2 * Copyright 2015 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <linux/completion.h>
28 #include <drm/drm_print.h>
29 #include <drm/gpu_scheduler.h>
31 #include "gpu_scheduler_trace.h"
33 #define to_drm_sched_job(sched_job) \
34 container_of((sched_job), struct drm_sched_job, queue_node)
37 * drm_sched_entity_init - Init a context entity used by scheduler when
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
48 * Note: the sched_list should have at least one element to schedule
51 * Returns 0 on success or a negative error code on failure.
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,
59 if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0])))
62 memset(entity, 0, sizeof(struct drm_sched_entity));
63 INIT_LIST_HEAD(&entity->list);
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;
72 entity->rq = &sched_list[0]->sched_rq[entity->priority];
74 init_completion(&entity->entity_idle);
76 spin_lock_init(&entity->rq_lock);
77 spsc_queue_init(&entity->job_queue);
79 atomic_set(&entity->fence_seq, 0);
80 entity->fence_context = dma_fence_context_alloc(2);
84 EXPORT_SYMBOL(drm_sched_entity_init);
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
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)
97 WARN_ON(!num_sched_list || !sched_list);
99 entity->sched_list = sched_list;
100 entity->num_sched_list = num_sched_list;
102 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
105 * drm_sched_entity_is_idle - Check if entity is idle
107 * @entity: scheduler entity
109 * Returns true if the entity does not have any unscheduled jobs.
111 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
113 rmb(); /* for list_empty to work without lock */
115 if (list_empty(&entity->list) ||
116 spsc_queue_count(&entity->job_queue) == 0 ||
124 * drm_sched_entity_is_ready - Check if entity is ready
126 * @entity: scheduler entity
128 * Return true if entity could provide a job.
130 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
132 if (spsc_queue_peek(&entity->job_queue) == NULL)
135 if (READ_ONCE(entity->dependency))
142 * drm_sched_entity_flush - Flush a context entity
144 * @entity: scheduler entity
145 * @timeout: time to wait in for Q to become empty in jiffies.
147 * Splitting drm_sched_entity_fini() into two functions, The first one does the
148 * waiting, removes the entity from the runqueue and returns an error when the
149 * process was killed.
151 * Returns the remaining time in jiffies left from the input timeout
153 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
155 struct drm_gpu_scheduler *sched;
156 struct task_struct *last_user;
162 sched = entity->rq->sched;
164 * The client will not queue more IBs during this fini, consume existing
165 * queued IBs or discard them on SIGKILL
167 if (current->flags & PF_EXITING) {
169 ret = wait_event_timeout(
170 sched->job_scheduled,
171 drm_sched_entity_is_idle(entity),
174 wait_event_killable(sched->job_scheduled,
175 drm_sched_entity_is_idle(entity));
178 /* For killed process disable any more IBs enqueue right now */
179 last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
180 if ((!last_user || last_user == current->group_leader) &&
181 (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
182 spin_lock(&entity->rq_lock);
183 entity->stopped = true;
184 drm_sched_rq_remove_entity(entity->rq, entity);
185 spin_unlock(&entity->rq_lock);
190 EXPORT_SYMBOL(drm_sched_entity_flush);
193 * drm_sched_entity_kill_jobs - helper for drm_sched_entity_kill_jobs
196 * @cb: our callback structure
198 * Signal the scheduler finished fence when the entity in question is killed.
200 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
201 struct dma_fence_cb *cb)
203 struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
206 drm_sched_fence_finished(job->s_fence);
207 WARN_ON(job->s_fence->parent);
208 job->sched->ops->free_job(job);
212 * drm_sched_entity_kill_jobs - Make sure all remaining jobs are killed
214 * @entity: entity which is cleaned up
216 * Makes sure that all remaining jobs in an entity are killed before it is
219 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
221 struct drm_sched_job *job;
225 while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
226 struct drm_sched_fence *s_fence = job->s_fence;
228 /* Wait for all dependencies to avoid data corruptions */
229 while ((f = job->sched->ops->dependency(job, entity)))
230 dma_fence_wait(f, false);
232 drm_sched_fence_scheduled(s_fence);
233 dma_fence_set_error(&s_fence->finished, -ESRCH);
236 * When pipe is hanged by older entity, new entity might
237 * not even have chance to submit it's first job to HW
238 * and so entity->last_scheduled will remain NULL
240 if (!entity->last_scheduled) {
241 drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
245 r = dma_fence_add_callback(entity->last_scheduled,
247 drm_sched_entity_kill_jobs_cb);
249 drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
251 DRM_ERROR("fence add callback failed (%d)\n", r);
256 * drm_sched_entity_cleanup - Destroy a context entity
258 * @entity: scheduler entity
260 * This should be called after @drm_sched_entity_do_release. It goes over the
261 * entity and signals all jobs with an error code if the process was killed.
264 void drm_sched_entity_fini(struct drm_sched_entity *entity)
266 struct drm_gpu_scheduler *sched = NULL;
269 sched = entity->rq->sched;
270 drm_sched_rq_remove_entity(entity->rq, entity);
273 /* Consumption of existing IBs wasn't completed. Forcefully
276 if (spsc_queue_count(&entity->job_queue)) {
279 * Wait for thread to idle to make sure it isn't processing
282 wait_for_completion(&entity->entity_idle);
285 if (entity->dependency) {
286 dma_fence_remove_callback(entity->dependency,
288 dma_fence_put(entity->dependency);
289 entity->dependency = NULL;
292 drm_sched_entity_kill_jobs(entity);
295 dma_fence_put(entity->last_scheduled);
296 entity->last_scheduled = NULL;
298 EXPORT_SYMBOL(drm_sched_entity_fini);
301 * drm_sched_entity_fini - Destroy a context entity
303 * @entity: scheduler entity
305 * Calls drm_sched_entity_do_release() and drm_sched_entity_cleanup()
307 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
309 drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
310 drm_sched_entity_fini(entity);
312 EXPORT_SYMBOL(drm_sched_entity_destroy);
315 * drm_sched_entity_clear_dep - callback to clear the entities dependency
317 static void drm_sched_entity_clear_dep(struct dma_fence *f,
318 struct dma_fence_cb *cb)
320 struct drm_sched_entity *entity =
321 container_of(cb, struct drm_sched_entity, cb);
323 entity->dependency = NULL;
328 * drm_sched_entity_clear_dep - callback to clear the entities dependency and
331 static void drm_sched_entity_wakeup(struct dma_fence *f,
332 struct dma_fence_cb *cb)
334 struct drm_sched_entity *entity =
335 container_of(cb, struct drm_sched_entity, cb);
337 drm_sched_entity_clear_dep(f, cb);
338 drm_sched_wakeup(entity->rq->sched);
342 * drm_sched_entity_set_priority - Sets priority of the entity
344 * @entity: scheduler entity
345 * @priority: scheduler priority
347 * Update the priority of runqueus used for the entity.
349 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
350 enum drm_sched_priority priority)
352 spin_lock(&entity->rq_lock);
353 entity->priority = priority;
354 spin_unlock(&entity->rq_lock);
356 EXPORT_SYMBOL(drm_sched_entity_set_priority);
359 * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
361 * @entity: entity with dependency
363 * Add a callback to the current dependency of the entity to wake up the
364 * scheduler when the entity becomes available.
366 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
368 struct drm_gpu_scheduler *sched = entity->rq->sched;
369 struct dma_fence *fence = entity->dependency;
370 struct drm_sched_fence *s_fence;
372 if (fence->context == entity->fence_context ||
373 fence->context == entity->fence_context + 1) {
375 * Fence is a scheduled/finished fence from a job
376 * which belongs to the same entity, we can ignore
377 * fences from ourself
379 dma_fence_put(entity->dependency);
383 s_fence = to_drm_sched_fence(fence);
384 if (s_fence && s_fence->sched == sched) {
387 * Fence is from the same scheduler, only need to wait for
390 fence = dma_fence_get(&s_fence->scheduled);
391 dma_fence_put(entity->dependency);
392 entity->dependency = fence;
393 if (!dma_fence_add_callback(fence, &entity->cb,
394 drm_sched_entity_clear_dep))
397 /* Ignore it when it is already scheduled */
398 dma_fence_put(fence);
402 if (!dma_fence_add_callback(entity->dependency, &entity->cb,
403 drm_sched_entity_wakeup))
406 dma_fence_put(entity->dependency);
411 * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
413 * @entity: entity to get the job from
415 * Process all dependencies and try to get one job from the entities queue.
417 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
419 struct drm_gpu_scheduler *sched = entity->rq->sched;
420 struct drm_sched_job *sched_job;
422 sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
426 while ((entity->dependency =
427 sched->ops->dependency(sched_job, entity))) {
428 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
430 if (drm_sched_entity_add_dependency_cb(entity))
434 /* skip jobs from entity that marked guilty */
435 if (entity->guilty && atomic_read(entity->guilty))
436 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
438 dma_fence_put(entity->last_scheduled);
439 entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
441 spsc_queue_pop(&entity->job_queue);
446 * drm_sched_entity_select_rq - select a new rq for the entity
448 * @entity: scheduler entity
450 * Check all prerequisites and select a new rq for the entity for load
453 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
455 struct dma_fence *fence;
456 struct drm_gpu_scheduler *sched;
457 struct drm_sched_rq *rq;
459 if (spsc_queue_count(&entity->job_queue) || entity->num_sched_list <= 1)
462 fence = READ_ONCE(entity->last_scheduled);
463 if (fence && !dma_fence_is_signaled(fence))
466 spin_lock(&entity->rq_lock);
467 sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
468 rq = sched ? &sched->sched_rq[entity->priority] : NULL;
469 if (rq != entity->rq) {
470 drm_sched_rq_remove_entity(entity->rq, entity);
474 spin_unlock(&entity->rq_lock);
478 * drm_sched_entity_push_job - Submit a job to the entity's job queue
480 * @sched_job: job to submit
481 * @entity: scheduler entity
483 * Note: To guarantee that the order of insertion to queue matches
484 * the job's fence sequence number this function should be
485 * called with drm_sched_job_init under common lock.
487 * Returns 0 for success, negative error code otherwise.
489 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
490 struct drm_sched_entity *entity)
494 trace_drm_sched_job(sched_job, entity);
495 atomic_inc(&entity->rq->sched->score);
496 WRITE_ONCE(entity->last_user, current->group_leader);
497 first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
499 /* first job wakes up scheduler */
501 /* Add the entity to the run queue */
502 spin_lock(&entity->rq_lock);
503 if (entity->stopped) {
504 spin_unlock(&entity->rq_lock);
506 DRM_ERROR("Trying to push to a killed entity\n");
509 drm_sched_rq_add_entity(entity->rq, entity);
510 spin_unlock(&entity->rq_lock);
511 drm_sched_wakeup(entity->rq->sched);
514 EXPORT_SYMBOL(drm_sched_entity_push_job);