xref: /openbmc/linux/include/drm/gpu_scheduler.h (revision 25ebbc57ca56df3cf9149e9da6b1d3169c8487db)
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 enum dma_resv_usage;
45 struct dma_resv;
46 struct drm_gem_object;
47 
48 struct drm_gpu_scheduler;
49 struct drm_sched_rq;
50 
51 struct drm_file;
52 
53 /* These are often used as an (initial) index
54  * to an array, and as such should start at 0.
55  */
56 enum drm_sched_priority {
57 	DRM_SCHED_PRIORITY_MIN,
58 	DRM_SCHED_PRIORITY_NORMAL,
59 	DRM_SCHED_PRIORITY_HIGH,
60 	DRM_SCHED_PRIORITY_KERNEL,
61 
62 	DRM_SCHED_PRIORITY_COUNT,
63 	DRM_SCHED_PRIORITY_UNSET = -2
64 };
65 
66 /* Used to chose between FIFO and RR jobs scheduling */
67 extern int drm_sched_policy;
68 
69 #define DRM_SCHED_POLICY_RR    0
70 #define DRM_SCHED_POLICY_FIFO  1
71 
72 /**
73  * struct drm_sched_entity - A wrapper around a job queue (typically
74  * attached to the DRM file_priv).
75  *
76  * Entities will emit jobs in order to their corresponding hardware
77  * ring, and the scheduler will alternate between entities based on
78  * scheduling policy.
79  */
80 struct drm_sched_entity {
81 	/**
82 	 * @list:
83 	 *
84 	 * Used to append this struct to the list of entities in the runqueue
85 	 * @rq under &drm_sched_rq.entities.
86 	 *
87 	 * Protected by &drm_sched_rq.lock of @rq.
88 	 */
89 	struct list_head		list;
90 
91 	/**
92 	 * @rq:
93 	 *
94 	 * Runqueue on which this entity is currently scheduled.
95 	 *
96 	 * FIXME: Locking is very unclear for this. Writers are protected by
97 	 * @rq_lock, but readers are generally lockless and seem to just race
98 	 * with not even a READ_ONCE.
99 	 */
100 	struct drm_sched_rq		*rq;
101 
102 	/**
103 	 * @sched_list:
104 	 *
105 	 * A list of schedulers (struct drm_gpu_scheduler).  Jobs from this entity can
106 	 * be scheduled on any scheduler on this list.
107 	 *
108 	 * This can be modified by calling drm_sched_entity_modify_sched().
109 	 * Locking is entirely up to the driver, see the above function for more
110 	 * details.
111 	 *
112 	 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
113 	 * set already.
114 	 *
115 	 * FIXME: This means priority changes through
116 	 * drm_sched_entity_set_priority() will be lost henceforth in this case.
117 	 */
118 	struct drm_gpu_scheduler        **sched_list;
119 
120 	/**
121 	 * @num_sched_list:
122 	 *
123 	 * Number of drm_gpu_schedulers in the @sched_list.
124 	 */
125 	unsigned int                    num_sched_list;
126 
127 	/**
128 	 * @priority:
129 	 *
130 	 * Priority of the entity. This can be modified by calling
131 	 * drm_sched_entity_set_priority(). Protected by &rq_lock.
132 	 */
133 	enum drm_sched_priority         priority;
134 
135 	/**
136 	 * @rq_lock:
137 	 *
138 	 * Lock to modify the runqueue to which this entity belongs.
139 	 */
140 	spinlock_t			rq_lock;
141 
142 	/**
143 	 * @job_queue: the list of jobs of this entity.
144 	 */
145 	struct spsc_queue		job_queue;
146 
147 	/**
148 	 * @fence_seq:
149 	 *
150 	 * A linearly increasing seqno incremented with each new
151 	 * &drm_sched_fence which is part of the entity.
152 	 *
153 	 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
154 	 * this doesn't need to be atomic.
155 	 */
156 	atomic_t			fence_seq;
157 
158 	/**
159 	 * @fence_context:
160 	 *
161 	 * A unique context for all the fences which belong to this entity.  The
162 	 * &drm_sched_fence.scheduled uses the fence_context but
163 	 * &drm_sched_fence.finished uses fence_context + 1.
164 	 */
165 	uint64_t			fence_context;
166 
167 	/**
168 	 * @dependency:
169 	 *
170 	 * The dependency fence of the job which is on the top of the job queue.
171 	 */
172 	struct dma_fence		*dependency;
173 
174 	/**
175 	 * @cb:
176 	 *
177 	 * Callback for the dependency fence above.
178 	 */
179 	struct dma_fence_cb		cb;
180 
181 	/**
182 	 * @guilty:
183 	 *
184 	 * Points to entities' guilty.
185 	 */
186 	atomic_t			*guilty;
187 
188 	/**
189 	 * @last_scheduled:
190 	 *
191 	 * Points to the finished fence of the last scheduled job. Only written
192 	 * by the scheduler thread, can be accessed locklessly from
193 	 * drm_sched_job_arm() iff the queue is empty.
194 	 */
195 	struct dma_fence                *last_scheduled;
196 
197 	/**
198 	 * @last_user: last group leader pushing a job into the entity.
199 	 */
200 	struct task_struct		*last_user;
201 
202 	/**
203 	 * @stopped:
204 	 *
205 	 * Marks the enity as removed from rq and destined for
206 	 * termination. This is set by calling drm_sched_entity_flush() and by
207 	 * drm_sched_fini().
208 	 */
209 	bool 				stopped;
210 
211 	/**
212 	 * @entity_idle:
213 	 *
214 	 * Signals when entity is not in use, used to sequence entity cleanup in
215 	 * drm_sched_entity_fini().
216 	 */
217 	struct completion		entity_idle;
218 
219 	/**
220 	 * @oldest_job_waiting:
221 	 *
222 	 * Marks earliest job waiting in SW queue
223 	 */
224 	ktime_t				oldest_job_waiting;
225 
226 	/**
227 	 * @rb_tree_node:
228 	 *
229 	 * The node used to insert this entity into time based priority queue
230 	 */
231 	struct rb_node			rb_tree_node;
232 
233 	/**
234 	 * @elapsed_ns:
235 	 *
236 	 * Records the amount of time where jobs from this entity were active
237 	 * on the GPU.
238 	 */
239 	uint64_t elapsed_ns;
240 };
241 
242 /**
243  * struct drm_sched_rq - queue of entities to be scheduled.
244  *
245  * @lock: to modify the entities list.
246  * @sched: the scheduler to which this rq belongs to.
247  * @entities: list of the entities to be scheduled.
248  * @current_entity: the entity which is to be scheduled.
249  * @rb_tree_root: root of time based priory queue of entities for FIFO scheduling
250  *
251  * Run queue is a set of entities scheduling command submissions for
252  * one specific ring. It implements the scheduling policy that selects
253  * the next entity to emit commands from.
254  */
255 struct drm_sched_rq {
256 	spinlock_t			lock;
257 	struct drm_gpu_scheduler	*sched;
258 	struct list_head		entities;
259 	struct drm_sched_entity		*current_entity;
260 	struct rb_root_cached		rb_tree_root;
261 };
262 
263 /**
264  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
265  */
266 struct drm_sched_fence {
267         /**
268          * @scheduled: this fence is what will be signaled by the scheduler
269          * when the job is scheduled.
270          */
271 	struct dma_fence		scheduled;
272 
273         /**
274          * @finished: this fence is what will be signaled by the scheduler
275          * when the job is completed.
276          *
277          * When setting up an out fence for the job, you should use
278          * this, since it's available immediately upon
279          * drm_sched_job_init(), and the fence returned by the driver
280          * from run_job() won't be created until the dependencies have
281          * resolved.
282          */
283 	struct dma_fence		finished;
284 
285         /**
286          * @parent: the fence returned by &drm_sched_backend_ops.run_job
287          * when scheduling the job on hardware. We signal the
288          * &drm_sched_fence.finished fence once parent is signalled.
289          */
290 	struct dma_fence		*parent;
291         /**
292          * @sched: the scheduler instance to which the job having this struct
293          * belongs to.
294          */
295 	struct drm_gpu_scheduler	*sched;
296         /**
297          * @lock: the lock used by the scheduled and the finished fences.
298          */
299 	spinlock_t			lock;
300         /**
301          * @owner: job owner for debugging
302          */
303 	void				*owner;
304 };
305 
306 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
307 
308 /**
309  * struct drm_sched_job - A job to be run by an entity.
310  *
311  * @queue_node: used to append this struct to the queue of jobs in an entity.
312  * @list: a job participates in a "pending" and "done" lists.
313  * @sched: the scheduler instance on which this job is scheduled.
314  * @s_fence: contains the fences for the scheduling of job.
315  * @finish_cb: the callback for the finished fence.
316  * @work: Helper to reschdeule job kill to different context.
317  * @id: a unique id assigned to each job scheduled on the scheduler.
318  * @karma: increment on every hang caused by this job. If this exceeds the hang
319  *         limit of the scheduler then the job is marked guilty and will not
320  *         be scheduled further.
321  * @s_priority: the priority of the job.
322  * @entity: the entity to which this job belongs.
323  * @cb: the callback for the parent fence in s_fence.
324  *
325  * A job is created by the driver using drm_sched_job_init(), and
326  * should call drm_sched_entity_push_job() once it wants the scheduler
327  * to schedule the job.
328  */
329 struct drm_sched_job {
330 	struct spsc_node		queue_node;
331 	struct list_head		list;
332 	struct drm_gpu_scheduler	*sched;
333 	struct drm_sched_fence		*s_fence;
334 
335 	/*
336 	 * work is used only after finish_cb has been used and will not be
337 	 * accessed anymore.
338 	 */
339 	union {
340 		struct dma_fence_cb		finish_cb;
341 		struct work_struct		work;
342 	};
343 
344 	uint64_t			id;
345 	atomic_t			karma;
346 	enum drm_sched_priority		s_priority;
347 	struct drm_sched_entity         *entity;
348 	struct dma_fence_cb		cb;
349 	/**
350 	 * @dependencies:
351 	 *
352 	 * Contains the dependencies as struct dma_fence for this job, see
353 	 * drm_sched_job_add_dependency() and
354 	 * drm_sched_job_add_implicit_dependencies().
355 	 */
356 	struct xarray			dependencies;
357 
358 	/** @last_dependency: tracks @dependencies as they signal */
359 	unsigned long			last_dependency;
360 
361 	/**
362 	 * @submit_ts:
363 	 *
364 	 * When the job was pushed into the entity queue.
365 	 */
366 	ktime_t                         submit_ts;
367 };
368 
369 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
370 					    int threshold)
371 {
372 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
373 }
374 
375 enum drm_gpu_sched_stat {
376 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
377 	DRM_GPU_SCHED_STAT_NOMINAL,
378 	DRM_GPU_SCHED_STAT_ENODEV,
379 };
380 
381 /**
382  * struct drm_sched_backend_ops - Define the backend operations
383  *	called by the scheduler
384  *
385  * These functions should be implemented in the driver side.
386  */
387 struct drm_sched_backend_ops {
388 	/**
389 	 * @prepare_job:
390 	 *
391 	 * Called when the scheduler is considering scheduling this job next, to
392 	 * get another struct dma_fence for this job to block on.  Once it
393 	 * returns NULL, run_job() may be called.
394 	 *
395 	 * Can be NULL if no additional preparation to the dependencies are
396 	 * necessary. Skipped when jobs are killed instead of run.
397 	 */
398 	struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job,
399 					 struct drm_sched_entity *s_entity);
400 
401 	/**
402          * @run_job: Called to execute the job once all of the dependencies
403          * have been resolved.  This may be called multiple times, if
404 	 * timedout_job() has happened and drm_sched_job_recovery()
405 	 * decides to try it again.
406 	 */
407 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
408 
409 	/**
410 	 * @timedout_job: Called when a job has taken too long to execute,
411 	 * to trigger GPU recovery.
412 	 *
413 	 * This method is called in a workqueue context.
414 	 *
415 	 * Drivers typically issue a reset to recover from GPU hangs, and this
416 	 * procedure usually follows the following workflow:
417 	 *
418 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
419 	 *    scheduler thread and cancel the timeout work, guaranteeing that
420 	 *    nothing is queued while we reset the hardware queue
421 	 * 2. Try to gracefully stop non-faulty jobs (optional)
422 	 * 3. Issue a GPU reset (driver-specific)
423 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
424 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
425 	 *    jobs can be queued, and the scheduler thread is unblocked
426 	 *
427 	 * Note that some GPUs have distinct hardware queues but need to reset
428 	 * the GPU globally, which requires extra synchronization between the
429 	 * timeout handler of the different &drm_gpu_scheduler. One way to
430 	 * achieve this synchronization is to create an ordered workqueue
431 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
432 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
433 	 * executed sequentially. The above workflow needs to be slightly
434 	 * adjusted in that case:
435 	 *
436 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
437 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
438 	 *    the reset (optional)
439 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
440 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
441 	 *    drm_sched_resubmit_jobs()
442 	 * 5. Restart all schedulers that were stopped in step #1 using
443 	 *    drm_sched_start()
444 	 *
445 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
446 	 * and the underlying driver has started or completed recovery.
447 	 *
448 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
449 	 * available, i.e. has been unplugged.
450 	 */
451 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
452 
453 	/**
454          * @free_job: Called once the job's finished fence has been signaled
455          * and it's time to clean it up.
456 	 */
457 	void (*free_job)(struct drm_sched_job *sched_job);
458 };
459 
460 /**
461  * struct drm_gpu_scheduler - scheduler instance-specific data
462  *
463  * @ops: backend operations provided by the driver.
464  * @hw_submission_limit: the max size of the hardware queue.
465  * @timeout: the time after which a job is removed from the scheduler.
466  * @name: name of the ring for which this scheduler is being used.
467  * @sched_rq: priority wise array of run queues.
468  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
469  *                  is ready to be scheduled.
470  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
471  *                 waits on this wait queue until all the scheduled jobs are
472  *                 finished.
473  * @hw_rq_count: the number of jobs currently in the hardware queue.
474  * @job_id_count: used to assign unique id to the each job.
475  * @timeout_wq: workqueue used to queue @work_tdr
476  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
477  *            timeout interval is over.
478  * @thread: the kthread on which the scheduler which run.
479  * @pending_list: the list of jobs which are currently in the job queue.
480  * @job_list_lock: lock to protect the pending_list.
481  * @hang_limit: once the hangs by a job crosses this limit then it is marked
482  *              guilty and it will no longer be considered for scheduling.
483  * @score: score to help loadbalancer pick a idle sched
484  * @_score: score used when the driver doesn't provide one
485  * @ready: marks if the underlying HW is ready to work
486  * @free_guilty: A hit to time out handler to free the guilty job.
487  * @dev: system &struct device
488  *
489  * One scheduler is implemented for each hardware ring.
490  */
491 struct drm_gpu_scheduler {
492 	const struct drm_sched_backend_ops	*ops;
493 	uint32_t			hw_submission_limit;
494 	long				timeout;
495 	const char			*name;
496 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
497 	wait_queue_head_t		wake_up_worker;
498 	wait_queue_head_t		job_scheduled;
499 	atomic_t			hw_rq_count;
500 	atomic64_t			job_id_count;
501 	struct workqueue_struct		*timeout_wq;
502 	struct delayed_work		work_tdr;
503 	struct task_struct		*thread;
504 	struct list_head		pending_list;
505 	spinlock_t			job_list_lock;
506 	int				hang_limit;
507 	atomic_t                        *score;
508 	atomic_t                        _score;
509 	bool				ready;
510 	bool				free_guilty;
511 	struct device			*dev;
512 };
513 
514 int drm_sched_init(struct drm_gpu_scheduler *sched,
515 		   const struct drm_sched_backend_ops *ops,
516 		   uint32_t hw_submission, unsigned hang_limit,
517 		   long timeout, struct workqueue_struct *timeout_wq,
518 		   atomic_t *score, const char *name, struct device *dev);
519 
520 void drm_sched_fini(struct drm_gpu_scheduler *sched);
521 int drm_sched_job_init(struct drm_sched_job *job,
522 		       struct drm_sched_entity *entity,
523 		       void *owner);
524 void drm_sched_job_arm(struct drm_sched_job *job);
525 int drm_sched_job_add_dependency(struct drm_sched_job *job,
526 				 struct dma_fence *fence);
527 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
528 					 struct drm_file *file,
529 					 u32 handle,
530 					 u32 point);
531 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
532 					struct dma_resv *resv,
533 					enum dma_resv_usage usage);
534 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
535 					    struct drm_gem_object *obj,
536 					    bool write);
537 
538 
539 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
540 				    struct drm_gpu_scheduler **sched_list,
541                                    unsigned int num_sched_list);
542 
543 void drm_sched_job_cleanup(struct drm_sched_job *job);
544 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
545 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
546 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
547 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
548 void drm_sched_increase_karma(struct drm_sched_job *bad);
549 void drm_sched_reset_karma(struct drm_sched_job *bad);
550 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
551 bool drm_sched_dependency_optimized(struct dma_fence* fence,
552 				    struct drm_sched_entity *entity);
553 void drm_sched_fault(struct drm_gpu_scheduler *sched);
554 
555 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
556 			     struct drm_sched_entity *entity);
557 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
558 				struct drm_sched_entity *entity);
559 
560 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts);
561 
562 int drm_sched_entity_init(struct drm_sched_entity *entity,
563 			  enum drm_sched_priority priority,
564 			  struct drm_gpu_scheduler **sched_list,
565 			  unsigned int num_sched_list,
566 			  atomic_t *guilty);
567 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
568 void drm_sched_entity_fini(struct drm_sched_entity *entity);
569 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
570 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
571 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
572 void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
573 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
574 				   enum drm_sched_priority priority);
575 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
576 
577 struct drm_sched_fence *drm_sched_fence_alloc(
578 	struct drm_sched_entity *s_entity, void *owner);
579 void drm_sched_fence_init(struct drm_sched_fence *fence,
580 			  struct drm_sched_entity *entity);
581 void drm_sched_fence_free(struct drm_sched_fence *fence);
582 
583 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
584 void drm_sched_fence_finished(struct drm_sched_fence *fence);
585 
586 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
587 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
588 		                unsigned long remaining);
589 struct drm_gpu_scheduler *
590 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
591 		     unsigned int num_sched_list);
592 
593 #endif
594