Loading...
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/slab.h>
26#include <linux/completion.h>
27
28#include <drm/drm_print.h>
29#include <drm/gpu_scheduler.h>
30
31#include "gpu_scheduler_trace.h"
32
33#define to_drm_sched_job(sched_job) \
34 container_of((sched_job), struct drm_sched_job, queue_node)
35
36/**
37 * drm_sched_entity_init - Init a context entity used by scheduler when
38 * submit to HW ring.
39 *
40 * @entity: scheduler entity to init
41 * @priority: priority of the entity
42 * @sched_list: the list of drm scheds on which jobs from this
43 * entity can be submitted
44 * @num_sched_list: number of drm sched in sched_list
45 * @guilty: atomic_t set to 1 when a job on this queue
46 * is found to be guilty causing a timeout
47 *
48 * Note that the &sched_list must have at least one element to schedule the entity.
49 *
50 * For changing @priority later on at runtime see
51 * drm_sched_entity_set_priority(). For changing the set of schedulers
52 * @sched_list at runtime see drm_sched_entity_modify_sched().
53 *
54 * An entity is cleaned up by calling drm_sched_entity_fini(). See also
55 * drm_sched_entity_destroy().
56 *
57 * Returns 0 on success or a negative error code on failure.
58 */
59int drm_sched_entity_init(struct drm_sched_entity *entity,
60 enum drm_sched_priority priority,
61 struct drm_gpu_scheduler **sched_list,
62 unsigned int num_sched_list,
63 atomic_t *guilty)
64{
65 if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0])))
66 return -EINVAL;
67
68 memset(entity, 0, sizeof(struct drm_sched_entity));
69 INIT_LIST_HEAD(&entity->list);
70 entity->rq = NULL;
71 entity->guilty = guilty;
72 entity->num_sched_list = num_sched_list;
73 entity->priority = priority;
74 /*
75 * It's perfectly valid to initialize an entity without having a valid
76 * scheduler attached. It's just not valid to use the scheduler before it
77 * is initialized itself.
78 */
79 entity->sched_list = num_sched_list > 1 ? sched_list : NULL;
80 RCU_INIT_POINTER(entity->last_scheduled, NULL);
81 RB_CLEAR_NODE(&entity->rb_tree_node);
82
83 if (num_sched_list && !sched_list[0]->sched_rq) {
84 /* Since every entry covered by num_sched_list
85 * should be non-NULL and therefore we warn drivers
86 * not to do this and to fix their DRM calling order.
87 */
88 pr_warn("%s: called with uninitialized scheduler\n", __func__);
89 } else if (num_sched_list) {
90 /* The "priority" of an entity cannot exceed the number of run-queues of a
91 * scheduler. Protect against num_rqs being 0, by converting to signed. Choose
92 * the lowest priority available.
93 */
94 if (entity->priority >= sched_list[0]->num_rqs) {
95 drm_err(sched_list[0], "entity with out-of-bounds priority:%u num_rqs:%u\n",
96 entity->priority, sched_list[0]->num_rqs);
97 entity->priority = max_t(s32, (s32) sched_list[0]->num_rqs - 1,
98 (s32) DRM_SCHED_PRIORITY_KERNEL);
99 }
100 entity->rq = sched_list[0]->sched_rq[entity->priority];
101 }
102
103 init_completion(&entity->entity_idle);
104
105 /* We start in an idle state. */
106 complete_all(&entity->entity_idle);
107
108 spin_lock_init(&entity->lock);
109 spsc_queue_init(&entity->job_queue);
110
111 atomic_set(&entity->fence_seq, 0);
112 entity->fence_context = dma_fence_context_alloc(2);
113
114 return 0;
115}
116EXPORT_SYMBOL(drm_sched_entity_init);
117
118/**
119 * drm_sched_entity_modify_sched - Modify sched of an entity
120 * @entity: scheduler entity to init
121 * @sched_list: the list of new drm scheds which will replace
122 * existing entity->sched_list
123 * @num_sched_list: number of drm sched in sched_list
124 *
125 * Note that this must be called under the same common lock for @entity as
126 * drm_sched_job_arm() and drm_sched_entity_push_job(), or the driver needs to
127 * guarantee through some other means that this is never called while new jobs
128 * can be pushed to @entity.
129 */
130void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
131 struct drm_gpu_scheduler **sched_list,
132 unsigned int num_sched_list)
133{
134 WARN_ON(!num_sched_list || !sched_list);
135
136 spin_lock(&entity->lock);
137 entity->sched_list = sched_list;
138 entity->num_sched_list = num_sched_list;
139 spin_unlock(&entity->lock);
140}
141EXPORT_SYMBOL(drm_sched_entity_modify_sched);
142
143static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
144{
145 rmb(); /* for list_empty to work without lock */
146
147 if (list_empty(&entity->list) ||
148 spsc_queue_count(&entity->job_queue) == 0 ||
149 entity->stopped)
150 return true;
151
152 return false;
153}
154
155/* Return true if entity could provide a job. */
156bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
157{
158 if (spsc_queue_peek(&entity->job_queue) == NULL)
159 return false;
160
161 if (READ_ONCE(entity->dependency))
162 return false;
163
164 return true;
165}
166
167/**
168 * drm_sched_entity_error - return error of last scheduled job
169 * @entity: scheduler entity to check
170 *
171 * Opportunistically return the error of the last scheduled job. Result can
172 * change any time when new jobs are pushed to the hw.
173 */
174int drm_sched_entity_error(struct drm_sched_entity *entity)
175{
176 struct dma_fence *fence;
177 int r;
178
179 rcu_read_lock();
180 fence = rcu_dereference(entity->last_scheduled);
181 r = fence ? fence->error : 0;
182 rcu_read_unlock();
183
184 return r;
185}
186EXPORT_SYMBOL(drm_sched_entity_error);
187
188static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)
189{
190 struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
191
192 drm_sched_fence_finished(job->s_fence, -ESRCH);
193 WARN_ON(job->s_fence->parent);
194 job->sched->ops->free_job(job);
195}
196
197/* Signal the scheduler finished fence when the entity in question is killed. */
198static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
199 struct dma_fence_cb *cb)
200{
201 struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
202 finish_cb);
203 unsigned long index;
204
205 dma_fence_put(f);
206
207 /* Wait for all dependencies to avoid data corruptions */
208 xa_for_each(&job->dependencies, index, f) {
209 struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
210
211 if (s_fence && f == &s_fence->scheduled) {
212 /* The dependencies array had a reference on the scheduled
213 * fence, and the finished fence refcount might have
214 * dropped to zero. Use dma_fence_get_rcu() so we get
215 * a NULL fence in that case.
216 */
217 f = dma_fence_get_rcu(&s_fence->finished);
218
219 /* Now that we have a reference on the finished fence,
220 * we can release the reference the dependencies array
221 * had on the scheduled fence.
222 */
223 dma_fence_put(&s_fence->scheduled);
224 }
225
226 xa_erase(&job->dependencies, index);
227 if (f && !dma_fence_add_callback(f, &job->finish_cb,
228 drm_sched_entity_kill_jobs_cb))
229 return;
230
231 dma_fence_put(f);
232 }
233
234 INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work);
235 schedule_work(&job->work);
236}
237
238/* Remove the entity from the scheduler and kill all pending jobs */
239static void drm_sched_entity_kill(struct drm_sched_entity *entity)
240{
241 struct drm_sched_job *job;
242 struct dma_fence *prev;
243
244 if (!entity->rq)
245 return;
246
247 spin_lock(&entity->lock);
248 entity->stopped = true;
249 drm_sched_rq_remove_entity(entity->rq, entity);
250 spin_unlock(&entity->lock);
251
252 /* Make sure this entity is not used by the scheduler at the moment */
253 wait_for_completion(&entity->entity_idle);
254
255 /* The entity is guaranteed to not be used by the scheduler */
256 prev = rcu_dereference_check(entity->last_scheduled, true);
257 dma_fence_get(prev);
258 while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
259 struct drm_sched_fence *s_fence = job->s_fence;
260
261 dma_fence_get(&s_fence->finished);
262 if (!prev || dma_fence_add_callback(prev, &job->finish_cb,
263 drm_sched_entity_kill_jobs_cb))
264 drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
265
266 prev = &s_fence->finished;
267 }
268 dma_fence_put(prev);
269}
270
271/**
272 * drm_sched_entity_flush - Flush a context entity
273 *
274 * @entity: scheduler entity
275 * @timeout: time to wait in for Q to become empty in jiffies.
276 *
277 * Splitting drm_sched_entity_fini() into two functions, The first one does the
278 * waiting, removes the entity from the runqueue and returns an error when the
279 * process was killed.
280 *
281 * Returns the remaining time in jiffies left from the input timeout
282 */
283long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
284{
285 struct drm_gpu_scheduler *sched;
286 struct task_struct *last_user;
287 long ret = timeout;
288
289 if (!entity->rq)
290 return 0;
291
292 sched = entity->rq->sched;
293 /**
294 * The client will not queue more IBs during this fini, consume existing
295 * queued IBs or discard them on SIGKILL
296 */
297 if (current->flags & PF_EXITING) {
298 if (timeout)
299 ret = wait_event_timeout(
300 sched->job_scheduled,
301 drm_sched_entity_is_idle(entity),
302 timeout);
303 } else {
304 wait_event_killable(sched->job_scheduled,
305 drm_sched_entity_is_idle(entity));
306 }
307
308 /* For killed process disable any more IBs enqueue right now */
309 last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
310 if ((!last_user || last_user == current->group_leader) &&
311 (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))
312 drm_sched_entity_kill(entity);
313
314 return ret;
315}
316EXPORT_SYMBOL(drm_sched_entity_flush);
317
318/**
319 * drm_sched_entity_fini - Destroy a context entity
320 *
321 * @entity: scheduler entity
322 *
323 * Cleanups up @entity which has been initialized by drm_sched_entity_init().
324 *
325 * If there are potentially job still in flight or getting newly queued
326 * drm_sched_entity_flush() must be called first. This function then goes over
327 * the entity and signals all jobs with an error code if the process was killed.
328 */
329void drm_sched_entity_fini(struct drm_sched_entity *entity)
330{
331 /*
332 * If consumption of existing IBs wasn't completed. Forcefully remove
333 * them here. Also makes sure that the scheduler won't touch this entity
334 * any more.
335 */
336 drm_sched_entity_kill(entity);
337
338 if (entity->dependency) {
339 dma_fence_remove_callback(entity->dependency, &entity->cb);
340 dma_fence_put(entity->dependency);
341 entity->dependency = NULL;
342 }
343
344 dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
345 RCU_INIT_POINTER(entity->last_scheduled, NULL);
346}
347EXPORT_SYMBOL(drm_sched_entity_fini);
348
349/**
350 * drm_sched_entity_destroy - Destroy a context entity
351 * @entity: scheduler entity
352 *
353 * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a
354 * convenience wrapper.
355 */
356void drm_sched_entity_destroy(struct drm_sched_entity *entity)
357{
358 drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
359 drm_sched_entity_fini(entity);
360}
361EXPORT_SYMBOL(drm_sched_entity_destroy);
362
363/* drm_sched_entity_clear_dep - callback to clear the entities dependency */
364static void drm_sched_entity_clear_dep(struct dma_fence *f,
365 struct dma_fence_cb *cb)
366{
367 struct drm_sched_entity *entity =
368 container_of(cb, struct drm_sched_entity, cb);
369
370 entity->dependency = NULL;
371 dma_fence_put(f);
372}
373
374/*
375 * drm_sched_entity_wakeup - callback to clear the entity's dependency and
376 * wake up the scheduler
377 */
378static void drm_sched_entity_wakeup(struct dma_fence *f,
379 struct dma_fence_cb *cb)
380{
381 struct drm_sched_entity *entity =
382 container_of(cb, struct drm_sched_entity, cb);
383
384 drm_sched_entity_clear_dep(f, cb);
385 drm_sched_wakeup(entity->rq->sched);
386}
387
388/**
389 * drm_sched_entity_set_priority - Sets priority of the entity
390 *
391 * @entity: scheduler entity
392 * @priority: scheduler priority
393 *
394 * Update the priority of runqueues used for the entity.
395 */
396void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
397 enum drm_sched_priority priority)
398{
399 spin_lock(&entity->lock);
400 entity->priority = priority;
401 spin_unlock(&entity->lock);
402}
403EXPORT_SYMBOL(drm_sched_entity_set_priority);
404
405/*
406 * Add a callback to the current dependency of the entity to wake up the
407 * scheduler when the entity becomes available.
408 */
409static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
410{
411 struct drm_gpu_scheduler *sched = entity->rq->sched;
412 struct dma_fence *fence = entity->dependency;
413 struct drm_sched_fence *s_fence;
414
415 if (fence->context == entity->fence_context ||
416 fence->context == entity->fence_context + 1) {
417 /*
418 * Fence is a scheduled/finished fence from a job
419 * which belongs to the same entity, we can ignore
420 * fences from ourself
421 */
422 dma_fence_put(entity->dependency);
423 return false;
424 }
425
426 s_fence = to_drm_sched_fence(fence);
427 if (!fence->error && s_fence && s_fence->sched == sched &&
428 !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {
429
430 /*
431 * Fence is from the same scheduler, only need to wait for
432 * it to be scheduled
433 */
434 fence = dma_fence_get(&s_fence->scheduled);
435 dma_fence_put(entity->dependency);
436 entity->dependency = fence;
437 if (!dma_fence_add_callback(fence, &entity->cb,
438 drm_sched_entity_clear_dep))
439 return true;
440
441 /* Ignore it when it is already scheduled */
442 dma_fence_put(fence);
443 return false;
444 }
445
446 if (!dma_fence_add_callback(entity->dependency, &entity->cb,
447 drm_sched_entity_wakeup))
448 return true;
449
450 dma_fence_put(entity->dependency);
451 return false;
452}
453
454static struct dma_fence *
455drm_sched_job_dependency(struct drm_sched_job *job,
456 struct drm_sched_entity *entity)
457{
458 struct dma_fence *f;
459
460 /* We keep the fence around, so we can iterate over all dependencies
461 * in drm_sched_entity_kill_jobs_cb() to ensure all deps are signaled
462 * before killing the job.
463 */
464 f = xa_load(&job->dependencies, job->last_dependency);
465 if (f) {
466 job->last_dependency++;
467 return dma_fence_get(f);
468 }
469
470 if (job->sched->ops->prepare_job)
471 return job->sched->ops->prepare_job(job, entity);
472
473 return NULL;
474}
475
476struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
477{
478 struct drm_sched_job *sched_job;
479
480 sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
481 if (!sched_job)
482 return NULL;
483
484 while ((entity->dependency =
485 drm_sched_job_dependency(sched_job, entity))) {
486 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
487
488 if (drm_sched_entity_add_dependency_cb(entity))
489 return NULL;
490 }
491
492 /* skip jobs from entity that marked guilty */
493 if (entity->guilty && atomic_read(entity->guilty))
494 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
495
496 dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
497 rcu_assign_pointer(entity->last_scheduled,
498 dma_fence_get(&sched_job->s_fence->finished));
499
500 /*
501 * If the queue is empty we allow drm_sched_entity_select_rq() to
502 * locklessly access ->last_scheduled. This only works if we set the
503 * pointer before we dequeue and if we a write barrier here.
504 */
505 smp_wmb();
506
507 spsc_queue_pop(&entity->job_queue);
508
509 /*
510 * Update the entity's location in the min heap according to
511 * the timestamp of the next job, if any.
512 */
513 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {
514 struct drm_sched_job *next;
515
516 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
517 if (next) {
518 struct drm_sched_rq *rq;
519
520 spin_lock(&entity->lock);
521 rq = entity->rq;
522 spin_lock(&rq->lock);
523 drm_sched_rq_update_fifo_locked(entity, rq,
524 next->submit_ts);
525 spin_unlock(&rq->lock);
526 spin_unlock(&entity->lock);
527 }
528 }
529
530 /* Jobs and entities might have different lifecycles. Since we're
531 * removing the job from the entities queue, set the jobs entity pointer
532 * to NULL to prevent any future access of the entity through this job.
533 */
534 sched_job->entity = NULL;
535
536 return sched_job;
537}
538
539void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
540{
541 struct dma_fence *fence;
542 struct drm_gpu_scheduler *sched;
543 struct drm_sched_rq *rq;
544
545 /* single possible engine and already selected */
546 if (!entity->sched_list)
547 return;
548
549 /* queue non-empty, stay on the same engine */
550 if (spsc_queue_count(&entity->job_queue))
551 return;
552
553 /*
554 * Only when the queue is empty are we guaranteed that the scheduler
555 * thread cannot change ->last_scheduled. To enforce ordering we need
556 * a read barrier here. See drm_sched_entity_pop_job() for the other
557 * side.
558 */
559 smp_rmb();
560
561 fence = rcu_dereference_check(entity->last_scheduled, true);
562
563 /* stay on the same engine if the previous job hasn't finished */
564 if (fence && !dma_fence_is_signaled(fence))
565 return;
566
567 spin_lock(&entity->lock);
568 sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
569 rq = sched ? sched->sched_rq[entity->priority] : NULL;
570 if (rq != entity->rq) {
571 drm_sched_rq_remove_entity(entity->rq, entity);
572 entity->rq = rq;
573 }
574 spin_unlock(&entity->lock);
575
576 if (entity->num_sched_list == 1)
577 entity->sched_list = NULL;
578}
579
580/**
581 * drm_sched_entity_push_job - Submit a job to the entity's job queue
582 * @sched_job: job to submit
583 *
584 * Note: To guarantee that the order of insertion to queue matches the job's
585 * fence sequence number this function should be called with drm_sched_job_arm()
586 * under common lock for the struct drm_sched_entity that was set up for
587 * @sched_job in drm_sched_job_init().
588 */
589void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
590{
591 struct drm_sched_entity *entity = sched_job->entity;
592 bool first;
593 ktime_t submit_ts;
594
595 trace_drm_sched_job(sched_job, entity);
596 atomic_inc(entity->rq->sched->score);
597 WRITE_ONCE(entity->last_user, current->group_leader);
598
599 /*
600 * After the sched_job is pushed into the entity queue, it may be
601 * completed and freed up at any time. We can no longer access it.
602 * Make sure to set the submit_ts first, to avoid a race.
603 */
604 sched_job->submit_ts = submit_ts = ktime_get();
605 first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
606
607 /* first job wakes up scheduler */
608 if (first) {
609 struct drm_gpu_scheduler *sched;
610 struct drm_sched_rq *rq;
611
612 /* Add the entity to the run queue */
613 spin_lock(&entity->lock);
614 if (entity->stopped) {
615 spin_unlock(&entity->lock);
616
617 DRM_ERROR("Trying to push to a killed entity\n");
618 return;
619 }
620
621 rq = entity->rq;
622 sched = rq->sched;
623
624 spin_lock(&rq->lock);
625 drm_sched_rq_add_entity(rq, entity);
626
627 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
628 drm_sched_rq_update_fifo_locked(entity, rq, submit_ts);
629
630 spin_unlock(&rq->lock);
631 spin_unlock(&entity->lock);
632
633 drm_sched_wakeup(sched);
634 }
635}
636EXPORT_SYMBOL(drm_sched_entity_push_job);
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/slab.h>
26#include <linux/completion.h>
27
28#include <drm/drm_print.h>
29#include <drm/gpu_scheduler.h>
30
31#include "gpu_scheduler_trace.h"
32
33#define to_drm_sched_job(sched_job) \
34 container_of((sched_job), struct drm_sched_job, queue_node)
35
36/**
37 * drm_sched_entity_init - Init a context entity used by scheduler when
38 * submit to HW ring.
39 *
40 * @entity: scheduler entity to init
41 * @priority: priority of the entity
42 * @sched_list: the list of drm scheds on which jobs from this
43 * entity can be submitted
44 * @num_sched_list: number of drm sched in sched_list
45 * @guilty: atomic_t set to 1 when a job on this queue
46 * is found to be guilty causing a timeout
47 *
48 * Note that the &sched_list must have at least one element to schedule the entity.
49 *
50 * For changing @priority later on at runtime see
51 * drm_sched_entity_set_priority(). For changing the set of schedulers
52 * @sched_list at runtime see drm_sched_entity_modify_sched().
53 *
54 * An entity is cleaned up by callind drm_sched_entity_fini(). See also
55 * drm_sched_entity_destroy().
56 *
57 * Returns 0 on success or a negative error code on failure.
58 */
59int drm_sched_entity_init(struct drm_sched_entity *entity,
60 enum drm_sched_priority priority,
61 struct drm_gpu_scheduler **sched_list,
62 unsigned int num_sched_list,
63 atomic_t *guilty)
64{
65 if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0])))
66 return -EINVAL;
67
68 memset(entity, 0, sizeof(struct drm_sched_entity));
69 INIT_LIST_HEAD(&entity->list);
70 entity->rq = NULL;
71 entity->guilty = guilty;
72 entity->num_sched_list = num_sched_list;
73 entity->priority = priority;
74 entity->sched_list = num_sched_list > 1 ? sched_list : NULL;
75 entity->last_scheduled = NULL;
76 RB_CLEAR_NODE(&entity->rb_tree_node);
77
78 if(num_sched_list)
79 entity->rq = &sched_list[0]->sched_rq[entity->priority];
80
81 init_completion(&entity->entity_idle);
82
83 /* We start in an idle state. */
84 complete_all(&entity->entity_idle);
85
86 spin_lock_init(&entity->rq_lock);
87 spsc_queue_init(&entity->job_queue);
88
89 atomic_set(&entity->fence_seq, 0);
90 entity->fence_context = dma_fence_context_alloc(2);
91
92 return 0;
93}
94EXPORT_SYMBOL(drm_sched_entity_init);
95
96/**
97 * drm_sched_entity_modify_sched - Modify sched of an entity
98 * @entity: scheduler entity to init
99 * @sched_list: the list of new drm scheds which will replace
100 * existing entity->sched_list
101 * @num_sched_list: number of drm sched in sched_list
102 *
103 * Note that this must be called under the same common lock for @entity as
104 * drm_sched_job_arm() and drm_sched_entity_push_job(), or the driver needs to
105 * guarantee through some other means that this is never called while new jobs
106 * can be pushed to @entity.
107 */
108void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
109 struct drm_gpu_scheduler **sched_list,
110 unsigned int num_sched_list)
111{
112 WARN_ON(!num_sched_list || !sched_list);
113
114 entity->sched_list = sched_list;
115 entity->num_sched_list = num_sched_list;
116}
117EXPORT_SYMBOL(drm_sched_entity_modify_sched);
118
119static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
120{
121 rmb(); /* for list_empty to work without lock */
122
123 if (list_empty(&entity->list) ||
124 spsc_queue_count(&entity->job_queue) == 0 ||
125 entity->stopped)
126 return true;
127
128 return false;
129}
130
131/* Return true if entity could provide a job. */
132bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
133{
134 if (spsc_queue_peek(&entity->job_queue) == NULL)
135 return false;
136
137 if (READ_ONCE(entity->dependency))
138 return false;
139
140 return true;
141}
142
143static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)
144{
145 struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
146
147 drm_sched_fence_finished(job->s_fence);
148 WARN_ON(job->s_fence->parent);
149 job->sched->ops->free_job(job);
150}
151
152/* Signal the scheduler finished fence when the entity in question is killed. */
153static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
154 struct dma_fence_cb *cb)
155{
156 struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
157 finish_cb);
158 int r;
159
160 dma_fence_put(f);
161
162 /* Wait for all dependencies to avoid data corruptions */
163 while (!xa_empty(&job->dependencies)) {
164 f = xa_erase(&job->dependencies, job->last_dependency++);
165 r = dma_fence_add_callback(f, &job->finish_cb,
166 drm_sched_entity_kill_jobs_cb);
167 if (!r)
168 return;
169
170 dma_fence_put(f);
171 }
172
173 INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work);
174 schedule_work(&job->work);
175}
176
177/* Remove the entity from the scheduler and kill all pending jobs */
178static void drm_sched_entity_kill(struct drm_sched_entity *entity)
179{
180 struct drm_sched_job *job;
181 struct dma_fence *prev;
182
183 if (!entity->rq)
184 return;
185
186 spin_lock(&entity->rq_lock);
187 entity->stopped = true;
188 drm_sched_rq_remove_entity(entity->rq, entity);
189 spin_unlock(&entity->rq_lock);
190
191 /* Make sure this entity is not used by the scheduler at the moment */
192 wait_for_completion(&entity->entity_idle);
193
194 prev = dma_fence_get(entity->last_scheduled);
195 while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
196 struct drm_sched_fence *s_fence = job->s_fence;
197
198 dma_fence_set_error(&s_fence->finished, -ESRCH);
199
200 dma_fence_get(&s_fence->finished);
201 if (!prev || dma_fence_add_callback(prev, &job->finish_cb,
202 drm_sched_entity_kill_jobs_cb))
203 drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
204
205 prev = &s_fence->finished;
206 }
207 dma_fence_put(prev);
208}
209
210/**
211 * drm_sched_entity_flush - Flush a context entity
212 *
213 * @entity: scheduler entity
214 * @timeout: time to wait in for Q to become empty in jiffies.
215 *
216 * Splitting drm_sched_entity_fini() into two functions, The first one does the
217 * waiting, removes the entity from the runqueue and returns an error when the
218 * process was killed.
219 *
220 * Returns the remaining time in jiffies left from the input timeout
221 */
222long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
223{
224 struct drm_gpu_scheduler *sched;
225 struct task_struct *last_user;
226 long ret = timeout;
227
228 if (!entity->rq)
229 return 0;
230
231 sched = entity->rq->sched;
232 /**
233 * The client will not queue more IBs during this fini, consume existing
234 * queued IBs or discard them on SIGKILL
235 */
236 if (current->flags & PF_EXITING) {
237 if (timeout)
238 ret = wait_event_timeout(
239 sched->job_scheduled,
240 drm_sched_entity_is_idle(entity),
241 timeout);
242 } else {
243 wait_event_killable(sched->job_scheduled,
244 drm_sched_entity_is_idle(entity));
245 }
246
247 /* For killed process disable any more IBs enqueue right now */
248 last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
249 if ((!last_user || last_user == current->group_leader) &&
250 (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))
251 drm_sched_entity_kill(entity);
252
253 return ret;
254}
255EXPORT_SYMBOL(drm_sched_entity_flush);
256
257/**
258 * drm_sched_entity_fini - Destroy a context entity
259 *
260 * @entity: scheduler entity
261 *
262 * Cleanups up @entity which has been initialized by drm_sched_entity_init().
263 *
264 * If there are potentially job still in flight or getting newly queued
265 * drm_sched_entity_flush() must be called first. This function then goes over
266 * the entity and signals all jobs with an error code if the process was killed.
267 */
268void drm_sched_entity_fini(struct drm_sched_entity *entity)
269{
270 /*
271 * If consumption of existing IBs wasn't completed. Forcefully remove
272 * them here. Also makes sure that the scheduler won't touch this entity
273 * any more.
274 */
275 drm_sched_entity_kill(entity);
276
277 if (entity->dependency) {
278 dma_fence_remove_callback(entity->dependency, &entity->cb);
279 dma_fence_put(entity->dependency);
280 entity->dependency = NULL;
281 }
282
283 dma_fence_put(entity->last_scheduled);
284 entity->last_scheduled = NULL;
285}
286EXPORT_SYMBOL(drm_sched_entity_fini);
287
288/**
289 * drm_sched_entity_destroy - Destroy a context entity
290 * @entity: scheduler entity
291 *
292 * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a
293 * convenience wrapper.
294 */
295void drm_sched_entity_destroy(struct drm_sched_entity *entity)
296{
297 drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
298 drm_sched_entity_fini(entity);
299}
300EXPORT_SYMBOL(drm_sched_entity_destroy);
301
302/* drm_sched_entity_clear_dep - callback to clear the entities dependency */
303static void drm_sched_entity_clear_dep(struct dma_fence *f,
304 struct dma_fence_cb *cb)
305{
306 struct drm_sched_entity *entity =
307 container_of(cb, struct drm_sched_entity, cb);
308
309 entity->dependency = NULL;
310 dma_fence_put(f);
311}
312
313/*
314 * drm_sched_entity_clear_dep - callback to clear the entities dependency and
315 * wake up scheduler
316 */
317static void drm_sched_entity_wakeup(struct dma_fence *f,
318 struct dma_fence_cb *cb)
319{
320 struct drm_sched_entity *entity =
321 container_of(cb, struct drm_sched_entity, cb);
322
323 drm_sched_entity_clear_dep(f, cb);
324 drm_sched_wakeup(entity->rq->sched);
325}
326
327/**
328 * drm_sched_entity_set_priority - Sets priority of the entity
329 *
330 * @entity: scheduler entity
331 * @priority: scheduler priority
332 *
333 * Update the priority of runqueus used for the entity.
334 */
335void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
336 enum drm_sched_priority priority)
337{
338 spin_lock(&entity->rq_lock);
339 entity->priority = priority;
340 spin_unlock(&entity->rq_lock);
341}
342EXPORT_SYMBOL(drm_sched_entity_set_priority);
343
344/*
345 * Add a callback to the current dependency of the entity to wake up the
346 * scheduler when the entity becomes available.
347 */
348static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
349{
350 struct drm_gpu_scheduler *sched = entity->rq->sched;
351 struct dma_fence *fence = entity->dependency;
352 struct drm_sched_fence *s_fence;
353
354 if (fence->context == entity->fence_context ||
355 fence->context == entity->fence_context + 1) {
356 /*
357 * Fence is a scheduled/finished fence from a job
358 * which belongs to the same entity, we can ignore
359 * fences from ourself
360 */
361 dma_fence_put(entity->dependency);
362 return false;
363 }
364
365 s_fence = to_drm_sched_fence(fence);
366 if (s_fence && s_fence->sched == sched &&
367 !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {
368
369 /*
370 * Fence is from the same scheduler, only need to wait for
371 * it to be scheduled
372 */
373 fence = dma_fence_get(&s_fence->scheduled);
374 dma_fence_put(entity->dependency);
375 entity->dependency = fence;
376 if (!dma_fence_add_callback(fence, &entity->cb,
377 drm_sched_entity_clear_dep))
378 return true;
379
380 /* Ignore it when it is already scheduled */
381 dma_fence_put(fence);
382 return false;
383 }
384
385 if (!dma_fence_add_callback(entity->dependency, &entity->cb,
386 drm_sched_entity_wakeup))
387 return true;
388
389 dma_fence_put(entity->dependency);
390 return false;
391}
392
393static struct dma_fence *
394drm_sched_job_dependency(struct drm_sched_job *job,
395 struct drm_sched_entity *entity)
396{
397 if (!xa_empty(&job->dependencies))
398 return xa_erase(&job->dependencies, job->last_dependency++);
399
400 if (job->sched->ops->prepare_job)
401 return job->sched->ops->prepare_job(job, entity);
402
403 return NULL;
404}
405
406struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
407{
408 struct drm_sched_job *sched_job;
409
410 sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
411 if (!sched_job)
412 return NULL;
413
414 while ((entity->dependency =
415 drm_sched_job_dependency(sched_job, entity))) {
416 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
417
418 if (drm_sched_entity_add_dependency_cb(entity))
419 return NULL;
420 }
421
422 /* skip jobs from entity that marked guilty */
423 if (entity->guilty && atomic_read(entity->guilty))
424 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
425
426 dma_fence_put(entity->last_scheduled);
427
428 entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
429
430 /*
431 * If the queue is empty we allow drm_sched_entity_select_rq() to
432 * locklessly access ->last_scheduled. This only works if we set the
433 * pointer before we dequeue and if we a write barrier here.
434 */
435 smp_wmb();
436
437 spsc_queue_pop(&entity->job_queue);
438
439 /*
440 * Update the entity's location in the min heap according to
441 * the timestamp of the next job, if any.
442 */
443 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {
444 struct drm_sched_job *next;
445
446 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
447 if (next)
448 drm_sched_rq_update_fifo(entity, next->submit_ts);
449 }
450
451 return sched_job;
452}
453
454void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
455{
456 struct dma_fence *fence;
457 struct drm_gpu_scheduler *sched;
458 struct drm_sched_rq *rq;
459
460 /* single possible engine and already selected */
461 if (!entity->sched_list)
462 return;
463
464 /* queue non-empty, stay on the same engine */
465 if (spsc_queue_count(&entity->job_queue))
466 return;
467
468 /*
469 * Only when the queue is empty are we guaranteed that the scheduler
470 * thread cannot change ->last_scheduled. To enforce ordering we need
471 * a read barrier here. See drm_sched_entity_pop_job() for the other
472 * side.
473 */
474 smp_rmb();
475
476 fence = entity->last_scheduled;
477
478 /* stay on the same engine if the previous job hasn't finished */
479 if (fence && !dma_fence_is_signaled(fence))
480 return;
481
482 spin_lock(&entity->rq_lock);
483 sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
484 rq = sched ? &sched->sched_rq[entity->priority] : NULL;
485 if (rq != entity->rq) {
486 drm_sched_rq_remove_entity(entity->rq, entity);
487 entity->rq = rq;
488 }
489 spin_unlock(&entity->rq_lock);
490
491 if (entity->num_sched_list == 1)
492 entity->sched_list = NULL;
493}
494
495/**
496 * drm_sched_entity_push_job - Submit a job to the entity's job queue
497 * @sched_job: job to submit
498 *
499 * Note: To guarantee that the order of insertion to queue matches the job's
500 * fence sequence number this function should be called with drm_sched_job_arm()
501 * under common lock for the struct drm_sched_entity that was set up for
502 * @sched_job in drm_sched_job_init().
503 *
504 * Returns 0 for success, negative error code otherwise.
505 */
506void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
507{
508 struct drm_sched_entity *entity = sched_job->entity;
509 bool first;
510
511 trace_drm_sched_job(sched_job, entity);
512 atomic_inc(entity->rq->sched->score);
513 WRITE_ONCE(entity->last_user, current->group_leader);
514 first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
515 sched_job->submit_ts = ktime_get();
516
517 /* first job wakes up scheduler */
518 if (first) {
519 /* Add the entity to the run queue */
520 spin_lock(&entity->rq_lock);
521 if (entity->stopped) {
522 spin_unlock(&entity->rq_lock);
523
524 DRM_ERROR("Trying to push to a killed entity\n");
525 return;
526 }
527
528 drm_sched_rq_add_entity(entity->rq, entity);
529 spin_unlock(&entity->rq_lock);
530
531 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
532 drm_sched_rq_update_fifo(entity, sched_job->submit_ts);
533
534 drm_sched_wakeup(entity->rq->sched);
535 }
536}
537EXPORT_SYMBOL(drm_sched_entity_push_job);