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: the sched_list should have at least one element to schedule
49  *       the entity
50  *
51  * Returns 0 on success or a negative error code on failure.
52  */
53 int drm_sched_entity_init(struct drm_sched_entity *entity,
54 			  enum drm_sched_priority priority,
55 			  struct drm_gpu_scheduler **sched_list,
56 			  unsigned int num_sched_list,
57 			  atomic_t *guilty)
58 {
59 	if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0])))
60 		return -EINVAL;
61 
62 	memset(entity, 0, sizeof(struct drm_sched_entity));
63 	INIT_LIST_HEAD(&entity->list);
64 	entity->rq = NULL;
65 	entity->guilty = guilty;
66 	entity->num_sched_list = num_sched_list;
67 	entity->priority = priority;
68 	entity->sched_list = num_sched_list > 1 ? sched_list : NULL;
69 	entity->last_scheduled = NULL;
70 
71 	if(num_sched_list)
72 		entity->rq = &sched_list[0]->sched_rq[entity->priority];
73 
74 	init_completion(&entity->entity_idle);
75 
76 	spin_lock_init(&entity->rq_lock);
77 	spsc_queue_init(&entity->job_queue);
78 
79 	atomic_set(&entity->fence_seq, 0);
80 	entity->fence_context = dma_fence_context_alloc(2);
81 
82 	return 0;
83 }
84 EXPORT_SYMBOL(drm_sched_entity_init);
85 
86 /**
87  * drm_sched_entity_modify_sched - Modify sched of an entity
88  * @entity: scheduler entity to init
89  * @sched_list: the list of new drm scheds which will replace
90  *		 existing entity->sched_list
91  * @num_sched_list: number of drm sched in sched_list
92  */
93 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
94 				    struct drm_gpu_scheduler **sched_list,
95 				    unsigned int num_sched_list)
96 {
97 	WARN_ON(!num_sched_list || !sched_list);
98 
99 	entity->sched_list = sched_list;
100 	entity->num_sched_list = num_sched_list;
101 }
102 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
103 
104 /**
105  * drm_sched_entity_is_idle - Check if entity is idle
106  *
107  * @entity: scheduler entity
108  *
109  * Returns true if the entity does not have any unscheduled jobs.
110  */
111 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
112 {
113 	rmb(); /* for list_empty to work without lock */
114 
115 	if (list_empty(&entity->list) ||
116 	    spsc_queue_count(&entity->job_queue) == 0)
117 		return true;
118 
119 	return false;
120 }
121 
122 /**
123  * drm_sched_entity_is_ready - Check if entity is ready
124  *
125  * @entity: scheduler entity
126  *
127  * Return true if entity could provide a job.
128  */
129 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
130 {
131 	if (spsc_queue_peek(&entity->job_queue) == NULL)
132 		return false;
133 
134 	if (READ_ONCE(entity->dependency))
135 		return false;
136 
137 	return true;
138 }
139 
140 /**
141  * drm_sched_entity_flush - Flush a context entity
142  *
143  * @entity: scheduler entity
144  * @timeout: time to wait in for Q to become empty in jiffies.
145  *
146  * Splitting drm_sched_entity_fini() into two functions, The first one does the
147  * waiting, removes the entity from the runqueue and returns an error when the
148  * process was killed.
149  *
150  * Returns the remaining time in jiffies left from the input timeout
151  */
152 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
153 {
154 	struct drm_gpu_scheduler *sched;
155 	struct task_struct *last_user;
156 	long ret = timeout;
157 
158 	if (!entity->rq)
159 		return 0;
160 
161 	sched = entity->rq->sched;
162 	/**
163 	 * The client will not queue more IBs during this fini, consume existing
164 	 * queued IBs or discard them on SIGKILL
165 	 */
166 	if (current->flags & PF_EXITING) {
167 		if (timeout)
168 			ret = wait_event_timeout(
169 					sched->job_scheduled,
170 					drm_sched_entity_is_idle(entity),
171 					timeout);
172 	} else {
173 		wait_event_killable(sched->job_scheduled,
174 				    drm_sched_entity_is_idle(entity));
175 	}
176 
177 	/* For killed process disable any more IBs enqueue right now */
178 	last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
179 	if ((!last_user || last_user == current->group_leader) &&
180 	    (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
181 		spin_lock(&entity->rq_lock);
182 		entity->stopped = true;
183 		drm_sched_rq_remove_entity(entity->rq, entity);
184 		spin_unlock(&entity->rq_lock);
185 	}
186 
187 	return ret;
188 }
189 EXPORT_SYMBOL(drm_sched_entity_flush);
190 
191 /**
192  * drm_sched_entity_kill_jobs - helper for drm_sched_entity_kill_jobs
193  *
194  * @f: signaled fence
195  * @cb: our callback structure
196  *
197  * Signal the scheduler finished fence when the entity in question is killed.
198  */
199 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
200 					  struct dma_fence_cb *cb)
201 {
202 	struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
203 						 finish_cb);
204 
205 	drm_sched_fence_finished(job->s_fence);
206 	WARN_ON(job->s_fence->parent);
207 	job->sched->ops->free_job(job);
208 }
209 
210 /**
211  * drm_sched_entity_kill_jobs - Make sure all remaining jobs are killed
212  *
213  * @entity: entity which is cleaned up
214  *
215  * Makes sure that all remaining jobs in an entity are killed before it is
216  * destroyed.
217  */
218 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
219 {
220 	struct drm_sched_job *job;
221 	int r;
222 
223 	while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
224 		struct drm_sched_fence *s_fence = job->s_fence;
225 
226 		drm_sched_fence_scheduled(s_fence);
227 		dma_fence_set_error(&s_fence->finished, -ESRCH);
228 
229 		/*
230 		 * When pipe is hanged by older entity, new entity might
231 		 * not even have chance to submit it's first job to HW
232 		 * and so entity->last_scheduled will remain NULL
233 		 */
234 		if (!entity->last_scheduled) {
235 			drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
236 			continue;
237 		}
238 
239 		r = dma_fence_add_callback(entity->last_scheduled,
240 					   &job->finish_cb,
241 					   drm_sched_entity_kill_jobs_cb);
242 		if (r == -ENOENT)
243 			drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
244 		else if (r)
245 			DRM_ERROR("fence add callback failed (%d)\n", r);
246 	}
247 }
248 
249 /**
250  * drm_sched_entity_cleanup - Destroy a context entity
251  *
252  * @entity: scheduler entity
253  *
254  * This should be called after @drm_sched_entity_do_release. It goes over the
255  * entity and signals all jobs with an error code if the process was killed.
256  *
257  */
258 void drm_sched_entity_fini(struct drm_sched_entity *entity)
259 {
260 	struct drm_gpu_scheduler *sched = NULL;
261 
262 	if (entity->rq) {
263 		sched = entity->rq->sched;
264 		drm_sched_rq_remove_entity(entity->rq, entity);
265 	}
266 
267 	/* Consumption of existing IBs wasn't completed. Forcefully
268 	 * remove them here.
269 	 */
270 	if (spsc_queue_count(&entity->job_queue)) {
271 		if (sched) {
272 			/*
273 			 * Wait for thread to idle to make sure it isn't processing
274 			 * this entity.
275 			 */
276 			wait_for_completion(&entity->entity_idle);
277 
278 		}
279 		if (entity->dependency) {
280 			dma_fence_remove_callback(entity->dependency,
281 						  &entity->cb);
282 			dma_fence_put(entity->dependency);
283 			entity->dependency = NULL;
284 		}
285 
286 		drm_sched_entity_kill_jobs(entity);
287 	}
288 
289 	dma_fence_put(entity->last_scheduled);
290 	entity->last_scheduled = NULL;
291 }
292 EXPORT_SYMBOL(drm_sched_entity_fini);
293 
294 /**
295  * drm_sched_entity_fini - Destroy a context entity
296  *
297  * @entity: scheduler entity
298  *
299  * Calls drm_sched_entity_do_release() and drm_sched_entity_cleanup()
300  */
301 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
302 {
303 	drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
304 	drm_sched_entity_fini(entity);
305 }
306 EXPORT_SYMBOL(drm_sched_entity_destroy);
307 
308 /**
309  * drm_sched_entity_clear_dep - callback to clear the entities dependency
310  */
311 static void drm_sched_entity_clear_dep(struct dma_fence *f,
312 				       struct dma_fence_cb *cb)
313 {
314 	struct drm_sched_entity *entity =
315 		container_of(cb, struct drm_sched_entity, cb);
316 
317 	entity->dependency = NULL;
318 	dma_fence_put(f);
319 }
320 
321 /**
322  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
323  * wake up scheduler
324  */
325 static void drm_sched_entity_wakeup(struct dma_fence *f,
326 				    struct dma_fence_cb *cb)
327 {
328 	struct drm_sched_entity *entity =
329 		container_of(cb, struct drm_sched_entity, cb);
330 
331 	drm_sched_entity_clear_dep(f, cb);
332 	drm_sched_wakeup(entity->rq->sched);
333 }
334 
335 /**
336  * drm_sched_entity_set_priority - Sets priority of the entity
337  *
338  * @entity: scheduler entity
339  * @priority: scheduler priority
340  *
341  * Update the priority of runqueus used for the entity.
342  */
343 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
344 				   enum drm_sched_priority priority)
345 {
346 	spin_lock(&entity->rq_lock);
347 	entity->priority = priority;
348 	spin_unlock(&entity->rq_lock);
349 }
350 EXPORT_SYMBOL(drm_sched_entity_set_priority);
351 
352 /**
353  * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
354  *
355  * @entity: entity with dependency
356  *
357  * Add a callback to the current dependency of the entity to wake up the
358  * scheduler when the entity becomes available.
359  */
360 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
361 {
362 	struct drm_gpu_scheduler *sched = entity->rq->sched;
363 	struct dma_fence *fence = entity->dependency;
364 	struct drm_sched_fence *s_fence;
365 
366 	if (fence->context == entity->fence_context ||
367 	    fence->context == entity->fence_context + 1) {
368 		/*
369 		 * Fence is a scheduled/finished fence from a job
370 		 * which belongs to the same entity, we can ignore
371 		 * fences from ourself
372 		 */
373 		dma_fence_put(entity->dependency);
374 		return false;
375 	}
376 
377 	s_fence = to_drm_sched_fence(fence);
378 	if (s_fence && s_fence->sched == sched) {
379 
380 		/*
381 		 * Fence is from the same scheduler, only need to wait for
382 		 * it to be scheduled
383 		 */
384 		fence = dma_fence_get(&s_fence->scheduled);
385 		dma_fence_put(entity->dependency);
386 		entity->dependency = fence;
387 		if (!dma_fence_add_callback(fence, &entity->cb,
388 					    drm_sched_entity_clear_dep))
389 			return true;
390 
391 		/* Ignore it when it is already scheduled */
392 		dma_fence_put(fence);
393 		return false;
394 	}
395 
396 	if (!dma_fence_add_callback(entity->dependency, &entity->cb,
397 				    drm_sched_entity_wakeup))
398 		return true;
399 
400 	dma_fence_put(entity->dependency);
401 	return false;
402 }
403 
404 /**
405  * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
406  *
407  * @entity: entity to get the job from
408  *
409  * Process all dependencies and try to get one job from the entities queue.
410  */
411 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
412 {
413 	struct drm_gpu_scheduler *sched = entity->rq->sched;
414 	struct drm_sched_job *sched_job;
415 
416 	sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
417 	if (!sched_job)
418 		return NULL;
419 
420 	while ((entity->dependency =
421 			sched->ops->dependency(sched_job, entity))) {
422 		trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
423 
424 		if (drm_sched_entity_add_dependency_cb(entity))
425 			return NULL;
426 	}
427 
428 	/* skip jobs from entity that marked guilty */
429 	if (entity->guilty && atomic_read(entity->guilty))
430 		dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
431 
432 	dma_fence_put(entity->last_scheduled);
433 	entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
434 
435 	spsc_queue_pop(&entity->job_queue);
436 	return sched_job;
437 }
438 
439 /**
440  * drm_sched_entity_select_rq - select a new rq for the entity
441  *
442  * @entity: scheduler entity
443  *
444  * Check all prerequisites and select a new rq for the entity for load
445  * balancing.
446  */
447 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
448 {
449 	struct dma_fence *fence;
450 	struct drm_gpu_scheduler *sched;
451 	struct drm_sched_rq *rq;
452 
453 	if (spsc_queue_count(&entity->job_queue) || entity->num_sched_list <= 1)
454 		return;
455 
456 	fence = READ_ONCE(entity->last_scheduled);
457 	if (fence && !dma_fence_is_signaled(fence))
458 		return;
459 
460 	spin_lock(&entity->rq_lock);
461 	sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
462 	rq = sched ? &sched->sched_rq[entity->priority] : NULL;
463 	if (rq != entity->rq) {
464 		drm_sched_rq_remove_entity(entity->rq, entity);
465 		entity->rq = rq;
466 	}
467 
468 	spin_unlock(&entity->rq_lock);
469 }
470 
471 /**
472  * drm_sched_entity_push_job - Submit a job to the entity's job queue
473  *
474  * @sched_job: job to submit
475  * @entity: scheduler entity
476  *
477  * Note: To guarantee that the order of insertion to queue matches
478  * the job's fence sequence number this function should be
479  * called with drm_sched_job_init under common lock.
480  *
481  * Returns 0 for success, negative error code otherwise.
482  */
483 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
484 			       struct drm_sched_entity *entity)
485 {
486 	bool first;
487 
488 	trace_drm_sched_job(sched_job, entity);
489 	atomic_inc(&entity->rq->sched->score);
490 	WRITE_ONCE(entity->last_user, current->group_leader);
491 	first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
492 
493 	/* first job wakes up scheduler */
494 	if (first) {
495 		/* Add the entity to the run queue */
496 		spin_lock(&entity->rq_lock);
497 		if (entity->stopped) {
498 			spin_unlock(&entity->rq_lock);
499 
500 			DRM_ERROR("Trying to push to a killed entity\n");
501 			return;
502 		}
503 		drm_sched_rq_add_entity(entity->rq, entity);
504 		spin_unlock(&entity->rq_lock);
505 		drm_sched_wakeup(entity->rq->sched);
506 	}
507 }
508 EXPORT_SYMBOL(drm_sched_entity_push_job);
509