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: 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 */
53int 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}
84EXPORT_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 */
93void 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}
102EXPORT_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 */
111static 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 */
129bool 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 */
152long 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}
189EXPORT_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 */
199static 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 */
218static 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 */
258void 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}
292EXPORT_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 */
301void 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}
306EXPORT_SYMBOL(drm_sched_entity_destroy);
307
308/**
309 * drm_sched_entity_clear_dep - callback to clear the entities dependency
310 */
311static 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 */
325static 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 */
343void 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}
350EXPORT_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 */
360static 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 */
411struct 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 */
447void 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 */
483void 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}
508EXPORT_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);