]> git.karo-electronics.de Git - karo-tx-linux.git/blob - drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
Merge tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
[karo-tx-linux.git] / drivers / gpu / drm / amd / scheduler / gpu_scheduler.c
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 <uapi/linux/sched/types.h>
28 #include <drm/drmP.h>
29 #include "gpu_scheduler.h"
30
31 #define CREATE_TRACE_POINTS
32 #include "gpu_sched_trace.h"
33
34 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity);
35 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched);
36 static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
37
38 /* Initialize a given run queue struct */
39 static void amd_sched_rq_init(struct amd_sched_rq *rq)
40 {
41         spin_lock_init(&rq->lock);
42         INIT_LIST_HEAD(&rq->entities);
43         rq->current_entity = NULL;
44 }
45
46 static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
47                                     struct amd_sched_entity *entity)
48 {
49         if (!list_empty(&entity->list))
50                 return;
51         spin_lock(&rq->lock);
52         list_add_tail(&entity->list, &rq->entities);
53         spin_unlock(&rq->lock);
54 }
55
56 static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
57                                        struct amd_sched_entity *entity)
58 {
59         if (list_empty(&entity->list))
60                 return;
61         spin_lock(&rq->lock);
62         list_del_init(&entity->list);
63         if (rq->current_entity == entity)
64                 rq->current_entity = NULL;
65         spin_unlock(&rq->lock);
66 }
67
68 /**
69  * Select an entity which could provide a job to run
70  *
71  * @rq          The run queue to check.
72  *
73  * Try to find a ready entity, returns NULL if none found.
74  */
75 static struct amd_sched_entity *
76 amd_sched_rq_select_entity(struct amd_sched_rq *rq)
77 {
78         struct amd_sched_entity *entity;
79
80         spin_lock(&rq->lock);
81
82         entity = rq->current_entity;
83         if (entity) {
84                 list_for_each_entry_continue(entity, &rq->entities, list) {
85                         if (amd_sched_entity_is_ready(entity)) {
86                                 rq->current_entity = entity;
87                                 spin_unlock(&rq->lock);
88                                 return entity;
89                         }
90                 }
91         }
92
93         list_for_each_entry(entity, &rq->entities, list) {
94
95                 if (amd_sched_entity_is_ready(entity)) {
96                         rq->current_entity = entity;
97                         spin_unlock(&rq->lock);
98                         return entity;
99                 }
100
101                 if (entity == rq->current_entity)
102                         break;
103         }
104
105         spin_unlock(&rq->lock);
106
107         return NULL;
108 }
109
110 /**
111  * Init a context entity used by scheduler when submit to HW ring.
112  *
113  * @sched       The pointer to the scheduler
114  * @entity      The pointer to a valid amd_sched_entity
115  * @rq          The run queue this entity belongs
116  * @kernel      If this is an entity for the kernel
117  * @jobs        The max number of jobs in the job queue
118  *
119  * return 0 if succeed. negative error code on failure
120 */
121 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
122                           struct amd_sched_entity *entity,
123                           struct amd_sched_rq *rq,
124                           uint32_t jobs)
125 {
126         int r;
127
128         if (!(sched && entity && rq))
129                 return -EINVAL;
130
131         memset(entity, 0, sizeof(struct amd_sched_entity));
132         INIT_LIST_HEAD(&entity->list);
133         entity->rq = rq;
134         entity->sched = sched;
135
136         spin_lock_init(&entity->queue_lock);
137         r = kfifo_alloc(&entity->job_queue, jobs * sizeof(void *), GFP_KERNEL);
138         if (r)
139                 return r;
140
141         atomic_set(&entity->fence_seq, 0);
142         entity->fence_context = dma_fence_context_alloc(2);
143
144         return 0;
145 }
146
147 /**
148  * Query if entity is initialized
149  *
150  * @sched       Pointer to scheduler instance
151  * @entity      The pointer to a valid scheduler entity
152  *
153  * return true if entity is initialized, false otherwise
154 */
155 static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
156                                             struct amd_sched_entity *entity)
157 {
158         return entity->sched == sched &&
159                 entity->rq != NULL;
160 }
161
162 /**
163  * Check if entity is idle
164  *
165  * @entity      The pointer to a valid scheduler entity
166  *
167  * Return true if entity don't has any unscheduled jobs.
168  */
169 static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
170 {
171         rmb();
172         if (kfifo_is_empty(&entity->job_queue))
173                 return true;
174
175         return false;
176 }
177
178 /**
179  * Check if entity is ready
180  *
181  * @entity      The pointer to a valid scheduler entity
182  *
183  * Return true if entity could provide a job.
184  */
185 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity)
186 {
187         if (kfifo_is_empty(&entity->job_queue))
188                 return false;
189
190         if (ACCESS_ONCE(entity->dependency))
191                 return false;
192
193         return true;
194 }
195
196 /**
197  * Destroy a context entity
198  *
199  * @sched       Pointer to scheduler instance
200  * @entity      The pointer to a valid scheduler entity
201  *
202  * Cleanup and free the allocated resources.
203  */
204 void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
205                            struct amd_sched_entity *entity)
206 {
207         struct amd_sched_rq *rq = entity->rq;
208
209         if (!amd_sched_entity_is_initialized(sched, entity))
210                 return;
211
212         /**
213          * The client will not queue more IBs during this fini, consume existing
214          * queued IBs
215         */
216         wait_event(sched->job_scheduled, amd_sched_entity_is_idle(entity));
217
218         amd_sched_rq_remove_entity(rq, entity);
219         kfifo_free(&entity->job_queue);
220 }
221
222 static void amd_sched_entity_wakeup(struct dma_fence *f, struct dma_fence_cb *cb)
223 {
224         struct amd_sched_entity *entity =
225                 container_of(cb, struct amd_sched_entity, cb);
226         entity->dependency = NULL;
227         dma_fence_put(f);
228         amd_sched_wakeup(entity->sched);
229 }
230
231 static void amd_sched_entity_clear_dep(struct dma_fence *f, struct dma_fence_cb *cb)
232 {
233         struct amd_sched_entity *entity =
234                 container_of(cb, struct amd_sched_entity, cb);
235         entity->dependency = NULL;
236         dma_fence_put(f);
237 }
238
239 bool amd_sched_dependency_optimized(struct dma_fence* fence,
240                                     struct amd_sched_entity *entity)
241 {
242         struct amd_gpu_scheduler *sched = entity->sched;
243         struct amd_sched_fence *s_fence;
244
245         if (!fence || dma_fence_is_signaled(fence))
246                 return false;
247         if (fence->context == entity->fence_context)
248                 return true;
249         s_fence = to_amd_sched_fence(fence);
250         if (s_fence && s_fence->sched == sched)
251                 return true;
252
253         return false;
254 }
255
256 static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
257 {
258         struct amd_gpu_scheduler *sched = entity->sched;
259         struct dma_fence * fence = entity->dependency;
260         struct amd_sched_fence *s_fence;
261
262         if (fence->context == entity->fence_context) {
263                 /* We can ignore fences from ourself */
264                 dma_fence_put(entity->dependency);
265                 return false;
266         }
267
268         s_fence = to_amd_sched_fence(fence);
269         if (s_fence && s_fence->sched == sched) {
270
271                 /*
272                  * Fence is from the same scheduler, only need to wait for
273                  * it to be scheduled
274                  */
275                 fence = dma_fence_get(&s_fence->scheduled);
276                 dma_fence_put(entity->dependency);
277                 entity->dependency = fence;
278                 if (!dma_fence_add_callback(fence, &entity->cb,
279                                             amd_sched_entity_clear_dep))
280                         return true;
281
282                 /* Ignore it when it is already scheduled */
283                 dma_fence_put(fence);
284                 return false;
285         }
286
287         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
288                                     amd_sched_entity_wakeup))
289                 return true;
290
291         dma_fence_put(entity->dependency);
292         return false;
293 }
294
295 static struct amd_sched_job *
296 amd_sched_entity_pop_job(struct amd_sched_entity *entity)
297 {
298         struct amd_gpu_scheduler *sched = entity->sched;
299         struct amd_sched_job *sched_job;
300
301         if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job)))
302                 return NULL;
303
304         while ((entity->dependency = sched->ops->dependency(sched_job)))
305                 if (amd_sched_entity_add_dependency_cb(entity))
306                         return NULL;
307
308         return sched_job;
309 }
310
311 /**
312  * Helper to submit a job to the job queue
313  *
314  * @sched_job           The pointer to job required to submit
315  *
316  * Returns true if we could submit the job.
317  */
318 static bool amd_sched_entity_in(struct amd_sched_job *sched_job)
319 {
320         struct amd_gpu_scheduler *sched = sched_job->sched;
321         struct amd_sched_entity *entity = sched_job->s_entity;
322         bool added, first = false;
323
324         spin_lock(&entity->queue_lock);
325         added = kfifo_in(&entity->job_queue, &sched_job,
326                         sizeof(sched_job)) == sizeof(sched_job);
327
328         if (added && kfifo_len(&entity->job_queue) == sizeof(sched_job))
329                 first = true;
330
331         spin_unlock(&entity->queue_lock);
332
333         /* first job wakes up scheduler */
334         if (first) {
335                 /* Add the entity to the run queue */
336                 amd_sched_rq_add_entity(entity->rq, entity);
337                 amd_sched_wakeup(sched);
338         }
339         return added;
340 }
341
342 /* job_finish is called after hw fence signaled, and
343  * the job had already been deleted from ring_mirror_list
344  */
345 static void amd_sched_job_finish(struct work_struct *work)
346 {
347         struct amd_sched_job *s_job = container_of(work, struct amd_sched_job,
348                                                    finish_work);
349         struct amd_gpu_scheduler *sched = s_job->sched;
350
351         /* remove job from ring_mirror_list */
352         spin_lock(&sched->job_list_lock);
353         list_del_init(&s_job->node);
354         if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
355                 struct amd_sched_job *next;
356
357                 spin_unlock(&sched->job_list_lock);
358                 cancel_delayed_work_sync(&s_job->work_tdr);
359                 spin_lock(&sched->job_list_lock);
360
361                 /* queue TDR for next job */
362                 next = list_first_entry_or_null(&sched->ring_mirror_list,
363                                                 struct amd_sched_job, node);
364
365                 if (next)
366                         schedule_delayed_work(&next->work_tdr, sched->timeout);
367         }
368         spin_unlock(&sched->job_list_lock);
369         sched->ops->free_job(s_job);
370 }
371
372 static void amd_sched_job_finish_cb(struct dma_fence *f,
373                                     struct dma_fence_cb *cb)
374 {
375         struct amd_sched_job *job = container_of(cb, struct amd_sched_job,
376                                                  finish_cb);
377         schedule_work(&job->finish_work);
378 }
379
380 static void amd_sched_job_begin(struct amd_sched_job *s_job)
381 {
382         struct amd_gpu_scheduler *sched = s_job->sched;
383
384         spin_lock(&sched->job_list_lock);
385         list_add_tail(&s_job->node, &sched->ring_mirror_list);
386         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
387             list_first_entry_or_null(&sched->ring_mirror_list,
388                                      struct amd_sched_job, node) == s_job)
389                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
390         spin_unlock(&sched->job_list_lock);
391 }
392
393 static void amd_sched_job_timedout(struct work_struct *work)
394 {
395         struct amd_sched_job *job = container_of(work, struct amd_sched_job,
396                                                  work_tdr.work);
397
398         job->sched->ops->timedout_job(job);
399 }
400
401 void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched)
402 {
403         struct amd_sched_job *s_job;
404
405         spin_lock(&sched->job_list_lock);
406         list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
407                 if (s_job->s_fence->parent &&
408                     dma_fence_remove_callback(s_job->s_fence->parent,
409                                               &s_job->s_fence->cb)) {
410                         dma_fence_put(s_job->s_fence->parent);
411                         s_job->s_fence->parent = NULL;
412                 }
413         }
414         atomic_set(&sched->hw_rq_count, 0);
415         spin_unlock(&sched->job_list_lock);
416 }
417
418 void amd_sched_job_recovery(struct amd_gpu_scheduler *sched)
419 {
420         struct amd_sched_job *s_job, *tmp;
421         int r;
422
423         spin_lock(&sched->job_list_lock);
424         s_job = list_first_entry_or_null(&sched->ring_mirror_list,
425                                          struct amd_sched_job, node);
426         if (s_job && sched->timeout != MAX_SCHEDULE_TIMEOUT)
427                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
428
429         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
430                 struct amd_sched_fence *s_fence = s_job->s_fence;
431                 struct dma_fence *fence;
432
433                 spin_unlock(&sched->job_list_lock);
434                 fence = sched->ops->run_job(s_job);
435                 atomic_inc(&sched->hw_rq_count);
436                 if (fence) {
437                         s_fence->parent = dma_fence_get(fence);
438                         r = dma_fence_add_callback(fence, &s_fence->cb,
439                                                    amd_sched_process_job);
440                         if (r == -ENOENT)
441                                 amd_sched_process_job(fence, &s_fence->cb);
442                         else if (r)
443                                 DRM_ERROR("fence add callback failed (%d)\n",
444                                           r);
445                         dma_fence_put(fence);
446                 } else {
447                         DRM_ERROR("Failed to run job!\n");
448                         amd_sched_process_job(NULL, &s_fence->cb);
449                 }
450                 spin_lock(&sched->job_list_lock);
451         }
452         spin_unlock(&sched->job_list_lock);
453 }
454
455 /**
456  * Submit a job to the job queue
457  *
458  * @sched_job           The pointer to job required to submit
459  *
460  * Returns 0 for success, negative error code otherwise.
461  */
462 void amd_sched_entity_push_job(struct amd_sched_job *sched_job)
463 {
464         struct amd_sched_entity *entity = sched_job->s_entity;
465
466         trace_amd_sched_job(sched_job);
467         dma_fence_add_callback(&sched_job->s_fence->finished, &sched_job->finish_cb,
468                                amd_sched_job_finish_cb);
469         wait_event(entity->sched->job_scheduled,
470                    amd_sched_entity_in(sched_job));
471 }
472
473 /* init a sched_job with basic field */
474 int amd_sched_job_init(struct amd_sched_job *job,
475                        struct amd_gpu_scheduler *sched,
476                        struct amd_sched_entity *entity,
477                        void *owner)
478 {
479         job->sched = sched;
480         job->s_entity = entity;
481         job->s_fence = amd_sched_fence_create(entity, owner);
482         if (!job->s_fence)
483                 return -ENOMEM;
484         job->id = atomic64_inc_return(&sched->job_id_count);
485
486         INIT_WORK(&job->finish_work, amd_sched_job_finish);
487         INIT_LIST_HEAD(&job->node);
488         INIT_DELAYED_WORK(&job->work_tdr, amd_sched_job_timedout);
489
490         return 0;
491 }
492
493 /**
494  * Return ture if we can push more jobs to the hw.
495  */
496 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
497 {
498         return atomic_read(&sched->hw_rq_count) <
499                 sched->hw_submission_limit;
500 }
501
502 /**
503  * Wake up the scheduler when it is ready
504  */
505 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched)
506 {
507         if (amd_sched_ready(sched))
508                 wake_up_interruptible(&sched->wake_up_worker);
509 }
510
511 /**
512  * Select next entity to process
513 */
514 static struct amd_sched_entity *
515 amd_sched_select_entity(struct amd_gpu_scheduler *sched)
516 {
517         struct amd_sched_entity *entity;
518         int i;
519
520         if (!amd_sched_ready(sched))
521                 return NULL;
522
523         /* Kernel run queue has higher priority than normal run queue*/
524         for (i = AMD_SCHED_PRIORITY_MAX - 1; i >= AMD_SCHED_PRIORITY_MIN; i--) {
525                 entity = amd_sched_rq_select_entity(&sched->sched_rq[i]);
526                 if (entity)
527                         break;
528         }
529
530         return entity;
531 }
532
533 static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
534 {
535         struct amd_sched_fence *s_fence =
536                 container_of(cb, struct amd_sched_fence, cb);
537         struct amd_gpu_scheduler *sched = s_fence->sched;
538
539         atomic_dec(&sched->hw_rq_count);
540         amd_sched_fence_finished(s_fence);
541
542         trace_amd_sched_process_job(s_fence);
543         dma_fence_put(&s_fence->finished);
544         wake_up_interruptible(&sched->wake_up_worker);
545 }
546
547 static bool amd_sched_blocked(struct amd_gpu_scheduler *sched)
548 {
549         if (kthread_should_park()) {
550                 kthread_parkme();
551                 return true;
552         }
553
554         return false;
555 }
556
557 static int amd_sched_main(void *param)
558 {
559         struct sched_param sparam = {.sched_priority = 1};
560         struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
561         int r, count;
562
563         sched_setscheduler(current, SCHED_FIFO, &sparam);
564
565         while (!kthread_should_stop()) {
566                 struct amd_sched_entity *entity = NULL;
567                 struct amd_sched_fence *s_fence;
568                 struct amd_sched_job *sched_job;
569                 struct dma_fence *fence;
570
571                 wait_event_interruptible(sched->wake_up_worker,
572                                          (!amd_sched_blocked(sched) &&
573                                           (entity = amd_sched_select_entity(sched))) ||
574                                          kthread_should_stop());
575
576                 if (!entity)
577                         continue;
578
579                 sched_job = amd_sched_entity_pop_job(entity);
580                 if (!sched_job)
581                         continue;
582
583                 s_fence = sched_job->s_fence;
584
585                 atomic_inc(&sched->hw_rq_count);
586                 amd_sched_job_begin(sched_job);
587
588                 fence = sched->ops->run_job(sched_job);
589                 amd_sched_fence_scheduled(s_fence);
590                 if (fence) {
591                         s_fence->parent = dma_fence_get(fence);
592                         r = dma_fence_add_callback(fence, &s_fence->cb,
593                                                    amd_sched_process_job);
594                         if (r == -ENOENT)
595                                 amd_sched_process_job(fence, &s_fence->cb);
596                         else if (r)
597                                 DRM_ERROR("fence add callback failed (%d)\n",
598                                           r);
599                         dma_fence_put(fence);
600                 } else {
601                         DRM_ERROR("Failed to run job!\n");
602                         amd_sched_process_job(NULL, &s_fence->cb);
603                 }
604
605                 count = kfifo_out(&entity->job_queue, &sched_job,
606                                 sizeof(sched_job));
607                 WARN_ON(count != sizeof(sched_job));
608                 wake_up(&sched->job_scheduled);
609         }
610         return 0;
611 }
612
613 /**
614  * Init a gpu scheduler instance
615  *
616  * @sched               The pointer to the scheduler
617  * @ops                 The backend operations for this scheduler.
618  * @hw_submissions      Number of hw submissions to do.
619  * @name                Name used for debugging
620  *
621  * Return 0 on success, otherwise error code.
622 */
623 int amd_sched_init(struct amd_gpu_scheduler *sched,
624                    const struct amd_sched_backend_ops *ops,
625                    unsigned hw_submission, long timeout, const char *name)
626 {
627         int i;
628         sched->ops = ops;
629         sched->hw_submission_limit = hw_submission;
630         sched->name = name;
631         sched->timeout = timeout;
632         for (i = AMD_SCHED_PRIORITY_MIN; i < AMD_SCHED_PRIORITY_MAX; i++)
633                 amd_sched_rq_init(&sched->sched_rq[i]);
634
635         init_waitqueue_head(&sched->wake_up_worker);
636         init_waitqueue_head(&sched->job_scheduled);
637         INIT_LIST_HEAD(&sched->ring_mirror_list);
638         spin_lock_init(&sched->job_list_lock);
639         atomic_set(&sched->hw_rq_count, 0);
640         atomic64_set(&sched->job_id_count, 0);
641
642         /* Each scheduler will run on a seperate kernel thread */
643         sched->thread = kthread_run(amd_sched_main, sched, sched->name);
644         if (IS_ERR(sched->thread)) {
645                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
646                 return PTR_ERR(sched->thread);
647         }
648
649         return 0;
650 }
651
652 /**
653  * Destroy a gpu scheduler
654  *
655  * @sched       The pointer to the scheduler
656  */
657 void amd_sched_fini(struct amd_gpu_scheduler *sched)
658 {
659         if (sched->thread)
660                 kthread_stop(sched->thread);
661 }