Loading...
Note: File does not exist in v6.9.4.
1/*
2 * Copyright 2015 Advanced Micro Devices, Inc.
3 *
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:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
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.
21 *
22 *
23 */
24#include <linux/kthread.h>
25#include <linux/wait.h>
26#include <linux/sched.h>
27#include <drm/drmP.h>
28#include "gpu_scheduler.h"
29
30#define CREATE_TRACE_POINTS
31#include "gpu_sched_trace.h"
32
33static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity);
34static void amd_sched_wakeup(struct amd_gpu_scheduler *sched);
35static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
36
37/* Initialize a given run queue struct */
38static void amd_sched_rq_init(struct amd_sched_rq *rq)
39{
40 spin_lock_init(&rq->lock);
41 INIT_LIST_HEAD(&rq->entities);
42 rq->current_entity = NULL;
43}
44
45static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
46 struct amd_sched_entity *entity)
47{
48 if (!list_empty(&entity->list))
49 return;
50 spin_lock(&rq->lock);
51 list_add_tail(&entity->list, &rq->entities);
52 spin_unlock(&rq->lock);
53}
54
55static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
56 struct amd_sched_entity *entity)
57{
58 if (list_empty(&entity->list))
59 return;
60 spin_lock(&rq->lock);
61 list_del_init(&entity->list);
62 if (rq->current_entity == entity)
63 rq->current_entity = NULL;
64 spin_unlock(&rq->lock);
65}
66
67/**
68 * Select an entity which could provide a job to run
69 *
70 * @rq The run queue to check.
71 *
72 * Try to find a ready entity, returns NULL if none found.
73 */
74static struct amd_sched_entity *
75amd_sched_rq_select_entity(struct amd_sched_rq *rq)
76{
77 struct amd_sched_entity *entity;
78
79 spin_lock(&rq->lock);
80
81 entity = rq->current_entity;
82 if (entity) {
83 list_for_each_entry_continue(entity, &rq->entities, list) {
84 if (amd_sched_entity_is_ready(entity)) {
85 rq->current_entity = entity;
86 spin_unlock(&rq->lock);
87 return entity;
88 }
89 }
90 }
91
92 list_for_each_entry(entity, &rq->entities, list) {
93
94 if (amd_sched_entity_is_ready(entity)) {
95 rq->current_entity = entity;
96 spin_unlock(&rq->lock);
97 return entity;
98 }
99
100 if (entity == rq->current_entity)
101 break;
102 }
103
104 spin_unlock(&rq->lock);
105
106 return NULL;
107}
108
109/**
110 * Init a context entity used by scheduler when submit to HW ring.
111 *
112 * @sched The pointer to the scheduler
113 * @entity The pointer to a valid amd_sched_entity
114 * @rq The run queue this entity belongs
115 * @kernel If this is an entity for the kernel
116 * @jobs The max number of jobs in the job queue
117 *
118 * return 0 if succeed. negative error code on failure
119*/
120int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
121 struct amd_sched_entity *entity,
122 struct amd_sched_rq *rq,
123 uint32_t jobs)
124{
125 int r;
126
127 if (!(sched && entity && rq))
128 return -EINVAL;
129
130 memset(entity, 0, sizeof(struct amd_sched_entity));
131 INIT_LIST_HEAD(&entity->list);
132 entity->rq = rq;
133 entity->sched = sched;
134
135 spin_lock_init(&entity->queue_lock);
136 r = kfifo_alloc(&entity->job_queue, jobs * sizeof(void *), GFP_KERNEL);
137 if (r)
138 return r;
139
140 atomic_set(&entity->fence_seq, 0);
141 entity->fence_context = dma_fence_context_alloc(2);
142
143 return 0;
144}
145
146/**
147 * Query if entity is initialized
148 *
149 * @sched Pointer to scheduler instance
150 * @entity The pointer to a valid scheduler entity
151 *
152 * return true if entity is initialized, false otherwise
153*/
154static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
155 struct amd_sched_entity *entity)
156{
157 return entity->sched == sched &&
158 entity->rq != NULL;
159}
160
161/**
162 * Check if entity is idle
163 *
164 * @entity The pointer to a valid scheduler entity
165 *
166 * Return true if entity don't has any unscheduled jobs.
167 */
168static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
169{
170 rmb();
171 if (kfifo_is_empty(&entity->job_queue))
172 return true;
173
174 return false;
175}
176
177/**
178 * Check if entity is ready
179 *
180 * @entity The pointer to a valid scheduler entity
181 *
182 * Return true if entity could provide a job.
183 */
184static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity)
185{
186 if (kfifo_is_empty(&entity->job_queue))
187 return false;
188
189 if (ACCESS_ONCE(entity->dependency))
190 return false;
191
192 return true;
193}
194
195/**
196 * Destroy a context entity
197 *
198 * @sched Pointer to scheduler instance
199 * @entity The pointer to a valid scheduler entity
200 *
201 * Cleanup and free the allocated resources.
202 */
203void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
204 struct amd_sched_entity *entity)
205{
206 struct amd_sched_rq *rq = entity->rq;
207
208 if (!amd_sched_entity_is_initialized(sched, entity))
209 return;
210
211 /**
212 * The client will not queue more IBs during this fini, consume existing
213 * queued IBs
214 */
215 wait_event(sched->job_scheduled, amd_sched_entity_is_idle(entity));
216
217 amd_sched_rq_remove_entity(rq, entity);
218 kfifo_free(&entity->job_queue);
219}
220
221static void amd_sched_entity_wakeup(struct dma_fence *f, struct dma_fence_cb *cb)
222{
223 struct amd_sched_entity *entity =
224 container_of(cb, struct amd_sched_entity, cb);
225 entity->dependency = NULL;
226 dma_fence_put(f);
227 amd_sched_wakeup(entity->sched);
228}
229
230static void amd_sched_entity_clear_dep(struct dma_fence *f, struct dma_fence_cb *cb)
231{
232 struct amd_sched_entity *entity =
233 container_of(cb, struct amd_sched_entity, cb);
234 entity->dependency = NULL;
235 dma_fence_put(f);
236}
237
238static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
239{
240 struct amd_gpu_scheduler *sched = entity->sched;
241 struct dma_fence * fence = entity->dependency;
242 struct amd_sched_fence *s_fence;
243
244 if (fence->context == entity->fence_context) {
245 /* We can ignore fences from ourself */
246 dma_fence_put(entity->dependency);
247 return false;
248 }
249
250 s_fence = to_amd_sched_fence(fence);
251 if (s_fence && s_fence->sched == sched) {
252
253 /*
254 * Fence is from the same scheduler, only need to wait for
255 * it to be scheduled
256 */
257 fence = dma_fence_get(&s_fence->scheduled);
258 dma_fence_put(entity->dependency);
259 entity->dependency = fence;
260 if (!dma_fence_add_callback(fence, &entity->cb,
261 amd_sched_entity_clear_dep))
262 return true;
263
264 /* Ignore it when it is already scheduled */
265 dma_fence_put(fence);
266 return false;
267 }
268
269 if (!dma_fence_add_callback(entity->dependency, &entity->cb,
270 amd_sched_entity_wakeup))
271 return true;
272
273 dma_fence_put(entity->dependency);
274 return false;
275}
276
277static struct amd_sched_job *
278amd_sched_entity_pop_job(struct amd_sched_entity *entity)
279{
280 struct amd_gpu_scheduler *sched = entity->sched;
281 struct amd_sched_job *sched_job;
282
283 if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job)))
284 return NULL;
285
286 while ((entity->dependency = sched->ops->dependency(sched_job)))
287 if (amd_sched_entity_add_dependency_cb(entity))
288 return NULL;
289
290 return sched_job;
291}
292
293/**
294 * Helper to submit a job to the job queue
295 *
296 * @sched_job The pointer to job required to submit
297 *
298 * Returns true if we could submit the job.
299 */
300static bool amd_sched_entity_in(struct amd_sched_job *sched_job)
301{
302 struct amd_gpu_scheduler *sched = sched_job->sched;
303 struct amd_sched_entity *entity = sched_job->s_entity;
304 bool added, first = false;
305
306 spin_lock(&entity->queue_lock);
307 added = kfifo_in(&entity->job_queue, &sched_job,
308 sizeof(sched_job)) == sizeof(sched_job);
309
310 if (added && kfifo_len(&entity->job_queue) == sizeof(sched_job))
311 first = true;
312
313 spin_unlock(&entity->queue_lock);
314
315 /* first job wakes up scheduler */
316 if (first) {
317 /* Add the entity to the run queue */
318 amd_sched_rq_add_entity(entity->rq, entity);
319 amd_sched_wakeup(sched);
320 }
321 return added;
322}
323
324/* job_finish is called after hw fence signaled, and
325 * the job had already been deleted from ring_mirror_list
326 */
327static void amd_sched_job_finish(struct work_struct *work)
328{
329 struct amd_sched_job *s_job = container_of(work, struct amd_sched_job,
330 finish_work);
331 struct amd_gpu_scheduler *sched = s_job->sched;
332
333 /* remove job from ring_mirror_list */
334 spin_lock(&sched->job_list_lock);
335 list_del_init(&s_job->node);
336 if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
337 struct amd_sched_job *next;
338
339 spin_unlock(&sched->job_list_lock);
340 cancel_delayed_work_sync(&s_job->work_tdr);
341 spin_lock(&sched->job_list_lock);
342
343 /* queue TDR for next job */
344 next = list_first_entry_or_null(&sched->ring_mirror_list,
345 struct amd_sched_job, node);
346
347 if (next)
348 schedule_delayed_work(&next->work_tdr, sched->timeout);
349 }
350 spin_unlock(&sched->job_list_lock);
351 sched->ops->free_job(s_job);
352}
353
354static void amd_sched_job_finish_cb(struct dma_fence *f,
355 struct dma_fence_cb *cb)
356{
357 struct amd_sched_job *job = container_of(cb, struct amd_sched_job,
358 finish_cb);
359 schedule_work(&job->finish_work);
360}
361
362static void amd_sched_job_begin(struct amd_sched_job *s_job)
363{
364 struct amd_gpu_scheduler *sched = s_job->sched;
365
366 spin_lock(&sched->job_list_lock);
367 list_add_tail(&s_job->node, &sched->ring_mirror_list);
368 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
369 list_first_entry_or_null(&sched->ring_mirror_list,
370 struct amd_sched_job, node) == s_job)
371 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
372 spin_unlock(&sched->job_list_lock);
373}
374
375static void amd_sched_job_timedout(struct work_struct *work)
376{
377 struct amd_sched_job *job = container_of(work, struct amd_sched_job,
378 work_tdr.work);
379
380 job->sched->ops->timedout_job(job);
381}
382
383void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched)
384{
385 struct amd_sched_job *s_job;
386
387 spin_lock(&sched->job_list_lock);
388 list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
389 if (dma_fence_remove_callback(s_job->s_fence->parent, &s_job->s_fence->cb)) {
390 dma_fence_put(s_job->s_fence->parent);
391 s_job->s_fence->parent = NULL;
392 }
393 }
394 atomic_set(&sched->hw_rq_count, 0);
395 spin_unlock(&sched->job_list_lock);
396}
397
398void amd_sched_job_recovery(struct amd_gpu_scheduler *sched)
399{
400 struct amd_sched_job *s_job, *tmp;
401 int r;
402
403 spin_lock(&sched->job_list_lock);
404 s_job = list_first_entry_or_null(&sched->ring_mirror_list,
405 struct amd_sched_job, node);
406 if (s_job && sched->timeout != MAX_SCHEDULE_TIMEOUT)
407 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
408
409 list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
410 struct amd_sched_fence *s_fence = s_job->s_fence;
411 struct dma_fence *fence;
412
413 spin_unlock(&sched->job_list_lock);
414 fence = sched->ops->run_job(s_job);
415 atomic_inc(&sched->hw_rq_count);
416 if (fence) {
417 s_fence->parent = dma_fence_get(fence);
418 r = dma_fence_add_callback(fence, &s_fence->cb,
419 amd_sched_process_job);
420 if (r == -ENOENT)
421 amd_sched_process_job(fence, &s_fence->cb);
422 else if (r)
423 DRM_ERROR("fence add callback failed (%d)\n",
424 r);
425 dma_fence_put(fence);
426 } else {
427 DRM_ERROR("Failed to run job!\n");
428 amd_sched_process_job(NULL, &s_fence->cb);
429 }
430 spin_lock(&sched->job_list_lock);
431 }
432 spin_unlock(&sched->job_list_lock);
433}
434
435/**
436 * Submit a job to the job queue
437 *
438 * @sched_job The pointer to job required to submit
439 *
440 * Returns 0 for success, negative error code otherwise.
441 */
442void amd_sched_entity_push_job(struct amd_sched_job *sched_job)
443{
444 struct amd_sched_entity *entity = sched_job->s_entity;
445
446 trace_amd_sched_job(sched_job);
447 dma_fence_add_callback(&sched_job->s_fence->finished, &sched_job->finish_cb,
448 amd_sched_job_finish_cb);
449 wait_event(entity->sched->job_scheduled,
450 amd_sched_entity_in(sched_job));
451}
452
453/* init a sched_job with basic field */
454int amd_sched_job_init(struct amd_sched_job *job,
455 struct amd_gpu_scheduler *sched,
456 struct amd_sched_entity *entity,
457 void *owner)
458{
459 job->sched = sched;
460 job->s_entity = entity;
461 job->s_fence = amd_sched_fence_create(entity, owner);
462 if (!job->s_fence)
463 return -ENOMEM;
464
465 INIT_WORK(&job->finish_work, amd_sched_job_finish);
466 INIT_LIST_HEAD(&job->node);
467 INIT_DELAYED_WORK(&job->work_tdr, amd_sched_job_timedout);
468
469 return 0;
470}
471
472/**
473 * Return ture if we can push more jobs to the hw.
474 */
475static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
476{
477 return atomic_read(&sched->hw_rq_count) <
478 sched->hw_submission_limit;
479}
480
481/**
482 * Wake up the scheduler when it is ready
483 */
484static void amd_sched_wakeup(struct amd_gpu_scheduler *sched)
485{
486 if (amd_sched_ready(sched))
487 wake_up_interruptible(&sched->wake_up_worker);
488}
489
490/**
491 * Select next entity to process
492*/
493static struct amd_sched_entity *
494amd_sched_select_entity(struct amd_gpu_scheduler *sched)
495{
496 struct amd_sched_entity *entity;
497 int i;
498
499 if (!amd_sched_ready(sched))
500 return NULL;
501
502 /* Kernel run queue has higher priority than normal run queue*/
503 for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++) {
504 entity = amd_sched_rq_select_entity(&sched->sched_rq[i]);
505 if (entity)
506 break;
507 }
508
509 return entity;
510}
511
512static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
513{
514 struct amd_sched_fence *s_fence =
515 container_of(cb, struct amd_sched_fence, cb);
516 struct amd_gpu_scheduler *sched = s_fence->sched;
517
518 atomic_dec(&sched->hw_rq_count);
519 amd_sched_fence_finished(s_fence);
520
521 trace_amd_sched_process_job(s_fence);
522 dma_fence_put(&s_fence->finished);
523 wake_up_interruptible(&sched->wake_up_worker);
524}
525
526static bool amd_sched_blocked(struct amd_gpu_scheduler *sched)
527{
528 if (kthread_should_park()) {
529 kthread_parkme();
530 return true;
531 }
532
533 return false;
534}
535
536static int amd_sched_main(void *param)
537{
538 struct sched_param sparam = {.sched_priority = 1};
539 struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
540 int r, count;
541
542 sched_setscheduler(current, SCHED_FIFO, &sparam);
543
544 while (!kthread_should_stop()) {
545 struct amd_sched_entity *entity = NULL;
546 struct amd_sched_fence *s_fence;
547 struct amd_sched_job *sched_job;
548 struct dma_fence *fence;
549
550 wait_event_interruptible(sched->wake_up_worker,
551 (!amd_sched_blocked(sched) &&
552 (entity = amd_sched_select_entity(sched))) ||
553 kthread_should_stop());
554
555 if (!entity)
556 continue;
557
558 sched_job = amd_sched_entity_pop_job(entity);
559 if (!sched_job)
560 continue;
561
562 s_fence = sched_job->s_fence;
563
564 atomic_inc(&sched->hw_rq_count);
565 amd_sched_job_begin(sched_job);
566
567 fence = sched->ops->run_job(sched_job);
568 amd_sched_fence_scheduled(s_fence);
569 if (fence) {
570 s_fence->parent = dma_fence_get(fence);
571 r = dma_fence_add_callback(fence, &s_fence->cb,
572 amd_sched_process_job);
573 if (r == -ENOENT)
574 amd_sched_process_job(fence, &s_fence->cb);
575 else if (r)
576 DRM_ERROR("fence add callback failed (%d)\n",
577 r);
578 dma_fence_put(fence);
579 } else {
580 DRM_ERROR("Failed to run job!\n");
581 amd_sched_process_job(NULL, &s_fence->cb);
582 }
583
584 count = kfifo_out(&entity->job_queue, &sched_job,
585 sizeof(sched_job));
586 WARN_ON(count != sizeof(sched_job));
587 wake_up(&sched->job_scheduled);
588 }
589 return 0;
590}
591
592/**
593 * Init a gpu scheduler instance
594 *
595 * @sched The pointer to the scheduler
596 * @ops The backend operations for this scheduler.
597 * @hw_submissions Number of hw submissions to do.
598 * @name Name used for debugging
599 *
600 * Return 0 on success, otherwise error code.
601*/
602int amd_sched_init(struct amd_gpu_scheduler *sched,
603 const struct amd_sched_backend_ops *ops,
604 unsigned hw_submission, long timeout, const char *name)
605{
606 int i;
607 sched->ops = ops;
608 sched->hw_submission_limit = hw_submission;
609 sched->name = name;
610 sched->timeout = timeout;
611 for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++)
612 amd_sched_rq_init(&sched->sched_rq[i]);
613
614 init_waitqueue_head(&sched->wake_up_worker);
615 init_waitqueue_head(&sched->job_scheduled);
616 INIT_LIST_HEAD(&sched->ring_mirror_list);
617 spin_lock_init(&sched->job_list_lock);
618 atomic_set(&sched->hw_rq_count, 0);
619
620 /* Each scheduler will run on a seperate kernel thread */
621 sched->thread = kthread_run(amd_sched_main, sched, sched->name);
622 if (IS_ERR(sched->thread)) {
623 DRM_ERROR("Failed to create scheduler for %s.\n", name);
624 return PTR_ERR(sched->thread);
625 }
626
627 return 0;
628}
629
630/**
631 * Destroy a gpu scheduler
632 *
633 * @sched The pointer to the scheduler
634 */
635void amd_sched_fini(struct amd_gpu_scheduler *sched)
636{
637 if (sched->thread)
638 kthread_stop(sched->thread);
639}