xref: /openbmc/linux/include/drm/gpu_scheduler.h (revision fa8391ad)
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 };
73 
74 /* Used to chose between FIFO and RR jobs 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 	 * @rq:
101 	 *
102 	 * Runqueue on which this entity is currently scheduled.
103 	 *
104 	 * FIXME: Locking is very unclear for this. Writers are protected by
105 	 * @rq_lock, but readers are generally lockless and seem to just race
106 	 * with not even a READ_ONCE.
107 	 */
108 	struct drm_sched_rq		*rq;
109 
110 	/**
111 	 * @sched_list:
112 	 *
113 	 * A list of schedulers (struct drm_gpu_scheduler).  Jobs from this entity can
114 	 * be scheduled on any scheduler on this list.
115 	 *
116 	 * This can be modified by calling drm_sched_entity_modify_sched().
117 	 * Locking is entirely up to the driver, see the above function for more
118 	 * details.
119 	 *
120 	 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
121 	 * set already.
122 	 *
123 	 * FIXME: This means priority changes through
124 	 * drm_sched_entity_set_priority() will be lost henceforth in this case.
125 	 */
126 	struct drm_gpu_scheduler        **sched_list;
127 
128 	/**
129 	 * @num_sched_list:
130 	 *
131 	 * Number of drm_gpu_schedulers in the @sched_list.
132 	 */
133 	unsigned int                    num_sched_list;
134 
135 	/**
136 	 * @priority:
137 	 *
138 	 * Priority of the entity. This can be modified by calling
139 	 * drm_sched_entity_set_priority(). Protected by &rq_lock.
140 	 */
141 	enum drm_sched_priority         priority;
142 
143 	/**
144 	 * @rq_lock:
145 	 *
146 	 * Lock to modify the runqueue to which this entity belongs.
147 	 */
148 	spinlock_t			rq_lock;
149 
150 	/**
151 	 * @job_queue: the list of jobs of this entity.
152 	 */
153 	struct spsc_queue		job_queue;
154 
155 	/**
156 	 * @fence_seq:
157 	 *
158 	 * A linearly increasing seqno incremented with each new
159 	 * &drm_sched_fence which is part of the entity.
160 	 *
161 	 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
162 	 * this doesn't need to be atomic.
163 	 */
164 	atomic_t			fence_seq;
165 
166 	/**
167 	 * @fence_context:
168 	 *
169 	 * A unique context for all the fences which belong to this entity.  The
170 	 * &drm_sched_fence.scheduled uses the fence_context but
171 	 * &drm_sched_fence.finished uses fence_context + 1.
172 	 */
173 	uint64_t			fence_context;
174 
175 	/**
176 	 * @dependency:
177 	 *
178 	 * The dependency fence of the job which is on the top of the job queue.
179 	 */
180 	struct dma_fence		*dependency;
181 
182 	/**
183 	 * @cb:
184 	 *
185 	 * Callback for the dependency fence above.
186 	 */
187 	struct dma_fence_cb		cb;
188 
189 	/**
190 	 * @guilty:
191 	 *
192 	 * Points to entities' guilty.
193 	 */
194 	atomic_t			*guilty;
195 
196 	/**
197 	 * @last_scheduled:
198 	 *
199 	 * Points to the finished fence of the last scheduled job. Only written
200 	 * by the scheduler thread, can be accessed locklessly from
201 	 * drm_sched_job_arm() iff the queue is empty.
202 	 */
203 	struct dma_fence __rcu		*last_scheduled;
204 
205 	/**
206 	 * @last_user: last group leader pushing a job into the entity.
207 	 */
208 	struct task_struct		*last_user;
209 
210 	/**
211 	 * @stopped:
212 	 *
213 	 * Marks the enity as removed from rq and destined for
214 	 * termination. This is set by calling drm_sched_entity_flush() and by
215 	 * drm_sched_fini().
216 	 */
217 	bool 				stopped;
218 
219 	/**
220 	 * @entity_idle:
221 	 *
222 	 * Signals when entity is not in use, used to sequence entity cleanup in
223 	 * drm_sched_entity_fini().
224 	 */
225 	struct completion		entity_idle;
226 
227 	/**
228 	 * @oldest_job_waiting:
229 	 *
230 	 * Marks earliest job waiting in SW queue
231 	 */
232 	ktime_t				oldest_job_waiting;
233 
234 	/**
235 	 * @rb_tree_node:
236 	 *
237 	 * The node used to insert this entity into time based priority queue
238 	 */
239 	struct rb_node			rb_tree_node;
240 
241 };
242 
243 /**
244  * struct drm_sched_rq - queue of entities to be scheduled.
245  *
246  * @lock: to modify the entities list.
247  * @sched: the scheduler to which this rq belongs to.
248  * @entities: list of the entities to be scheduled.
249  * @current_entity: the entity which is to be scheduled.
250  * @rb_tree_root: root of time based priory queue of entities for FIFO scheduling
251  *
252  * Run queue is a set of entities scheduling command submissions for
253  * one specific ring. It implements the scheduling policy that selects
254  * the next entity to emit commands from.
255  */
256 struct drm_sched_rq {
257 	spinlock_t			lock;
258 	struct drm_gpu_scheduler	*sched;
259 	struct list_head		entities;
260 	struct drm_sched_entity		*current_entity;
261 	struct rb_root_cached		rb_tree_root;
262 };
263 
264 /**
265  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
266  */
267 struct drm_sched_fence {
268         /**
269          * @scheduled: this fence is what will be signaled by the scheduler
270          * when the job is scheduled.
271          */
272 	struct dma_fence		scheduled;
273 
274         /**
275          * @finished: this fence is what will be signaled by the scheduler
276          * when the job is completed.
277          *
278          * When setting up an out fence for the job, you should use
279          * this, since it's available immediately upon
280          * drm_sched_job_init(), and the fence returned by the driver
281          * from run_job() won't be created until the dependencies have
282          * resolved.
283          */
284 	struct dma_fence		finished;
285 
286 	/**
287 	 * @deadline: deadline set on &drm_sched_fence.finished which
288 	 * potentially needs to be propagated to &drm_sched_fence.parent
289 	 */
290 	ktime_t				deadline;
291 
292         /**
293          * @parent: the fence returned by &drm_sched_backend_ops.run_job
294          * when scheduling the job on hardware. We signal the
295          * &drm_sched_fence.finished fence once parent is signalled.
296          */
297 	struct dma_fence		*parent;
298         /**
299          * @sched: the scheduler instance to which the job having this struct
300          * belongs to.
301          */
302 	struct drm_gpu_scheduler	*sched;
303         /**
304          * @lock: the lock used by the scheduled and the finished fences.
305          */
306 	spinlock_t			lock;
307         /**
308          * @owner: job owner for debugging
309          */
310 	void				*owner;
311 };
312 
313 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
314 
315 /**
316  * struct drm_sched_job - A job to be run by an entity.
317  *
318  * @queue_node: used to append this struct to the queue of jobs in an entity.
319  * @list: a job participates in a "pending" and "done" lists.
320  * @sched: the scheduler instance on which this job is scheduled.
321  * @s_fence: contains the fences for the scheduling of job.
322  * @finish_cb: the callback for the finished fence.
323  * @work: Helper to reschdeule job kill to different context.
324  * @id: a unique id assigned to each job scheduled on the scheduler.
325  * @karma: increment on every hang caused by this job. If this exceeds the hang
326  *         limit of the scheduler then the job is marked guilty and will not
327  *         be scheduled further.
328  * @s_priority: the priority of the job.
329  * @entity: the entity to which this job belongs.
330  * @cb: the callback for the parent fence in s_fence.
331  *
332  * A job is created by the driver using drm_sched_job_init(), and
333  * should call drm_sched_entity_push_job() once it wants the scheduler
334  * to schedule the job.
335  */
336 struct drm_sched_job {
337 	struct spsc_node		queue_node;
338 	struct list_head		list;
339 	struct drm_gpu_scheduler	*sched;
340 	struct drm_sched_fence		*s_fence;
341 
342 	/*
343 	 * work is used only after finish_cb has been used and will not be
344 	 * accessed anymore.
345 	 */
346 	union {
347 		struct dma_fence_cb		finish_cb;
348 		struct work_struct		work;
349 	};
350 
351 	uint64_t			id;
352 	atomic_t			karma;
353 	enum drm_sched_priority		s_priority;
354 	struct drm_sched_entity         *entity;
355 	struct dma_fence_cb		cb;
356 	/**
357 	 * @dependencies:
358 	 *
359 	 * Contains the dependencies as struct dma_fence for this job, see
360 	 * drm_sched_job_add_dependency() and
361 	 * drm_sched_job_add_implicit_dependencies().
362 	 */
363 	struct xarray			dependencies;
364 
365 	/** @last_dependency: tracks @dependencies as they signal */
366 	unsigned long			last_dependency;
367 
368 	/**
369 	 * @submit_ts:
370 	 *
371 	 * When the job was pushed into the entity queue.
372 	 */
373 	ktime_t                         submit_ts;
374 };
375 
drm_sched_invalidate_job(struct drm_sched_job * s_job,int threshold)376 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
377 					    int threshold)
378 {
379 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
380 }
381 
382 enum drm_gpu_sched_stat {
383 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
384 	DRM_GPU_SCHED_STAT_NOMINAL,
385 	DRM_GPU_SCHED_STAT_ENODEV,
386 };
387 
388 /**
389  * struct drm_sched_backend_ops - Define the backend operations
390  *	called by the scheduler
391  *
392  * These functions should be implemented in the driver side.
393  */
394 struct drm_sched_backend_ops {
395 	/**
396 	 * @prepare_job:
397 	 *
398 	 * Called when the scheduler is considering scheduling this job next, to
399 	 * get another struct dma_fence for this job to block on.  Once it
400 	 * returns NULL, run_job() may be called.
401 	 *
402 	 * Can be NULL if no additional preparation to the dependencies are
403 	 * necessary. Skipped when jobs are killed instead of run.
404 	 */
405 	struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job,
406 					 struct drm_sched_entity *s_entity);
407 
408 	/**
409          * @run_job: Called to execute the job once all of the dependencies
410          * have been resolved.  This may be called multiple times, if
411 	 * timedout_job() has happened and drm_sched_job_recovery()
412 	 * decides to try it again.
413 	 */
414 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
415 
416 	/**
417 	 * @timedout_job: Called when a job has taken too long to execute,
418 	 * to trigger GPU recovery.
419 	 *
420 	 * This method is called in a workqueue context.
421 	 *
422 	 * Drivers typically issue a reset to recover from GPU hangs, and this
423 	 * procedure usually follows the following workflow:
424 	 *
425 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
426 	 *    scheduler thread and cancel the timeout work, guaranteeing that
427 	 *    nothing is queued while we reset the hardware queue
428 	 * 2. Try to gracefully stop non-faulty jobs (optional)
429 	 * 3. Issue a GPU reset (driver-specific)
430 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
431 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
432 	 *    jobs can be queued, and the scheduler thread is unblocked
433 	 *
434 	 * Note that some GPUs have distinct hardware queues but need to reset
435 	 * the GPU globally, which requires extra synchronization between the
436 	 * timeout handler of the different &drm_gpu_scheduler. One way to
437 	 * achieve this synchronization is to create an ordered workqueue
438 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
439 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
440 	 * executed sequentially. The above workflow needs to be slightly
441 	 * adjusted in that case:
442 	 *
443 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
444 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
445 	 *    the reset (optional)
446 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
447 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
448 	 *    drm_sched_resubmit_jobs()
449 	 * 5. Restart all schedulers that were stopped in step #1 using
450 	 *    drm_sched_start()
451 	 *
452 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
453 	 * and the underlying driver has started or completed recovery.
454 	 *
455 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
456 	 * available, i.e. has been unplugged.
457 	 */
458 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
459 
460 	/**
461          * @free_job: Called once the job's finished fence has been signaled
462          * and it's time to clean it up.
463 	 */
464 	void (*free_job)(struct drm_sched_job *sched_job);
465 };
466 
467 /**
468  * struct drm_gpu_scheduler - scheduler instance-specific data
469  *
470  * @ops: backend operations provided by the driver.
471  * @hw_submission_limit: the max size of the hardware queue.
472  * @timeout: the time after which a job is removed from the scheduler.
473  * @name: name of the ring for which this scheduler is being used.
474  * @sched_rq: priority wise array of run queues.
475  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
476  *                  is ready to be scheduled.
477  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
478  *                 waits on this wait queue until all the scheduled jobs are
479  *                 finished.
480  * @hw_rq_count: the number of jobs currently in the hardware queue.
481  * @job_id_count: used to assign unique id to the each job.
482  * @timeout_wq: workqueue used to queue @work_tdr
483  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
484  *            timeout interval is over.
485  * @thread: the kthread on which the scheduler which run.
486  * @pending_list: the list of jobs which are currently in the job queue.
487  * @job_list_lock: lock to protect the pending_list.
488  * @hang_limit: once the hangs by a job crosses this limit then it is marked
489  *              guilty and it will no longer be considered for scheduling.
490  * @score: score to help loadbalancer pick a idle sched
491  * @_score: score used when the driver doesn't provide one
492  * @ready: marks if the underlying HW is ready to work
493  * @free_guilty: A hit to time out handler to free the guilty job.
494  * @dev: system &struct device
495  *
496  * One scheduler is implemented for each hardware ring.
497  */
498 struct drm_gpu_scheduler {
499 	const struct drm_sched_backend_ops	*ops;
500 	uint32_t			hw_submission_limit;
501 	long				timeout;
502 	const char			*name;
503 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
504 	wait_queue_head_t		wake_up_worker;
505 	wait_queue_head_t		job_scheduled;
506 	atomic_t			hw_rq_count;
507 	atomic64_t			job_id_count;
508 	struct workqueue_struct		*timeout_wq;
509 	struct delayed_work		work_tdr;
510 	struct task_struct		*thread;
511 	struct list_head		pending_list;
512 	spinlock_t			job_list_lock;
513 	int				hang_limit;
514 	atomic_t                        *score;
515 	atomic_t                        _score;
516 	bool				ready;
517 	bool				free_guilty;
518 	struct device			*dev;
519 };
520 
521 int drm_sched_init(struct drm_gpu_scheduler *sched,
522 		   const struct drm_sched_backend_ops *ops,
523 		   uint32_t hw_submission, unsigned hang_limit,
524 		   long timeout, struct workqueue_struct *timeout_wq,
525 		   atomic_t *score, const char *name, struct device *dev);
526 
527 void drm_sched_fini(struct drm_gpu_scheduler *sched);
528 int drm_sched_job_init(struct drm_sched_job *job,
529 		       struct drm_sched_entity *entity,
530 		       void *owner);
531 void drm_sched_job_arm(struct drm_sched_job *job);
532 int drm_sched_job_add_dependency(struct drm_sched_job *job,
533 				 struct dma_fence *fence);
534 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
535 					 struct drm_file *file,
536 					 u32 handle,
537 					 u32 point);
538 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
539 					struct dma_resv *resv,
540 					enum dma_resv_usage usage);
541 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
542 					    struct drm_gem_object *obj,
543 					    bool write);
544 
545 
546 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
547 				    struct drm_gpu_scheduler **sched_list,
548                                    unsigned int num_sched_list);
549 
550 void drm_sched_job_cleanup(struct drm_sched_job *job);
551 void drm_sched_wakeup_if_can_queue(struct drm_gpu_scheduler *sched);
552 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
553 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
554 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
555 void drm_sched_increase_karma(struct drm_sched_job *bad);
556 void drm_sched_reset_karma(struct drm_sched_job *bad);
557 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
558 bool drm_sched_dependency_optimized(struct dma_fence* fence,
559 				    struct drm_sched_entity *entity);
560 void drm_sched_fault(struct drm_gpu_scheduler *sched);
561 
562 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
563 			     struct drm_sched_entity *entity);
564 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
565 				struct drm_sched_entity *entity);
566 
567 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts);
568 
569 int drm_sched_entity_init(struct drm_sched_entity *entity,
570 			  enum drm_sched_priority priority,
571 			  struct drm_gpu_scheduler **sched_list,
572 			  unsigned int num_sched_list,
573 			  atomic_t *guilty);
574 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
575 void drm_sched_entity_fini(struct drm_sched_entity *entity);
576 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
577 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
578 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
579 void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
580 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
581 				   enum drm_sched_priority priority);
582 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
583 int drm_sched_entity_error(struct drm_sched_entity *entity);
584 
585 struct drm_sched_fence *drm_sched_fence_alloc(
586 	struct drm_sched_entity *s_entity, void *owner);
587 void drm_sched_fence_init(struct drm_sched_fence *fence,
588 			  struct drm_sched_entity *entity);
589 void drm_sched_fence_free(struct drm_sched_fence *fence);
590 
591 void drm_sched_fence_scheduled(struct drm_sched_fence *fence,
592 			       struct dma_fence *parent);
593 void drm_sched_fence_finished(struct drm_sched_fence *fence, int result);
594 
595 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
596 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
597 		                unsigned long remaining);
598 struct drm_gpu_scheduler *
599 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
600 		     unsigned int num_sched_list);
601 
602 #endif
603