sched_job_pre_schedule(sched, sched_job);
499 fence = sched->ops->run_job(sched_job);
500 amd_sched_fence_scheduled(s_fence);
502 r = fence_add_callback(fence, &s_fence->cb,
503 amd_sched_process_job);
505 amd_sched_process_job(fence, &s_fence->cb);
507 DRM_ERROR("fence add callback failed (%d)\n", r);
510 DRM_ERROR("Failed to run job!\n");
511 amd_sched_process_job(NULL, &s_fence->cb);
514 count = kfifo_out(&entity->job_queue, &sched_job,
516 WARN_ON(count != sizeof(sched_job));
517 wake_up(&sched->job_scheduled);
523 * Init a gpu scheduler instance
525 * @sched The pointer to the scheduler
526 * @ops The backend operations for this scheduler.
527 * @hw_submissions Number of hw submissions to do.
528 * @name Name used for debugging
530 * Return 0 on success, otherwise error code.
532 int amd_sched_init(struct amd_gpu_scheduler *sched,
533 const struct amd_sched_backend_ops *ops,
534 unsigned hw_submission, long timeout, const char *name)
538 sched->hw_submission_limit = hw_submission;
540 sched->timeout = timeout;
541 for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++)
542 amd_sched_rq_init(&sched->sched_rq[i]);
544 init_waitqueue_head(&sched->wake_up_worker);
545 init_waitqueue_head(&sched->job_scheduled);
546 INIT_LIST_HEAD(&sched->ring_mirror_list);
547 spin_lock_init(&sched->job_list_lock);
548 atomic_set(&sched->hw_rq_count, 0);
549 if (atomic_inc_return(&sched_fence_slab_ref) == 1) {
550 sched_fence_slab = kmem_cache_create(
551 "amd_sched_fence", sizeof(struct amd_sched_fence), 0,
552 SLAB_HWCACHE_ALIGN, NULL);
553 if (!sched_fence_slab)
557 /* Each scheduler will run on a seperate kernel thread */
558 sched->thread = kthread_run(amd_sched_main, sched, sched->name);
559 if (IS_ERR(sched->thread)) {
560 DRM_ERROR("Failed to create scheduler for %s.\n", name);
561 return PTR_ERR(sched->thread);
568 * Destroy a gpu scheduler
570 * @sched The pointer to the scheduler
572 void amd_sched_fini(struct amd_gpu_scheduler *sched)
575 kthread_stop(sched->thread);
576 if (atomic_dec_and_test(&sched_fence_slab_ref))
577 kmem_cache_destroy(sched_fence_slab);