xref: /openbmc/linux/include/drm/gpu_scheduler.h (revision 47ebd031)
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 /* These are often used as an (initial) index
52  * to an array, and as such should start at 0.
53  */
54 enum drm_sched_priority {
55 	DRM_SCHED_PRIORITY_MIN,
56 	DRM_SCHED_PRIORITY_NORMAL,
57 	DRM_SCHED_PRIORITY_HIGH,
58 	DRM_SCHED_PRIORITY_KERNEL,
59 
60 	DRM_SCHED_PRIORITY_COUNT,
61 	DRM_SCHED_PRIORITY_UNSET = -2
62 };
63 
64 /* Used to chose between FIFO and RR jobs scheduling */
65 extern int drm_sched_policy;
66 
67 #define DRM_SCHED_POLICY_RR    0
68 #define DRM_SCHED_POLICY_FIFO  1
69 
70 /**
71  * struct drm_sched_entity - A wrapper around a job queue (typically
72  * attached to the DRM file_priv).
73  *
74  * Entities will emit jobs in order to their corresponding hardware
75  * ring, and the scheduler will alternate between entities based on
76  * scheduling policy.
77  */
78 struct drm_sched_entity {
79 	/**
80 	 * @list:
81 	 *
82 	 * Used to append this struct to the list of entities in the runqueue
83 	 * @rq under &drm_sched_rq.entities.
84 	 *
85 	 * Protected by &drm_sched_rq.lock of @rq.
86 	 */
87 	struct list_head		list;
88 
89 	/**
90 	 * @rq:
91 	 *
92 	 * Runqueue on which this entity is currently scheduled.
93 	 *
94 	 * FIXME: Locking is very unclear for this. Writers are protected by
95 	 * @rq_lock, but readers are generally lockless and seem to just race
96 	 * with not even a READ_ONCE.
97 	 */
98 	struct drm_sched_rq		*rq;
99 
100 	/**
101 	 * @sched_list:
102 	 *
103 	 * A list of schedulers (struct drm_gpu_scheduler).  Jobs from this entity can
104 	 * be scheduled on any scheduler on this list.
105 	 *
106 	 * This can be modified by calling drm_sched_entity_modify_sched().
107 	 * Locking is entirely up to the driver, see the above function for more
108 	 * details.
109 	 *
110 	 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
111 	 * set already.
112 	 *
113 	 * FIXME: This means priority changes through
114 	 * drm_sched_entity_set_priority() will be lost henceforth in this case.
115 	 */
116 	struct drm_gpu_scheduler        **sched_list;
117 
118 	/**
119 	 * @num_sched_list:
120 	 *
121 	 * Number of drm_gpu_schedulers in the @sched_list.
122 	 */
123 	unsigned int                    num_sched_list;
124 
125 	/**
126 	 * @priority:
127 	 *
128 	 * Priority of the entity. This can be modified by calling
129 	 * drm_sched_entity_set_priority(). Protected by &rq_lock.
130 	 */
131 	enum drm_sched_priority         priority;
132 
133 	/**
134 	 * @rq_lock:
135 	 *
136 	 * Lock to modify the runqueue to which this entity belongs.
137 	 */
138 	spinlock_t			rq_lock;
139 
140 	/**
141 	 * @job_queue: the list of jobs of this entity.
142 	 */
143 	struct spsc_queue		job_queue;
144 
145 	/**
146 	 * @fence_seq:
147 	 *
148 	 * A linearly increasing seqno incremented with each new
149 	 * &drm_sched_fence which is part of the entity.
150 	 *
151 	 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
152 	 * this doesn't need to be atomic.
153 	 */
154 	atomic_t			fence_seq;
155 
156 	/**
157 	 * @fence_context:
158 	 *
159 	 * A unique context for all the fences which belong to this entity.  The
160 	 * &drm_sched_fence.scheduled uses the fence_context but
161 	 * &drm_sched_fence.finished uses fence_context + 1.
162 	 */
163 	uint64_t			fence_context;
164 
165 	/**
166 	 * @dependency:
167 	 *
168 	 * The dependency fence of the job which is on the top of the job queue.
169 	 */
170 	struct dma_fence		*dependency;
171 
172 	/**
173 	 * @cb:
174 	 *
175 	 * Callback for the dependency fence above.
176 	 */
177 	struct dma_fence_cb		cb;
178 
179 	/**
180 	 * @guilty:
181 	 *
182 	 * Points to entities' guilty.
183 	 */
184 	atomic_t			*guilty;
185 
186 	/**
187 	 * @last_scheduled:
188 	 *
189 	 * Points to the finished fence of the last scheduled job. Only written
190 	 * by the scheduler thread, can be accessed locklessly from
191 	 * drm_sched_job_arm() iff the queue is empty.
192 	 */
193 	struct dma_fence                *last_scheduled;
194 
195 	/**
196 	 * @last_user: last group leader pushing a job into the entity.
197 	 */
198 	struct task_struct		*last_user;
199 
200 	/**
201 	 * @stopped:
202 	 *
203 	 * Marks the enity as removed from rq and destined for
204 	 * termination. This is set by calling drm_sched_entity_flush() and by
205 	 * drm_sched_fini().
206 	 */
207 	bool 				stopped;
208 
209 	/**
210 	 * @entity_idle:
211 	 *
212 	 * Signals when entity is not in use, used to sequence entity cleanup in
213 	 * drm_sched_entity_fini().
214 	 */
215 	struct completion		entity_idle;
216 
217 	/**
218 	 * @oldest_job_waiting:
219 	 *
220 	 * Marks earliest job waiting in SW queue
221 	 */
222 	ktime_t				oldest_job_waiting;
223 
224 	/**
225 	 * @rb_tree_node:
226 	 *
227 	 * The node used to insert this entity into time based priority queue
228 	 */
229 	struct rb_node			rb_tree_node;
230 
231 	/**
232 	 * @elapsed_ns:
233 	 *
234 	 * Records the amount of time where jobs from this entity were active
235 	 * on the GPU.
236 	 */
237 	uint64_t elapsed_ns;
238 };
239 
240 /**
241  * struct drm_sched_rq - queue of entities to be scheduled.
242  *
243  * @lock: to modify the entities list.
244  * @sched: the scheduler to which this rq belongs to.
245  * @entities: list of the entities to be scheduled.
246  * @current_entity: the entity which is to be scheduled.
247  * @rb_tree_root: root of time based priory queue of entities for FIFO scheduling
248  *
249  * Run queue is a set of entities scheduling command submissions for
250  * one specific ring. It implements the scheduling policy that selects
251  * the next entity to emit commands from.
252  */
253 struct drm_sched_rq {
254 	spinlock_t			lock;
255 	struct drm_gpu_scheduler	*sched;
256 	struct list_head		entities;
257 	struct drm_sched_entity		*current_entity;
258 	struct rb_root_cached		rb_tree_root;
259 };
260 
261 /**
262  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
263  */
264 struct drm_sched_fence {
265         /**
266          * @scheduled: this fence is what will be signaled by the scheduler
267          * when the job is scheduled.
268          */
269 	struct dma_fence		scheduled;
270 
271         /**
272          * @finished: this fence is what will be signaled by the scheduler
273          * when the job is completed.
274          *
275          * When setting up an out fence for the job, you should use
276          * this, since it's available immediately upon
277          * drm_sched_job_init(), and the fence returned by the driver
278          * from run_job() won't be created until the dependencies have
279          * resolved.
280          */
281 	struct dma_fence		finished;
282 
283         /**
284          * @parent: the fence returned by &drm_sched_backend_ops.run_job
285          * when scheduling the job on hardware. We signal the
286          * &drm_sched_fence.finished fence once parent is signalled.
287          */
288 	struct dma_fence		*parent;
289         /**
290          * @sched: the scheduler instance to which the job having this struct
291          * belongs to.
292          */
293 	struct drm_gpu_scheduler	*sched;
294         /**
295          * @lock: the lock used by the scheduled and the finished fences.
296          */
297 	spinlock_t			lock;
298         /**
299          * @owner: job owner for debugging
300          */
301 	void				*owner;
302 };
303 
304 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
305 
306 /**
307  * struct drm_sched_job - A job to be run by an entity.
308  *
309  * @queue_node: used to append this struct to the queue of jobs in an entity.
310  * @list: a job participates in a "pending" and "done" lists.
311  * @sched: the scheduler instance on which this job is scheduled.
312  * @s_fence: contains the fences for the scheduling of job.
313  * @finish_cb: the callback for the finished fence.
314  * @work: Helper to reschdeule job kill to different context.
315  * @id: a unique id assigned to each job scheduled on the scheduler.
316  * @karma: increment on every hang caused by this job. If this exceeds the hang
317  *         limit of the scheduler then the job is marked guilty and will not
318  *         be scheduled further.
319  * @s_priority: the priority of the job.
320  * @entity: the entity to which this job belongs.
321  * @cb: the callback for the parent fence in s_fence.
322  *
323  * A job is created by the driver using drm_sched_job_init(), and
324  * should call drm_sched_entity_push_job() once it wants the scheduler
325  * to schedule the job.
326  */
327 struct drm_sched_job {
328 	struct spsc_node		queue_node;
329 	struct list_head		list;
330 	struct drm_gpu_scheduler	*sched;
331 	struct drm_sched_fence		*s_fence;
332 
333 	/*
334 	 * work is used only after finish_cb has been used and will not be
335 	 * accessed anymore.
336 	 */
337 	union {
338 		struct dma_fence_cb		finish_cb;
339 		struct work_struct		work;
340 	};
341 
342 	uint64_t			id;
343 	atomic_t			karma;
344 	enum drm_sched_priority		s_priority;
345 	struct drm_sched_entity         *entity;
346 	struct dma_fence_cb		cb;
347 	/**
348 	 * @dependencies:
349 	 *
350 	 * Contains the dependencies as struct dma_fence for this job, see
351 	 * drm_sched_job_add_dependency() and
352 	 * drm_sched_job_add_implicit_dependencies().
353 	 */
354 	struct xarray			dependencies;
355 
356 	/** @last_dependency: tracks @dependencies as they signal */
357 	unsigned long			last_dependency;
358 
359 	/**
360 	 * @submit_ts:
361 	 *
362 	 * When the job was pushed into the entity queue.
363 	 */
364 	ktime_t                         submit_ts;
365 };
366 
367 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
368 					    int threshold)
369 {
370 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
371 }
372 
373 enum drm_gpu_sched_stat {
374 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
375 	DRM_GPU_SCHED_STAT_NOMINAL,
376 	DRM_GPU_SCHED_STAT_ENODEV,
377 };
378 
379 /**
380  * struct drm_sched_backend_ops - Define the backend operations
381  *	called by the scheduler
382  *
383  * These functions should be implemented in the driver side.
384  */
385 struct drm_sched_backend_ops {
386 	/**
387 	 * @prepare_job:
388 	 *
389 	 * Called when the scheduler is considering scheduling this job next, to
390 	 * get another struct dma_fence for this job to block on.  Once it
391 	 * returns NULL, run_job() may be called.
392 	 *
393 	 * Can be NULL if no additional preparation to the dependencies are
394 	 * necessary. Skipped when jobs are killed instead of run.
395 	 */
396 	struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job,
397 					 struct drm_sched_entity *s_entity);
398 
399 	/**
400          * @run_job: Called to execute the job once all of the dependencies
401          * have been resolved.  This may be called multiple times, if
402 	 * timedout_job() has happened and drm_sched_job_recovery()
403 	 * decides to try it again.
404 	 */
405 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
406 
407 	/**
408 	 * @timedout_job: Called when a job has taken too long to execute,
409 	 * to trigger GPU recovery.
410 	 *
411 	 * This method is called in a workqueue context.
412 	 *
413 	 * Drivers typically issue a reset to recover from GPU hangs, and this
414 	 * procedure usually follows the following workflow:
415 	 *
416 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
417 	 *    scheduler thread and cancel the timeout work, guaranteeing that
418 	 *    nothing is queued while we reset the hardware queue
419 	 * 2. Try to gracefully stop non-faulty jobs (optional)
420 	 * 3. Issue a GPU reset (driver-specific)
421 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
422 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
423 	 *    jobs can be queued, and the scheduler thread is unblocked
424 	 *
425 	 * Note that some GPUs have distinct hardware queues but need to reset
426 	 * the GPU globally, which requires extra synchronization between the
427 	 * timeout handler of the different &drm_gpu_scheduler. One way to
428 	 * achieve this synchronization is to create an ordered workqueue
429 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
430 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
431 	 * executed sequentially. The above workflow needs to be slightly
432 	 * adjusted in that case:
433 	 *
434 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
435 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
436 	 *    the reset (optional)
437 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
438 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
439 	 *    drm_sched_resubmit_jobs()
440 	 * 5. Restart all schedulers that were stopped in step #1 using
441 	 *    drm_sched_start()
442 	 *
443 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
444 	 * and the underlying driver has started or completed recovery.
445 	 *
446 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
447 	 * available, i.e. has been unplugged.
448 	 */
449 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
450 
451 	/**
452          * @free_job: Called once the job's finished fence has been signaled
453          * and it's time to clean it up.
454 	 */
455 	void (*free_job)(struct drm_sched_job *sched_job);
456 };
457 
458 /**
459  * struct drm_gpu_scheduler - scheduler instance-specific data
460  *
461  * @ops: backend operations provided by the driver.
462  * @hw_submission_limit: the max size of the hardware queue.
463  * @timeout: the time after which a job is removed from the scheduler.
464  * @name: name of the ring for which this scheduler is being used.
465  * @sched_rq: priority wise array of run queues.
466  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
467  *                  is ready to be scheduled.
468  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
469  *                 waits on this wait queue until all the scheduled jobs are
470  *                 finished.
471  * @hw_rq_count: the number of jobs currently in the hardware queue.
472  * @job_id_count: used to assign unique id to the each job.
473  * @timeout_wq: workqueue used to queue @work_tdr
474  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
475  *            timeout interval is over.
476  * @thread: the kthread on which the scheduler which run.
477  * @pending_list: the list of jobs which are currently in the job queue.
478  * @job_list_lock: lock to protect the pending_list.
479  * @hang_limit: once the hangs by a job crosses this limit then it is marked
480  *              guilty and it will no longer be considered for scheduling.
481  * @score: score to help loadbalancer pick a idle sched
482  * @_score: score used when the driver doesn't provide one
483  * @ready: marks if the underlying HW is ready to work
484  * @free_guilty: A hit to time out handler to free the guilty job.
485  * @dev: system &struct device
486  *
487  * One scheduler is implemented for each hardware ring.
488  */
489 struct drm_gpu_scheduler {
490 	const struct drm_sched_backend_ops	*ops;
491 	uint32_t			hw_submission_limit;
492 	long				timeout;
493 	const char			*name;
494 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
495 	wait_queue_head_t		wake_up_worker;
496 	wait_queue_head_t		job_scheduled;
497 	atomic_t			hw_rq_count;
498 	atomic64_t			job_id_count;
499 	struct workqueue_struct		*timeout_wq;
500 	struct delayed_work		work_tdr;
501 	struct task_struct		*thread;
502 	struct list_head		pending_list;
503 	spinlock_t			job_list_lock;
504 	int				hang_limit;
505 	atomic_t                        *score;
506 	atomic_t                        _score;
507 	bool				ready;
508 	bool				free_guilty;
509 	struct device			*dev;
510 };
511 
512 int drm_sched_init(struct drm_gpu_scheduler *sched,
513 		   const struct drm_sched_backend_ops *ops,
514 		   uint32_t hw_submission, unsigned hang_limit,
515 		   long timeout, struct workqueue_struct *timeout_wq,
516 		   atomic_t *score, const char *name, struct device *dev);
517 
518 void drm_sched_fini(struct drm_gpu_scheduler *sched);
519 int drm_sched_job_init(struct drm_sched_job *job,
520 		       struct drm_sched_entity *entity,
521 		       void *owner);
522 void drm_sched_job_arm(struct drm_sched_job *job);
523 int drm_sched_job_add_dependency(struct drm_sched_job *job,
524 				 struct dma_fence *fence);
525 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
526 					struct dma_resv *resv,
527 					enum dma_resv_usage usage);
528 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
529 					    struct drm_gem_object *obj,
530 					    bool write);
531 
532 
533 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
534 				    struct drm_gpu_scheduler **sched_list,
535                                    unsigned int num_sched_list);
536 
537 void drm_sched_job_cleanup(struct drm_sched_job *job);
538 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
539 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
540 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
541 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
542 void drm_sched_increase_karma(struct drm_sched_job *bad);
543 void drm_sched_reset_karma(struct drm_sched_job *bad);
544 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
545 bool drm_sched_dependency_optimized(struct dma_fence* fence,
546 				    struct drm_sched_entity *entity);
547 void drm_sched_fault(struct drm_gpu_scheduler *sched);
548 
549 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
550 			     struct drm_sched_entity *entity);
551 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
552 				struct drm_sched_entity *entity);
553 
554 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts);
555 
556 int drm_sched_entity_init(struct drm_sched_entity *entity,
557 			  enum drm_sched_priority priority,
558 			  struct drm_gpu_scheduler **sched_list,
559 			  unsigned int num_sched_list,
560 			  atomic_t *guilty);
561 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
562 void drm_sched_entity_fini(struct drm_sched_entity *entity);
563 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
564 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
565 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
566 void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
567 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
568 				   enum drm_sched_priority priority);
569 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
570 
571 struct drm_sched_fence *drm_sched_fence_alloc(
572 	struct drm_sched_entity *s_entity, void *owner);
573 void drm_sched_fence_init(struct drm_sched_fence *fence,
574 			  struct drm_sched_entity *entity);
575 void drm_sched_fence_free(struct drm_sched_fence *fence);
576 
577 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
578 void drm_sched_fence_finished(struct drm_sched_fence *fence);
579 
580 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
581 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
582 		                unsigned long remaining);
583 struct drm_gpu_scheduler *
584 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
585 		     unsigned int num_sched_list);
586 
587 #endif
588