xref: /linux/include/drm/gpu_scheduler.h (revision 4ae68b26c3ab5a82aa271e6e9fc9b1a06e1d6b40)
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 - Prefent 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  * dependen 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_MIN,
67 	DRM_SCHED_PRIORITY_NORMAL,
68 	DRM_SCHED_PRIORITY_HIGH,
69 	DRM_SCHED_PRIORITY_KERNEL,
70 
71 	DRM_SCHED_PRIORITY_COUNT,
72 	DRM_SCHED_PRIORITY_UNSET = -2
73 };
74 
75 /* Used to chose between FIFO and RR jobs scheduling */
76 extern int drm_sched_policy;
77 
78 #define DRM_SCHED_POLICY_RR    0
79 #define DRM_SCHED_POLICY_FIFO  1
80 
81 /**
82  * struct drm_sched_entity - A wrapper around a job queue (typically
83  * attached to the DRM file_priv).
84  *
85  * Entities will emit jobs in order to their corresponding hardware
86  * ring, and the scheduler will alternate between entities based on
87  * scheduling policy.
88  */
89 struct drm_sched_entity {
90 	/**
91 	 * @list:
92 	 *
93 	 * Used to append this struct to the list of entities in the runqueue
94 	 * @rq under &drm_sched_rq.entities.
95 	 *
96 	 * Protected by &drm_sched_rq.lock of @rq.
97 	 */
98 	struct list_head		list;
99 
100 	/**
101 	 * @rq:
102 	 *
103 	 * Runqueue on which this entity is currently scheduled.
104 	 *
105 	 * FIXME: Locking is very unclear for this. Writers are protected by
106 	 * @rq_lock, but readers are generally lockless and seem to just race
107 	 * with not even a READ_ONCE.
108 	 */
109 	struct drm_sched_rq		*rq;
110 
111 	/**
112 	 * @sched_list:
113 	 *
114 	 * A list of schedulers (struct drm_gpu_scheduler).  Jobs from this entity can
115 	 * be scheduled on any scheduler on this list.
116 	 *
117 	 * This can be modified by calling drm_sched_entity_modify_sched().
118 	 * Locking is entirely up to the driver, see the above function for more
119 	 * details.
120 	 *
121 	 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
122 	 * set already.
123 	 *
124 	 * FIXME: This means priority changes through
125 	 * drm_sched_entity_set_priority() will be lost henceforth in this case.
126 	 */
127 	struct drm_gpu_scheduler        **sched_list;
128 
129 	/**
130 	 * @num_sched_list:
131 	 *
132 	 * Number of drm_gpu_schedulers in the @sched_list.
133 	 */
134 	unsigned int                    num_sched_list;
135 
136 	/**
137 	 * @priority:
138 	 *
139 	 * Priority of the entity. This can be modified by calling
140 	 * drm_sched_entity_set_priority(). Protected by &rq_lock.
141 	 */
142 	enum drm_sched_priority         priority;
143 
144 	/**
145 	 * @rq_lock:
146 	 *
147 	 * Lock to modify the runqueue to which this entity belongs.
148 	 */
149 	spinlock_t			rq_lock;
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() iff 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  * @lock: to modify the entities list.
248  * @sched: the scheduler to which this rq belongs to.
249  * @entities: list of the entities to be scheduled.
250  * @current_entity: the entity which is to be scheduled.
251  * @rb_tree_root: root of time based priory 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 	spinlock_t			lock;
259 	struct drm_gpu_scheduler	*sched;
260 	struct list_head		entities;
261 	struct drm_sched_entity		*current_entity;
262 	struct rb_root_cached		rb_tree_root;
263 };
264 
265 /**
266  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
267  */
268 struct drm_sched_fence {
269         /**
270          * @scheduled: this fence is what will be signaled by the scheduler
271          * when the job is scheduled.
272          */
273 	struct dma_fence		scheduled;
274 
275         /**
276          * @finished: this fence is what will be signaled by the scheduler
277          * when the job is completed.
278          *
279          * When setting up an out fence for the job, you should use
280          * this, since it's available immediately upon
281          * drm_sched_job_init(), and the fence returned by the driver
282          * from run_job() won't be created until the dependencies have
283          * resolved.
284          */
285 	struct dma_fence		finished;
286 
287 	/**
288 	 * @deadline: deadline set on &drm_sched_fence.finished which
289 	 * potentially needs to be propagated to &drm_sched_fence.parent
290 	 */
291 	ktime_t				deadline;
292 
293         /**
294          * @parent: the fence returned by &drm_sched_backend_ops.run_job
295          * when scheduling the job on hardware. We signal the
296          * &drm_sched_fence.finished fence once parent is signalled.
297          */
298 	struct dma_fence		*parent;
299         /**
300          * @sched: the scheduler instance to which the job having this struct
301          * belongs to.
302          */
303 	struct drm_gpu_scheduler	*sched;
304         /**
305          * @lock: the lock used by the scheduled and the finished fences.
306          */
307 	spinlock_t			lock;
308         /**
309          * @owner: job owner for debugging
310          */
311 	void				*owner;
312 };
313 
314 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
315 
316 /**
317  * struct drm_sched_job - A job to be run by an entity.
318  *
319  * @queue_node: used to append this struct to the queue of jobs in an entity.
320  * @list: a job participates in a "pending" and "done" lists.
321  * @sched: the scheduler instance on which this job is scheduled.
322  * @s_fence: contains the fences for the scheduling of job.
323  * @finish_cb: the callback for the finished fence.
324  * @work: Helper to reschdeule job kill to different context.
325  * @id: a unique id assigned to each job scheduled on the scheduler.
326  * @karma: increment on every hang caused by this job. If this exceeds the hang
327  *         limit of the scheduler then the job is marked guilty and will not
328  *         be scheduled further.
329  * @s_priority: the priority of the job.
330  * @entity: the entity to which this job belongs.
331  * @cb: the callback for the parent fence in s_fence.
332  *
333  * A job is created by the driver using drm_sched_job_init(), and
334  * should call drm_sched_entity_push_job() once it wants the scheduler
335  * to schedule the job.
336  */
337 struct drm_sched_job {
338 	struct spsc_node		queue_node;
339 	struct list_head		list;
340 	struct drm_gpu_scheduler	*sched;
341 	struct drm_sched_fence		*s_fence;
342 
343 	/*
344 	 * work is used only after finish_cb has been used and will not be
345 	 * accessed anymore.
346 	 */
347 	union {
348 		struct dma_fence_cb		finish_cb;
349 		struct work_struct		work;
350 	};
351 
352 	uint64_t			id;
353 	atomic_t			karma;
354 	enum drm_sched_priority		s_priority;
355 	struct drm_sched_entity         *entity;
356 	struct dma_fence_cb		cb;
357 	/**
358 	 * @dependencies:
359 	 *
360 	 * Contains the dependencies as struct dma_fence for this job, see
361 	 * drm_sched_job_add_dependency() and
362 	 * drm_sched_job_add_implicit_dependencies().
363 	 */
364 	struct xarray			dependencies;
365 
366 	/** @last_dependency: tracks @dependencies as they signal */
367 	unsigned long			last_dependency;
368 
369 	/**
370 	 * @submit_ts:
371 	 *
372 	 * When the job was pushed into the entity queue.
373 	 */
374 	ktime_t                         submit_ts;
375 };
376 
377 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
378 					    int threshold)
379 {
380 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
381 }
382 
383 enum drm_gpu_sched_stat {
384 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
385 	DRM_GPU_SCHED_STAT_NOMINAL,
386 	DRM_GPU_SCHED_STAT_ENODEV,
387 };
388 
389 /**
390  * struct drm_sched_backend_ops - Define the backend operations
391  *	called by the scheduler
392  *
393  * These functions should be implemented in the driver side.
394  */
395 struct drm_sched_backend_ops {
396 	/**
397 	 * @prepare_job:
398 	 *
399 	 * Called when the scheduler is considering scheduling this job next, to
400 	 * get another struct dma_fence for this job to block on.  Once it
401 	 * returns NULL, run_job() may be called.
402 	 *
403 	 * Can be NULL if no additional preparation to the dependencies are
404 	 * necessary. Skipped when jobs are killed instead of run.
405 	 */
406 	struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job,
407 					 struct drm_sched_entity *s_entity);
408 
409 	/**
410          * @run_job: Called to execute the job once all of the dependencies
411          * have been resolved.  This may be called multiple times, if
412 	 * timedout_job() has happened and drm_sched_job_recovery()
413 	 * decides to try it again.
414 	 */
415 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
416 
417 	/**
418 	 * @timedout_job: Called when a job has taken too long to execute,
419 	 * to trigger GPU recovery.
420 	 *
421 	 * This method is called in a workqueue context.
422 	 *
423 	 * Drivers typically issue a reset to recover from GPU hangs, and this
424 	 * procedure usually follows the following workflow:
425 	 *
426 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
427 	 *    scheduler thread and cancel the timeout work, guaranteeing that
428 	 *    nothing is queued while we reset the hardware queue
429 	 * 2. Try to gracefully stop non-faulty jobs (optional)
430 	 * 3. Issue a GPU reset (driver-specific)
431 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
432 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
433 	 *    jobs can be queued, and the scheduler thread is unblocked
434 	 *
435 	 * Note that some GPUs have distinct hardware queues but need to reset
436 	 * the GPU globally, which requires extra synchronization between the
437 	 * timeout handler of the different &drm_gpu_scheduler. One way to
438 	 * achieve this synchronization is to create an ordered workqueue
439 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
440 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
441 	 * executed sequentially. The above workflow needs to be slightly
442 	 * adjusted in that case:
443 	 *
444 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
445 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
446 	 *    the reset (optional)
447 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
448 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
449 	 *    drm_sched_resubmit_jobs()
450 	 * 5. Restart all schedulers that were stopped in step #1 using
451 	 *    drm_sched_start()
452 	 *
453 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
454 	 * and the underlying driver has started or completed recovery.
455 	 *
456 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
457 	 * available, i.e. has been unplugged.
458 	 */
459 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
460 
461 	/**
462          * @free_job: Called once the job's finished fence has been signaled
463          * and it's time to clean it up.
464 	 */
465 	void (*free_job)(struct drm_sched_job *sched_job);
466 };
467 
468 /**
469  * struct drm_gpu_scheduler - scheduler instance-specific data
470  *
471  * @ops: backend operations provided by the driver.
472  * @hw_submission_limit: the max size of the hardware queue.
473  * @timeout: the time after which a job is removed from the scheduler.
474  * @name: name of the ring for which this scheduler is being used.
475  * @sched_rq: priority wise array of run queues.
476  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
477  *                  is ready to be scheduled.
478  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
479  *                 waits on this wait queue until all the scheduled jobs are
480  *                 finished.
481  * @hw_rq_count: the number of jobs currently in the hardware queue.
482  * @job_id_count: used to assign unique id to the each job.
483  * @timeout_wq: workqueue used to queue @work_tdr
484  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
485  *            timeout interval is over.
486  * @thread: the kthread on which the scheduler which run.
487  * @pending_list: the list of jobs which are currently in the job queue.
488  * @job_list_lock: lock to protect the pending_list.
489  * @hang_limit: once the hangs by a job crosses this limit then it is marked
490  *              guilty and it will no longer be considered for scheduling.
491  * @score: score to help loadbalancer pick a idle sched
492  * @_score: score used when the driver doesn't provide one
493  * @ready: marks if the underlying HW is ready to work
494  * @free_guilty: A hit to time out handler to free the guilty job.
495  * @dev: system &struct device
496  *
497  * One scheduler is implemented for each hardware ring.
498  */
499 struct drm_gpu_scheduler {
500 	const struct drm_sched_backend_ops	*ops;
501 	uint32_t			hw_submission_limit;
502 	long				timeout;
503 	const char			*name;
504 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
505 	wait_queue_head_t		wake_up_worker;
506 	wait_queue_head_t		job_scheduled;
507 	atomic_t			hw_rq_count;
508 	atomic64_t			job_id_count;
509 	struct workqueue_struct		*timeout_wq;
510 	struct delayed_work		work_tdr;
511 	struct task_struct		*thread;
512 	struct list_head		pending_list;
513 	spinlock_t			job_list_lock;
514 	int				hang_limit;
515 	atomic_t                        *score;
516 	atomic_t                        _score;
517 	bool				ready;
518 	bool				free_guilty;
519 	struct device			*dev;
520 };
521 
522 int drm_sched_init(struct drm_gpu_scheduler *sched,
523 		   const struct drm_sched_backend_ops *ops,
524 		   uint32_t hw_submission, unsigned hang_limit,
525 		   long timeout, struct workqueue_struct *timeout_wq,
526 		   atomic_t *score, const char *name, struct device *dev);
527 
528 void drm_sched_fini(struct drm_gpu_scheduler *sched);
529 int drm_sched_job_init(struct drm_sched_job *job,
530 		       struct drm_sched_entity *entity,
531 		       void *owner);
532 void drm_sched_job_arm(struct drm_sched_job *job);
533 int drm_sched_job_add_dependency(struct drm_sched_job *job,
534 				 struct dma_fence *fence);
535 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
536 					 struct drm_file *file,
537 					 u32 handle,
538 					 u32 point);
539 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
540 					struct dma_resv *resv,
541 					enum dma_resv_usage usage);
542 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
543 					    struct drm_gem_object *obj,
544 					    bool write);
545 
546 
547 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
548 				    struct drm_gpu_scheduler **sched_list,
549                                    unsigned int num_sched_list);
550 
551 void drm_sched_job_cleanup(struct drm_sched_job *job);
552 void drm_sched_wakeup_if_can_queue(struct drm_gpu_scheduler *sched);
553 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
554 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
555 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
556 void drm_sched_increase_karma(struct drm_sched_job *bad);
557 void drm_sched_reset_karma(struct drm_sched_job *bad);
558 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
559 bool drm_sched_dependency_optimized(struct dma_fence* fence,
560 				    struct drm_sched_entity *entity);
561 void drm_sched_fault(struct drm_gpu_scheduler *sched);
562 
563 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
564 			     struct drm_sched_entity *entity);
565 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
566 				struct drm_sched_entity *entity);
567 
568 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts);
569 
570 int drm_sched_entity_init(struct drm_sched_entity *entity,
571 			  enum drm_sched_priority priority,
572 			  struct drm_gpu_scheduler **sched_list,
573 			  unsigned int num_sched_list,
574 			  atomic_t *guilty);
575 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
576 void drm_sched_entity_fini(struct drm_sched_entity *entity);
577 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
578 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
579 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
580 void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
581 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
582 				   enum drm_sched_priority priority);
583 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
584 int drm_sched_entity_error(struct drm_sched_entity *entity);
585 
586 struct drm_sched_fence *drm_sched_fence_alloc(
587 	struct drm_sched_entity *s_entity, void *owner);
588 void drm_sched_fence_init(struct drm_sched_fence *fence,
589 			  struct drm_sched_entity *entity);
590 void drm_sched_fence_free(struct drm_sched_fence *fence);
591 
592 void drm_sched_fence_scheduled(struct drm_sched_fence *fence,
593 			       struct dma_fence *parent);
594 void drm_sched_fence_finished(struct drm_sched_fence *fence, int result);
595 
596 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
597 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
598 		                unsigned long remaining);
599 struct drm_gpu_scheduler *
600 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
601 		     unsigned int num_sched_list);
602 
603 #endif
604