sched_entity.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/export.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 "sched_internal.h"
 32
 33#include "gpu_scheduler_trace.h"
 34
 35/**
 36 * drm_sched_entity_init - Init a context entity used by scheduler when
 37 * submit to HW ring.
 38 *
 39 * @entity: scheduler entity to init
 40 * @priority: priority of the entity
 41 * @sched_list: the list of drm scheds on which jobs from this
 42 *           entity can be submitted
 43 * @num_sched_list: number of drm sched in sched_list
 44 * @guilty: atomic_t set to 1 when a job on this queue
 45 *          is found to be guilty causing a timeout
 46 *
 47 * Note that the &sched_list must have at least one element to schedule the entity.
 48 *
 49 * For changing @priority later on at runtime see
 50 * drm_sched_entity_set_priority(). For changing the set of schedulers
 51 * @sched_list at runtime see drm_sched_entity_modify_sched().
 52 *
 53 * An entity is cleaned up by calling drm_sched_entity_fini(). See also
 54 * drm_sched_entity_destroy().
 55 *
 56 * Returns 0 on success or a negative error code on failure.
 57 */
 58int drm_sched_entity_init(struct drm_sched_entity *entity,
 59			  enum drm_sched_priority priority,
 60			  struct drm_gpu_scheduler **sched_list,
 61			  unsigned int num_sched_list,
 62			  atomic_t *guilty)
 63{
 64	if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0])))
 65		return -EINVAL;
 66
 67	memset(entity, 0, sizeof(struct drm_sched_entity));
 68	INIT_LIST_HEAD(&entity->list);
 69	entity->rq = NULL;
 70	entity->guilty = guilty;
 71	entity->num_sched_list = num_sched_list;
 72	entity->priority = priority;
 73	entity->last_user = current->group_leader;
 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			dev_err(sched_list[0]->dev, "entity has 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/**
156 * drm_sched_entity_error - return error of last scheduled job
157 * @entity: scheduler entity to check
158 *
159 * Opportunistically return the error of the last scheduled job. Result can
160 * change any time when new jobs are pushed to the hw.
161 */
162int drm_sched_entity_error(struct drm_sched_entity *entity)
163{
164	struct dma_fence *fence;
165	int r;
166
167	rcu_read_lock();
168	fence = rcu_dereference(entity->last_scheduled);
169	r = fence ? fence->error : 0;
170	rcu_read_unlock();
171
172	return r;
173}
174EXPORT_SYMBOL(drm_sched_entity_error);
175
176static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
177					  struct dma_fence_cb *cb);
178
179static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)
180{
181	struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
182	struct dma_fence *f;
183	unsigned long index;
184
185	/* Wait for all dependencies to avoid data corruptions */
186	xa_for_each(&job->dependencies, index, f) {
187		struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
188
189		if (s_fence && f == &s_fence->scheduled) {
190			/* The dependencies array had a reference on the scheduled
191			 * fence, and the finished fence refcount might have
192			 * dropped to zero. Use dma_fence_get_rcu() so we get
193			 * a NULL fence in that case.
194			 */
195			f = dma_fence_get_rcu(&s_fence->finished);
196
197			/* Now that we have a reference on the finished fence,
198			 * we can release the reference the dependencies array
199			 * had on the scheduled fence.
200			 */
201			dma_fence_put(&s_fence->scheduled);
202		}
203
204		xa_erase(&job->dependencies, index);
205		if (f && !dma_fence_add_callback(f, &job->finish_cb,
206						 drm_sched_entity_kill_jobs_cb))
207			return;
208
209		dma_fence_put(f);
210	}
211
212	drm_sched_fence_scheduled(job->s_fence, NULL);
213	drm_sched_fence_finished(job->s_fence, -ESRCH);
214	WARN_ON(job->s_fence->parent);
215	job->sched->ops->free_job(job);
216}
217
218/* Signal the scheduler finished fence when the entity in question is killed. */
219static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
220					  struct dma_fence_cb *cb)
221{
222	struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
223						 finish_cb);
224
225	dma_fence_put(f);
226
227	INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work);
228	schedule_work(&job->work);
229}
230
231/* Remove the entity from the scheduler and kill all pending jobs */
232static void drm_sched_entity_kill(struct drm_sched_entity *entity)
233{
234	struct drm_sched_job *job;
235	struct dma_fence *prev;
236
237	if (!entity->rq)
238		return;
239
240	spin_lock(&entity->lock);
241	entity->stopped = true;
242	drm_sched_rq_remove_entity(entity->rq, entity);
243	spin_unlock(&entity->lock);
244
245	/* Make sure this entity is not used by the scheduler at the moment */
246	wait_for_completion(&entity->entity_idle);
247
248	/* The entity is guaranteed to not be used by the scheduler */
249	prev = rcu_dereference_check(entity->last_scheduled, true);
250	dma_fence_get(prev);
251	while ((job = drm_sched_entity_queue_pop(entity))) {
252		struct drm_sched_fence *s_fence = job->s_fence;
253
254		dma_fence_get(&s_fence->finished);
255		if (!prev ||
256		    dma_fence_add_callback(prev, &job->finish_cb,
257					   drm_sched_entity_kill_jobs_cb)) {
258			/*
259			 * Adding callback above failed.
260			 * dma_fence_put() checks for NULL.
261			 */
262			dma_fence_put(prev);
263			drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
264		}
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 jobs during this fini - consume
295	 * existing queued ones, 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 a killed process disallow further enqueueing of jobs. */
309	last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
310	if (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 jobs wasn't completed forcefully remove
333	 * them. Also makes sure that the scheduler won't touch this entity any
334	 * 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/*
364 * drm_sched_entity_wakeup - callback to clear the entity's dependency and
365 * wake up the scheduler
366 */
367static void drm_sched_entity_wakeup(struct dma_fence *f,
368				    struct dma_fence_cb *cb)
369{
370	struct drm_sched_entity *entity =
371		container_of(cb, struct drm_sched_entity, cb);
372
373	entity->dependency = NULL;
374	dma_fence_put(f);
375	drm_sched_wakeup(entity->rq->sched);
376}
377
378/**
379 * drm_sched_entity_set_priority - Sets priority of the entity
380 *
381 * @entity: scheduler entity
382 * @priority: scheduler priority
383 *
384 * Update the priority of runqueues used for the entity.
385 */
386void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
387				   enum drm_sched_priority priority)
388{
389	spin_lock(&entity->lock);
390	entity->priority = priority;
391	spin_unlock(&entity->lock);
392}
393EXPORT_SYMBOL(drm_sched_entity_set_priority);
394
395/*
396 * Add a callback to the current dependency of the entity to wake up the
397 * scheduler when the entity becomes available.
398 */
399static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity,
400					       struct drm_sched_job *sched_job)
401{
402	struct drm_gpu_scheduler *sched = entity->rq->sched;
403	struct dma_fence *fence = entity->dependency;
404	struct drm_sched_fence *s_fence;
405
406	if (fence->context == entity->fence_context ||
407	    fence->context == entity->fence_context + 1) {
408		/*
409		 * Fence is a scheduled/finished fence from a job
410		 * which belongs to the same entity, we can ignore
411		 * fences from ourself
412		 */
413		dma_fence_put(entity->dependency);
414		return false;
415	}
416
417	s_fence = to_drm_sched_fence(fence);
418	if (!fence->error && s_fence && s_fence->sched == sched &&
419	    !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {
420
421		/*
422		 * Fence is from the same scheduler, only need to wait for
423		 * it to be scheduled
424		 */
425		fence = dma_fence_get(&s_fence->scheduled);
426		dma_fence_put(entity->dependency);
427		entity->dependency = fence;
428	}
429
430	if (trace_drm_sched_job_unschedulable_enabled() &&
431	    !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &entity->dependency->flags))
432		trace_drm_sched_job_unschedulable(sched_job, entity->dependency);
433
434	if (!dma_fence_add_callback(entity->dependency, &entity->cb,
435				    drm_sched_entity_wakeup))
436		return true;
437
438	dma_fence_put(entity->dependency);
439	return false;
440}
441
442static struct dma_fence *
443drm_sched_job_dependency(struct drm_sched_job *job,
444			 struct drm_sched_entity *entity)
445{
446	struct dma_fence *f;
447
448	/* We keep the fence around, so we can iterate over all dependencies
449	 * in drm_sched_entity_kill_jobs_cb() to ensure all deps are signaled
450	 * before killing the job.
451	 */
452	f = xa_load(&job->dependencies, job->last_dependency);
453	if (f) {
454		job->last_dependency++;
455		return dma_fence_get(f);
456	}
457
458	if (job->sched->ops->prepare_job)
459		return job->sched->ops->prepare_job(job, entity);
460
461	return NULL;
462}
463
464struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
465{
466	struct drm_sched_job *sched_job;
467
468	sched_job = drm_sched_entity_queue_peek(entity);
469	if (!sched_job)
470		return NULL;
471
472	while ((entity->dependency =
473			drm_sched_job_dependency(sched_job, entity))) {
474		if (drm_sched_entity_add_dependency_cb(entity, sched_job))
475			return NULL;
476	}
477
478	/* skip jobs from entity that marked guilty */
479	if (entity->guilty && atomic_read(entity->guilty))
480		dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
481
482	dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
483	rcu_assign_pointer(entity->last_scheduled,
484			   dma_fence_get(&sched_job->s_fence->finished));
485
486	/*
487	 * If the queue is empty we allow drm_sched_entity_select_rq() to
488	 * locklessly access ->last_scheduled. This only works if we set the
489	 * pointer before we dequeue and if we a write barrier here.
490	 */
491	smp_wmb();
492
493	spsc_queue_pop(&entity->job_queue);
494
495	/*
496	 * Update the entity's location in the min heap according to
497	 * the timestamp of the next job, if any.
498	 */
499	if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {
500		struct drm_sched_job *next;
501
502		next = drm_sched_entity_queue_peek(entity);
503		if (next) {
504			struct drm_sched_rq *rq;
505
506			spin_lock(&entity->lock);
507			rq = entity->rq;
508			spin_lock(&rq->lock);
509			drm_sched_rq_update_fifo_locked(entity, rq,
510							next->submit_ts);
511			spin_unlock(&rq->lock);
512			spin_unlock(&entity->lock);
513		}
514	}
515
516	/* Jobs and entities might have different lifecycles. Since we're
517	 * removing the job from the entities queue, set the jobs entity pointer
518	 * to NULL to prevent any future access of the entity through this job.
519	 */
520	sched_job->entity = NULL;
521
522	return sched_job;
523}
524
525void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
526{
527	struct dma_fence *fence;
528	struct drm_gpu_scheduler *sched;
529	struct drm_sched_rq *rq;
530
531	/* single possible engine and already selected */
532	if (!entity->sched_list)
533		return;
534
535	/* queue non-empty, stay on the same engine */
536	if (spsc_queue_count(&entity->job_queue))
537		return;
538
539	/*
540	 * Only when the queue is empty are we guaranteed that
541	 * drm_sched_run_job_work() cannot change entity->last_scheduled. To
542	 * enforce ordering we need a read barrier here. See
543	 * drm_sched_entity_pop_job() for the other side.
544	 */
545	smp_rmb();
546
547	fence = rcu_dereference_check(entity->last_scheduled, true);
548
549	/* stay on the same engine if the previous job hasn't finished */
550	if (fence && !dma_fence_is_signaled(fence))
551		return;
552
553	spin_lock(&entity->lock);
554	sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
555	rq = sched ? sched->sched_rq[entity->priority] : NULL;
556	if (rq != entity->rq) {
557		drm_sched_rq_remove_entity(entity->rq, entity);
558		entity->rq = rq;
559	}
560
561	if (entity->num_sched_list == 1)
562		entity->sched_list = NULL;
563
564	spin_unlock(&entity->lock);
565}
566
567/**
568 * drm_sched_entity_push_job - Submit a job to the entity's job queue
569 * @sched_job: job to submit
570 *
571 * Note: To guarantee that the order of insertion to queue matches the job's
572 * fence sequence number this function should be called with drm_sched_job_arm()
573 * under common lock for the struct drm_sched_entity that was set up for
574 * @sched_job in drm_sched_job_init().
575 */
576void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
577{
578	struct drm_sched_entity *entity = sched_job->entity;
579	bool first;
580	ktime_t submit_ts;
581
582	trace_drm_sched_job_queue(sched_job, entity);
583
584	if (trace_drm_sched_job_add_dep_enabled()) {
585		struct dma_fence *entry;
586		unsigned long index;
587
588		xa_for_each(&sched_job->dependencies, index, entry)
589			trace_drm_sched_job_add_dep(sched_job, entry);
590	}
591	atomic_inc(entity->rq->sched->score);
592	WRITE_ONCE(entity->last_user, current->group_leader);
593
594	/*
595	 * After the sched_job is pushed into the entity queue, it may be
596	 * completed and freed up at any time. We can no longer access it.
597	 * Make sure to set the submit_ts first, to avoid a race.
598	 */
599	sched_job->submit_ts = submit_ts = ktime_get();
600	first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
601
602	/* first job wakes up scheduler */
603	if (first) {
604		struct drm_gpu_scheduler *sched;
605		struct drm_sched_rq *rq;
606
607		/* Add the entity to the run queue */
608		spin_lock(&entity->lock);
609		if (entity->stopped) {
610			spin_unlock(&entity->lock);
611
612			DRM_ERROR("Trying to push to a killed entity\n");
613			return;
614		}
615
616		rq = entity->rq;
617		sched = rq->sched;
618
619		spin_lock(&rq->lock);
620		drm_sched_rq_add_entity(rq, entity);
621
622		if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
623			drm_sched_rq_update_fifo_locked(entity, rq, submit_ts);
624
625		spin_unlock(&rq->lock);
626		spin_unlock(&entity->lock);
627
628		drm_sched_wakeup(sched);
629	}
630}
631EXPORT_SYMBOL(drm_sched_entity_push_job);