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 #ifndef _DRM_GPU_SCHEDULER_H_
25 #define _DRM_GPU_SCHEDULER_H_
26
27 #include <drm/spsc_queue.h>
28 #include <linux/dma-fence.h>
29 #include <linux/completion.h>
30 #include <linux/xarray.h>
31 #include <linux/workqueue.h>
32
33 #define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
34
35 /**
36 * DRM_SCHED_FENCE_DONT_PIPELINE - Prevent dependency pipelining
37 *
38 * Setting this flag on a scheduler fence prevents pipelining of jobs depending
39 * on this fence. In other words we always insert a full CPU round trip before
40 * dependent jobs are pushed to the hw queue.
41 */
42 #define DRM_SCHED_FENCE_DONT_PIPELINE DMA_FENCE_FLAG_USER_BITS
43
44 /**
45 * DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT - A fence deadline hint has been set
46 *
47 * Because we could have a deadline hint can be set before the backing hw
48 * fence is created, we need to keep track of whether a deadline has already
49 * been set.
50 */
51 #define DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT (DMA_FENCE_FLAG_USER_BITS + 1)
52
53 enum dma_resv_usage;
54 struct dma_resv;
55 struct drm_gem_object;
56
57 struct drm_gpu_scheduler;
58 struct drm_sched_rq;
59
60 struct drm_file;
61
62 /* These are often used as an (initial) index
63 * to an array, and as such should start at 0.
64 */
65 enum drm_sched_priority {
66 DRM_SCHED_PRIORITY_KERNEL,
67 DRM_SCHED_PRIORITY_HIGH,
68 DRM_SCHED_PRIORITY_NORMAL,
69 DRM_SCHED_PRIORITY_LOW,
70
71 DRM_SCHED_PRIORITY_COUNT
72 };
73
74 /* Used to choose between FIFO and RR job-scheduling */
75 extern int drm_sched_policy;
76
77 #define DRM_SCHED_POLICY_RR 0
78 #define DRM_SCHED_POLICY_FIFO 1
79
80 /**
81 * struct drm_sched_entity - A wrapper around a job queue (typically
82 * attached to the DRM file_priv).
83 *
84 * Entities will emit jobs in order to their corresponding hardware
85 * ring, and the scheduler will alternate between entities based on
86 * scheduling policy.
87 */
88 struct drm_sched_entity {
89 /**
90 * @list:
91 *
92 * Used to append this struct to the list of entities in the runqueue
93 * @rq under &drm_sched_rq.entities.
94 *
95 * Protected by &drm_sched_rq.lock of @rq.
96 */
97 struct list_head list;
98
99 /**
100 * @lock:
101 *
102 * Lock protecting the run-queue (@rq) to which this entity belongs,
103 * @priority and the list of schedulers (@sched_list, @num_sched_list).
104 */
105 spinlock_t lock;
106
107 /**
108 * @rq:
109 *
110 * Runqueue on which this entity is currently scheduled.
111 *
112 * FIXME: Locking is very unclear for this. Writers are protected by
113 * @lock, but readers are generally lockless and seem to just race with
114 * not even a READ_ONCE.
115 */
116 struct drm_sched_rq *rq;
117
118 /**
119 * @sched_list:
120 *
121 * A list of schedulers (struct drm_gpu_scheduler). Jobs from this entity can
122 * be scheduled on any scheduler on this list.
123 *
124 * This can be modified by calling drm_sched_entity_modify_sched().
125 * Locking is entirely up to the driver, see the above function for more
126 * details.
127 *
128 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
129 * set already.
130 *
131 * FIXME: This means priority changes through
132 * drm_sched_entity_set_priority() will be lost henceforth in this case.
133 */
134 struct drm_gpu_scheduler **sched_list;
135
136 /**
137 * @num_sched_list:
138 *
139 * Number of drm_gpu_schedulers in the @sched_list.
140 */
141 unsigned int num_sched_list;
142
143 /**
144 * @priority:
145 *
146 * Priority of the entity. This can be modified by calling
147 * drm_sched_entity_set_priority(). Protected by @lock.
148 */
149 enum drm_sched_priority priority;
150
151 /**
152 * @job_queue: the list of jobs of this entity.
153 */
154 struct spsc_queue job_queue;
155
156 /**
157 * @fence_seq:
158 *
159 * A linearly increasing seqno incremented with each new
160 * &drm_sched_fence which is part of the entity.
161 *
162 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
163 * this doesn't need to be atomic.
164 */
165 atomic_t fence_seq;
166
167 /**
168 * @fence_context:
169 *
170 * A unique context for all the fences which belong to this entity. The
171 * &drm_sched_fence.scheduled uses the fence_context but
172 * &drm_sched_fence.finished uses fence_context + 1.
173 */
174 uint64_t fence_context;
175
176 /**
177 * @dependency:
178 *
179 * The dependency fence of the job which is on the top of the job queue.
180 */
181 struct dma_fence *dependency;
182
183 /**
184 * @cb:
185 *
186 * Callback for the dependency fence above.
187 */
188 struct dma_fence_cb cb;
189
190 /**
191 * @guilty:
192 *
193 * Points to entities' guilty.
194 */
195 atomic_t *guilty;
196
197 /**
198 * @last_scheduled:
199 *
200 * Points to the finished fence of the last scheduled job. Only written
201 * by the scheduler thread, can be accessed locklessly from
202 * drm_sched_job_arm() if the queue is empty.
203 */
204 struct dma_fence __rcu *last_scheduled;
205
206 /**
207 * @last_user: last group leader pushing a job into the entity.
208 */
209 struct task_struct *last_user;
210
211 /**
212 * @stopped:
213 *
214 * Marks the enity as removed from rq and destined for
215 * termination. This is set by calling drm_sched_entity_flush() and by
216 * drm_sched_fini().
217 */
218 bool stopped;
219
220 /**
221 * @entity_idle:
222 *
223 * Signals when entity is not in use, used to sequence entity cleanup in
224 * drm_sched_entity_fini().
225 */
226 struct completion entity_idle;
227
228 /**
229 * @oldest_job_waiting:
230 *
231 * Marks earliest job waiting in SW queue
232 */
233 ktime_t oldest_job_waiting;
234
235 /**
236 * @rb_tree_node:
237 *
238 * The node used to insert this entity into time based priority queue
239 */
240 struct rb_node rb_tree_node;
241
242 };
243
244 /**
245 * struct drm_sched_rq - queue of entities to be scheduled.
246 *
247 * @sched: the scheduler to which this rq belongs to.
248 * @lock: protects @entities, @rb_tree_root and @current_entity.
249 * @current_entity: the entity which is to be scheduled.
250 * @entities: list of the entities to be scheduled.
251 * @rb_tree_root: root of time based priority queue of entities for FIFO scheduling
252 *
253 * Run queue is a set of entities scheduling command submissions for
254 * one specific ring. It implements the scheduling policy that selects
255 * the next entity to emit commands from.
256 */
257 struct drm_sched_rq {
258 struct drm_gpu_scheduler *sched;
259
260 spinlock_t lock;
261 /* Following members are protected by the @lock: */
262 struct drm_sched_entity *current_entity;
263 struct list_head entities;
264 struct rb_root_cached rb_tree_root;
265 };
266
267 /**
268 * struct drm_sched_fence - fences corresponding to the scheduling of a job.
269 */
270 struct drm_sched_fence {
271 /**
272 * @scheduled: this fence is what will be signaled by the scheduler
273 * when the job is scheduled.
274 */
275 struct dma_fence scheduled;
276
277 /**
278 * @finished: this fence is what will be signaled by the scheduler
279 * when the job is completed.
280 *
281 * When setting up an out fence for the job, you should use
282 * this, since it's available immediately upon
283 * drm_sched_job_init(), and the fence returned by the driver
284 * from run_job() won't be created until the dependencies have
285 * resolved.
286 */
287 struct dma_fence finished;
288
289 /**
290 * @deadline: deadline set on &drm_sched_fence.finished which
291 * potentially needs to be propagated to &drm_sched_fence.parent
292 */
293 ktime_t deadline;
294
295 /**
296 * @parent: the fence returned by &drm_sched_backend_ops.run_job
297 * when scheduling the job on hardware. We signal the
298 * &drm_sched_fence.finished fence once parent is signalled.
299 */
300 struct dma_fence *parent;
301 /**
302 * @sched: the scheduler instance to which the job having this struct
303 * belongs to.
304 */
305 struct drm_gpu_scheduler *sched;
306 /**
307 * @lock: the lock used by the scheduled and the finished fences.
308 */
309 spinlock_t lock;
310 /**
311 * @owner: job owner for debugging
312 */
313 void *owner;
314 };
315
316 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
317
318 /**
319 * struct drm_sched_job - A job to be run by an entity.
320 *
321 * @queue_node: used to append this struct to the queue of jobs in an entity.
322 * @list: a job participates in a "pending" and "done" lists.
323 * @sched: the scheduler instance on which this job is scheduled.
324 * @s_fence: contains the fences for the scheduling of job.
325 * @finish_cb: the callback for the finished fence.
326 * @credits: the number of credits this job contributes to the scheduler
327 * @work: Helper to reschedule job kill to different context.
328 * @id: a unique id assigned to each job scheduled on the scheduler.
329 * @karma: increment on every hang caused by this job. If this exceeds the hang
330 * limit of the scheduler then the job is marked guilty and will not
331 * be scheduled further.
332 * @s_priority: the priority of the job.
333 * @entity: the entity to which this job belongs.
334 * @cb: the callback for the parent fence in s_fence.
335 *
336 * A job is created by the driver using drm_sched_job_init(), and
337 * should call drm_sched_entity_push_job() once it wants the scheduler
338 * to schedule the job.
339 */
340 struct drm_sched_job {
341 struct spsc_node queue_node;
342 struct list_head list;
343
344 /**
345 * @sched:
346 *
347 * The scheduler this job is or will be scheduled on. Gets set by
348 * drm_sched_job_arm(). Valid until drm_sched_backend_ops.free_job()
349 * has finished.
350 */
351 struct drm_gpu_scheduler *sched;
352 struct drm_sched_fence *s_fence;
353
354 u32 credits;
355
356 /*
357 * work is used only after finish_cb has been used and will not be
358 * accessed anymore.
359 */
360 union {
361 struct dma_fence_cb finish_cb;
362 struct work_struct work;
363 };
364
365 uint64_t id;
366 atomic_t karma;
367 enum drm_sched_priority s_priority;
368 struct drm_sched_entity *entity;
369 struct dma_fence_cb cb;
370 /**
371 * @dependencies:
372 *
373 * Contains the dependencies as struct dma_fence for this job, see
374 * drm_sched_job_add_dependency() and
375 * drm_sched_job_add_implicit_dependencies().
376 */
377 struct xarray dependencies;
378
379 /** @last_dependency: tracks @dependencies as they signal */
380 unsigned long last_dependency;
381
382 /**
383 * @submit_ts:
384 *
385 * When the job was pushed into the entity queue.
386 */
387 ktime_t submit_ts;
388 };
389
drm_sched_invalidate_job(struct drm_sched_job * s_job,int threshold)390 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
391 int threshold)
392 {
393 return s_job && atomic_inc_return(&s_job->karma) > threshold;
394 }
395
396 enum drm_gpu_sched_stat {
397 DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
398 DRM_GPU_SCHED_STAT_NOMINAL,
399 DRM_GPU_SCHED_STAT_ENODEV,
400 };
401
402 /**
403 * struct drm_sched_backend_ops - Define the backend operations
404 * called by the scheduler
405 *
406 * These functions should be implemented in the driver side.
407 */
408 struct drm_sched_backend_ops {
409 /**
410 * @prepare_job:
411 *
412 * Called when the scheduler is considering scheduling this job next, to
413 * get another struct dma_fence for this job to block on. Once it
414 * returns NULL, run_job() may be called.
415 *
416 * Can be NULL if no additional preparation to the dependencies are
417 * necessary. Skipped when jobs are killed instead of run.
418 */
419 struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job,
420 struct drm_sched_entity *s_entity);
421
422 /**
423 * @run_job: Called to execute the job once all of the dependencies
424 * have been resolved. This may be called multiple times, if
425 * timedout_job() has happened and drm_sched_job_recovery()
426 * decides to try it again.
427 */
428 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
429
430 /**
431 * @timedout_job: Called when a job has taken too long to execute,
432 * to trigger GPU recovery.
433 *
434 * This method is called in a workqueue context.
435 *
436 * Drivers typically issue a reset to recover from GPU hangs, and this
437 * procedure usually follows the following workflow:
438 *
439 * 1. Stop the scheduler using drm_sched_stop(). This will park the
440 * scheduler thread and cancel the timeout work, guaranteeing that
441 * nothing is queued while we reset the hardware queue
442 * 2. Try to gracefully stop non-faulty jobs (optional)
443 * 3. Issue a GPU reset (driver-specific)
444 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
445 * 5. Restart the scheduler using drm_sched_start(). At that point, new
446 * jobs can be queued, and the scheduler thread is unblocked
447 *
448 * Note that some GPUs have distinct hardware queues but need to reset
449 * the GPU globally, which requires extra synchronization between the
450 * timeout handler of the different &drm_gpu_scheduler. One way to
451 * achieve this synchronization is to create an ordered workqueue
452 * (using alloc_ordered_workqueue()) at the driver level, and pass this
453 * queue to drm_sched_init(), to guarantee that timeout handlers are
454 * executed sequentially. The above workflow needs to be slightly
455 * adjusted in that case:
456 *
457 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
458 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
459 * the reset (optional)
460 * 3. Issue a GPU reset on all faulty queues (driver-specific)
461 * 4. Re-submit jobs on all schedulers impacted by the reset using
462 * drm_sched_resubmit_jobs()
463 * 5. Restart all schedulers that were stopped in step #1 using
464 * drm_sched_start()
465 *
466 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
467 * and the underlying driver has started or completed recovery.
468 *
469 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
470 * available, i.e. has been unplugged.
471 */
472 enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
473
474 /**
475 * @free_job: Called once the job's finished fence has been signaled
476 * and it's time to clean it up.
477 */
478 void (*free_job)(struct drm_sched_job *sched_job);
479
480 /**
481 * @update_job_credits: Called when the scheduler is considering this
482 * job for execution.
483 *
484 * This callback returns the number of credits the job would take if
485 * pushed to the hardware. Drivers may use this to dynamically update
486 * the job's credit count. For instance, deduct the number of credits
487 * for already signalled native fences.
488 *
489 * This callback is optional.
490 */
491 u32 (*update_job_credits)(struct drm_sched_job *sched_job);
492 };
493
494 /**
495 * struct drm_gpu_scheduler - scheduler instance-specific data
496 *
497 * @ops: backend operations provided by the driver.
498 * @credit_limit: the credit limit of this scheduler
499 * @credit_count: the current credit count of this scheduler
500 * @timeout: the time after which a job is removed from the scheduler.
501 * @name: name of the ring for which this scheduler is being used.
502 * @num_rqs: Number of run-queues. This is at most DRM_SCHED_PRIORITY_COUNT,
503 * as there's usually one run-queue per priority, but could be less.
504 * @sched_rq: An allocated array of run-queues of size @num_rqs;
505 * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
506 * waits on this wait queue until all the scheduled jobs are
507 * finished.
508 * @job_id_count: used to assign unique id to the each job.
509 * @submit_wq: workqueue used to queue @work_run_job and @work_free_job
510 * @timeout_wq: workqueue used to queue @work_tdr
511 * @work_run_job: work which calls run_job op of each scheduler.
512 * @work_free_job: work which calls free_job op of each scheduler.
513 * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
514 * timeout interval is over.
515 * @pending_list: the list of jobs which are currently in the job queue.
516 * @job_list_lock: lock to protect the pending_list.
517 * @hang_limit: once the hangs by a job crosses this limit then it is marked
518 * guilty and it will no longer be considered for scheduling.
519 * @score: score to help loadbalancer pick a idle sched
520 * @_score: score used when the driver doesn't provide one
521 * @ready: marks if the underlying HW is ready to work
522 * @free_guilty: A hit to time out handler to free the guilty job.
523 * @pause_submit: pause queuing of @work_run_job on @submit_wq
524 * @own_submit_wq: scheduler owns allocation of @submit_wq
525 * @dev: system &struct device
526 *
527 * One scheduler is implemented for each hardware ring.
528 */
529 struct drm_gpu_scheduler {
530 const struct drm_sched_backend_ops *ops;
531 u32 credit_limit;
532 atomic_t credit_count;
533 long timeout;
534 const char *name;
535 u32 num_rqs;
536 struct drm_sched_rq **sched_rq;
537 wait_queue_head_t job_scheduled;
538 atomic64_t job_id_count;
539 struct workqueue_struct *submit_wq;
540 struct workqueue_struct *timeout_wq;
541 struct work_struct work_run_job;
542 struct work_struct work_free_job;
543 struct delayed_work work_tdr;
544 struct list_head pending_list;
545 spinlock_t job_list_lock;
546 int hang_limit;
547 atomic_t *score;
548 atomic_t _score;
549 bool ready;
550 bool free_guilty;
551 bool pause_submit;
552 bool own_submit_wq;
553 struct device *dev;
554 };
555
556 int drm_sched_init(struct drm_gpu_scheduler *sched,
557 const struct drm_sched_backend_ops *ops,
558 struct workqueue_struct *submit_wq,
559 u32 num_rqs, u32 credit_limit, unsigned int hang_limit,
560 long timeout, struct workqueue_struct *timeout_wq,
561 atomic_t *score, const char *name, struct device *dev);
562
563 void drm_sched_fini(struct drm_gpu_scheduler *sched);
564 int drm_sched_job_init(struct drm_sched_job *job,
565 struct drm_sched_entity *entity,
566 u32 credits, void *owner);
567 void drm_sched_job_arm(struct drm_sched_job *job);
568 int drm_sched_job_add_dependency(struct drm_sched_job *job,
569 struct dma_fence *fence);
570 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
571 struct drm_file *file,
572 u32 handle,
573 u32 point);
574 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
575 struct dma_resv *resv,
576 enum dma_resv_usage usage);
577 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
578 struct drm_gem_object *obj,
579 bool write);
580
581
582 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
583 struct drm_gpu_scheduler **sched_list,
584 unsigned int num_sched_list);
585
586 void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched);
587 void drm_sched_job_cleanup(struct drm_sched_job *job);
588 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
589 bool drm_sched_wqueue_ready(struct drm_gpu_scheduler *sched);
590 void drm_sched_wqueue_stop(struct drm_gpu_scheduler *sched);
591 void drm_sched_wqueue_start(struct drm_gpu_scheduler *sched);
592 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
593 void drm_sched_start(struct drm_gpu_scheduler *sched, int errno);
594 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
595 void drm_sched_increase_karma(struct drm_sched_job *bad);
596 void drm_sched_reset_karma(struct drm_sched_job *bad);
597 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
598 bool drm_sched_dependency_optimized(struct dma_fence* fence,
599 struct drm_sched_entity *entity);
600 void drm_sched_fault(struct drm_gpu_scheduler *sched);
601
602 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
603 struct drm_sched_entity *entity);
604 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
605 struct drm_sched_entity *entity);
606
607 void drm_sched_rq_update_fifo_locked(struct drm_sched_entity *entity,
608 struct drm_sched_rq *rq, ktime_t ts);
609
610 int drm_sched_entity_init(struct drm_sched_entity *entity,
611 enum drm_sched_priority priority,
612 struct drm_gpu_scheduler **sched_list,
613 unsigned int num_sched_list,
614 atomic_t *guilty);
615 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
616 void drm_sched_entity_fini(struct drm_sched_entity *entity);
617 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
618 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
619 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
620 void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
621 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
622 enum drm_sched_priority priority);
623 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
624 int drm_sched_entity_error(struct drm_sched_entity *entity);
625
626 struct drm_sched_fence *drm_sched_fence_alloc(
627 struct drm_sched_entity *s_entity, void *owner);
628 void drm_sched_fence_init(struct drm_sched_fence *fence,
629 struct drm_sched_entity *entity);
630 void drm_sched_fence_free(struct drm_sched_fence *fence);
631
632 void drm_sched_fence_scheduled(struct drm_sched_fence *fence,
633 struct dma_fence *parent);
634 void drm_sched_fence_finished(struct drm_sched_fence *fence, int result);
635
636 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
637 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
638 unsigned long remaining);
639 struct drm_gpu_scheduler *
640 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
641 unsigned int num_sched_list);
642
643 #endif
644