xref: /openbmc/linux/include/drm/gpu_scheduler.h (revision 83b975b5)
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 struct drm_gem_object;
36 
37 struct drm_gpu_scheduler;
38 struct drm_sched_rq;
39 
40 /* These are often used as an (initial) index
41  * to an array, and as such should start at 0.
42  */
43 enum drm_sched_priority {
44 	DRM_SCHED_PRIORITY_MIN,
45 	DRM_SCHED_PRIORITY_NORMAL,
46 	DRM_SCHED_PRIORITY_HIGH,
47 	DRM_SCHED_PRIORITY_KERNEL,
48 
49 	DRM_SCHED_PRIORITY_COUNT,
50 	DRM_SCHED_PRIORITY_UNSET = -2
51 };
52 
53 /**
54  * struct drm_sched_entity - A wrapper around a job queue (typically
55  * attached to the DRM file_priv).
56  *
57  * Entities will emit jobs in order to their corresponding hardware
58  * ring, and the scheduler will alternate between entities based on
59  * scheduling policy.
60  */
61 struct drm_sched_entity {
62 	/**
63 	 * @list:
64 	 *
65 	 * Used to append this struct to the list of entities in the runqueue
66 	 * @rq under &drm_sched_rq.entities.
67 	 *
68 	 * Protected by &drm_sched_rq.lock of @rq.
69 	 */
70 	struct list_head		list;
71 
72 	/**
73 	 * @rq:
74 	 *
75 	 * Runqueue on which this entity is currently scheduled.
76 	 *
77 	 * FIXME: Locking is very unclear for this. Writers are protected by
78 	 * @rq_lock, but readers are generally lockless and seem to just race
79 	 * with not even a READ_ONCE.
80 	 */
81 	struct drm_sched_rq		*rq;
82 
83 	/**
84 	 * @sched_list:
85 	 *
86 	 * A list of schedulers (struct drm_gpu_scheduler).  Jobs from this entity can
87 	 * be scheduled on any scheduler on this list.
88 	 *
89 	 * This can be modified by calling drm_sched_entity_modify_sched().
90 	 * Locking is entirely up to the driver, see the above function for more
91 	 * details.
92 	 *
93 	 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
94 	 * set already.
95 	 *
96 	 * FIXME: This means priority changes through
97 	 * drm_sched_entity_set_priority() will be lost henceforth in this case.
98 	 */
99 	struct drm_gpu_scheduler        **sched_list;
100 
101 	/**
102 	 * @num_sched_list:
103 	 *
104 	 * Number of drm_gpu_schedulers in the @sched_list.
105 	 */
106 	unsigned int                    num_sched_list;
107 
108 	/**
109 	 * @priority:
110 	 *
111 	 * Priority of the entity. This can be modified by calling
112 	 * drm_sched_entity_set_priority(). Protected by &rq_lock.
113 	 */
114 	enum drm_sched_priority         priority;
115 
116 	/**
117 	 * @rq_lock:
118 	 *
119 	 * Lock to modify the runqueue to which this entity belongs.
120 	 */
121 	spinlock_t			rq_lock;
122 
123 	/**
124 	 * @job_queue: the list of jobs of this entity.
125 	 */
126 	struct spsc_queue		job_queue;
127 
128 	/**
129 	 * @fence_seq:
130 	 *
131 	 * A linearly increasing seqno incremented with each new
132 	 * &drm_sched_fence which is part of the entity.
133 	 *
134 	 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
135 	 * this doesn't need to be atomic.
136 	 */
137 	atomic_t			fence_seq;
138 
139 	/**
140 	 * @fence_context:
141 	 *
142 	 * A unique context for all the fences which belong to this entity.  The
143 	 * &drm_sched_fence.scheduled uses the fence_context but
144 	 * &drm_sched_fence.finished uses fence_context + 1.
145 	 */
146 	uint64_t			fence_context;
147 
148 	/**
149 	 * @dependency:
150 	 *
151 	 * The dependency fence of the job which is on the top of the job queue.
152 	 */
153 	struct dma_fence		*dependency;
154 
155 	/**
156 	 * @cb:
157 	 *
158 	 * Callback for the dependency fence above.
159 	 */
160 	struct dma_fence_cb		cb;
161 
162 	/**
163 	 * @guilty:
164 	 *
165 	 * Points to entities' guilty.
166 	 */
167 	atomic_t			*guilty;
168 
169 	/**
170 	 * @last_scheduled:
171 	 *
172 	 * Points to the finished fence of the last scheduled job. Only written
173 	 * by the scheduler thread, can be accessed locklessly from
174 	 * drm_sched_job_arm() iff the queue is empty.
175 	 */
176 	struct dma_fence                *last_scheduled;
177 
178 	/**
179 	 * @last_user: last group leader pushing a job into the entity.
180 	 */
181 	struct task_struct		*last_user;
182 
183 	/**
184 	 * @stopped:
185 	 *
186 	 * Marks the enity as removed from rq and destined for
187 	 * termination. This is set by calling drm_sched_entity_flush() and by
188 	 * drm_sched_fini().
189 	 */
190 	bool 				stopped;
191 
192 	/**
193 	 * @entity_idle:
194 	 *
195 	 * Signals when entity is not in use, used to sequence entity cleanup in
196 	 * drm_sched_entity_fini().
197 	 */
198 	struct completion		entity_idle;
199 };
200 
201 /**
202  * struct drm_sched_rq - queue of entities to be scheduled.
203  *
204  * @lock: to modify the entities list.
205  * @sched: the scheduler to which this rq belongs to.
206  * @entities: list of the entities to be scheduled.
207  * @current_entity: the entity which is to be scheduled.
208  *
209  * Run queue is a set of entities scheduling command submissions for
210  * one specific ring. It implements the scheduling policy that selects
211  * the next entity to emit commands from.
212  */
213 struct drm_sched_rq {
214 	spinlock_t			lock;
215 	struct drm_gpu_scheduler	*sched;
216 	struct list_head		entities;
217 	struct drm_sched_entity		*current_entity;
218 };
219 
220 /**
221  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
222  */
223 struct drm_sched_fence {
224         /**
225          * @scheduled: this fence is what will be signaled by the scheduler
226          * when the job is scheduled.
227          */
228 	struct dma_fence		scheduled;
229 
230         /**
231          * @finished: this fence is what will be signaled by the scheduler
232          * when the job is completed.
233          *
234          * When setting up an out fence for the job, you should use
235          * this, since it's available immediately upon
236          * drm_sched_job_init(), and the fence returned by the driver
237          * from run_job() won't be created until the dependencies have
238          * resolved.
239          */
240 	struct dma_fence		finished;
241 
242         /**
243          * @parent: the fence returned by &drm_sched_backend_ops.run_job
244          * when scheduling the job on hardware. We signal the
245          * &drm_sched_fence.finished fence once parent is signalled.
246          */
247 	struct dma_fence		*parent;
248         /**
249          * @sched: the scheduler instance to which the job having this struct
250          * belongs to.
251          */
252 	struct drm_gpu_scheduler	*sched;
253         /**
254          * @lock: the lock used by the scheduled and the finished fences.
255          */
256 	spinlock_t			lock;
257         /**
258          * @owner: job owner for debugging
259          */
260 	void				*owner;
261 };
262 
263 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
264 
265 /**
266  * struct drm_sched_job - A job to be run by an entity.
267  *
268  * @queue_node: used to append this struct to the queue of jobs in an entity.
269  * @list: a job participates in a "pending" and "done" lists.
270  * @sched: the scheduler instance on which this job is scheduled.
271  * @s_fence: contains the fences for the scheduling of job.
272  * @finish_cb: the callback for the finished fence.
273  * @work: Helper to reschdeule job kill to different context.
274  * @id: a unique id assigned to each job scheduled on the scheduler.
275  * @karma: increment on every hang caused by this job. If this exceeds the hang
276  *         limit of the scheduler then the job is marked guilty and will not
277  *         be scheduled further.
278  * @s_priority: the priority of the job.
279  * @entity: the entity to which this job belongs.
280  * @cb: the callback for the parent fence in s_fence.
281  *
282  * A job is created by the driver using drm_sched_job_init(), and
283  * should call drm_sched_entity_push_job() once it wants the scheduler
284  * to schedule the job.
285  */
286 struct drm_sched_job {
287 	struct spsc_node		queue_node;
288 	struct list_head		list;
289 	struct drm_gpu_scheduler	*sched;
290 	struct drm_sched_fence		*s_fence;
291 
292 	/*
293 	 * work is used only after finish_cb has been used and will not be
294 	 * accessed anymore.
295 	 */
296 	union {
297 		struct dma_fence_cb		finish_cb;
298 		struct work_struct 		work;
299 	};
300 
301 	uint64_t			id;
302 	atomic_t			karma;
303 	enum drm_sched_priority		s_priority;
304 	struct drm_sched_entity         *entity;
305 	struct dma_fence_cb		cb;
306 	/**
307 	 * @dependencies:
308 	 *
309 	 * Contains the dependencies as struct dma_fence for this job, see
310 	 * drm_sched_job_add_dependency() and
311 	 * drm_sched_job_add_implicit_dependencies().
312 	 */
313 	struct xarray			dependencies;
314 
315 	/** @last_dependency: tracks @dependencies as they signal */
316 	unsigned long			last_dependency;
317 };
318 
319 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
320 					    int threshold)
321 {
322 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
323 }
324 
325 enum drm_gpu_sched_stat {
326 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
327 	DRM_GPU_SCHED_STAT_NOMINAL,
328 	DRM_GPU_SCHED_STAT_ENODEV,
329 };
330 
331 /**
332  * struct drm_sched_backend_ops - Define the backend operations
333  *	called by the scheduler
334  *
335  * These functions should be implemented in the driver side.
336  */
337 struct drm_sched_backend_ops {
338 	/**
339 	 * @dependency:
340 	 *
341 	 * Called when the scheduler is considering scheduling this job next, to
342 	 * get another struct dma_fence for this job to block on.  Once it
343 	 * returns NULL, run_job() may be called.
344 	 *
345 	 * If a driver exclusively uses drm_sched_job_add_dependency() and
346 	 * drm_sched_job_add_implicit_dependencies() this can be ommitted and
347 	 * left as NULL.
348 	 */
349 	struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
350 					struct drm_sched_entity *s_entity);
351 
352 	/**
353          * @run_job: Called to execute the job once all of the dependencies
354          * have been resolved.  This may be called multiple times, if
355 	 * timedout_job() has happened and drm_sched_job_recovery()
356 	 * decides to try it again.
357 	 */
358 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
359 
360 	/**
361 	 * @timedout_job: Called when a job has taken too long to execute,
362 	 * to trigger GPU recovery.
363 	 *
364 	 * This method is called in a workqueue context.
365 	 *
366 	 * Drivers typically issue a reset to recover from GPU hangs, and this
367 	 * procedure usually follows the following workflow:
368 	 *
369 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
370 	 *    scheduler thread and cancel the timeout work, guaranteeing that
371 	 *    nothing is queued while we reset the hardware queue
372 	 * 2. Try to gracefully stop non-faulty jobs (optional)
373 	 * 3. Issue a GPU reset (driver-specific)
374 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
375 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
376 	 *    jobs can be queued, and the scheduler thread is unblocked
377 	 *
378 	 * Note that some GPUs have distinct hardware queues but need to reset
379 	 * the GPU globally, which requires extra synchronization between the
380 	 * timeout handler of the different &drm_gpu_scheduler. One way to
381 	 * achieve this synchronization is to create an ordered workqueue
382 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
383 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
384 	 * executed sequentially. The above workflow needs to be slightly
385 	 * adjusted in that case:
386 	 *
387 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
388 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
389 	 *    the reset (optional)
390 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
391 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
392 	 *    drm_sched_resubmit_jobs()
393 	 * 5. Restart all schedulers that were stopped in step #1 using
394 	 *    drm_sched_start()
395 	 *
396 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
397 	 * and the underlying driver has started or completed recovery.
398 	 *
399 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
400 	 * available, i.e. has been unplugged.
401 	 */
402 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
403 
404 	/**
405          * @free_job: Called once the job's finished fence has been signaled
406          * and it's time to clean it up.
407 	 */
408 	void (*free_job)(struct drm_sched_job *sched_job);
409 };
410 
411 /**
412  * struct drm_gpu_scheduler - scheduler instance-specific data
413  *
414  * @ops: backend operations provided by the driver.
415  * @hw_submission_limit: the max size of the hardware queue.
416  * @timeout: the time after which a job is removed from the scheduler.
417  * @name: name of the ring for which this scheduler is being used.
418  * @sched_rq: priority wise array of run queues.
419  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
420  *                  is ready to be scheduled.
421  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
422  *                 waits on this wait queue until all the scheduled jobs are
423  *                 finished.
424  * @hw_rq_count: the number of jobs currently in the hardware queue.
425  * @job_id_count: used to assign unique id to the each job.
426  * @timeout_wq: workqueue used to queue @work_tdr
427  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
428  *            timeout interval is over.
429  * @thread: the kthread on which the scheduler which run.
430  * @pending_list: the list of jobs which are currently in the job queue.
431  * @job_list_lock: lock to protect the pending_list.
432  * @hang_limit: once the hangs by a job crosses this limit then it is marked
433  *              guilty and it will no longer be considered for scheduling.
434  * @score: score to help loadbalancer pick a idle sched
435  * @_score: score used when the driver doesn't provide one
436  * @ready: marks if the underlying HW is ready to work
437  * @free_guilty: A hit to time out handler to free the guilty job.
438  * @dev: system &struct device
439  *
440  * One scheduler is implemented for each hardware ring.
441  */
442 struct drm_gpu_scheduler {
443 	const struct drm_sched_backend_ops	*ops;
444 	uint32_t			hw_submission_limit;
445 	long				timeout;
446 	const char			*name;
447 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
448 	wait_queue_head_t		wake_up_worker;
449 	wait_queue_head_t		job_scheduled;
450 	atomic_t			hw_rq_count;
451 	atomic64_t			job_id_count;
452 	struct workqueue_struct		*timeout_wq;
453 	struct delayed_work		work_tdr;
454 	struct task_struct		*thread;
455 	struct list_head		pending_list;
456 	spinlock_t			job_list_lock;
457 	int				hang_limit;
458 	atomic_t                        *score;
459 	atomic_t                        _score;
460 	bool				ready;
461 	bool				free_guilty;
462 	struct device			*dev;
463 };
464 
465 int drm_sched_init(struct drm_gpu_scheduler *sched,
466 		   const struct drm_sched_backend_ops *ops,
467 		   uint32_t hw_submission, unsigned hang_limit,
468 		   long timeout, struct workqueue_struct *timeout_wq,
469 		   atomic_t *score, const char *name, struct device *dev);
470 
471 void drm_sched_fini(struct drm_gpu_scheduler *sched);
472 int drm_sched_job_init(struct drm_sched_job *job,
473 		       struct drm_sched_entity *entity,
474 		       void *owner);
475 void drm_sched_job_arm(struct drm_sched_job *job);
476 int drm_sched_job_add_dependency(struct drm_sched_job *job,
477 				 struct dma_fence *fence);
478 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
479 					    struct drm_gem_object *obj,
480 					    bool write);
481 
482 
483 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
484 				    struct drm_gpu_scheduler **sched_list,
485                                    unsigned int num_sched_list);
486 
487 void drm_sched_job_cleanup(struct drm_sched_job *job);
488 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
489 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
490 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
491 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
492 void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max);
493 void drm_sched_increase_karma(struct drm_sched_job *bad);
494 void drm_sched_reset_karma(struct drm_sched_job *bad);
495 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
496 bool drm_sched_dependency_optimized(struct dma_fence* fence,
497 				    struct drm_sched_entity *entity);
498 void drm_sched_fault(struct drm_gpu_scheduler *sched);
499 void drm_sched_job_kickout(struct drm_sched_job *s_job);
500 
501 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
502 			     struct drm_sched_entity *entity);
503 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
504 				struct drm_sched_entity *entity);
505 
506 int drm_sched_entity_init(struct drm_sched_entity *entity,
507 			  enum drm_sched_priority priority,
508 			  struct drm_gpu_scheduler **sched_list,
509 			  unsigned int num_sched_list,
510 			  atomic_t *guilty);
511 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
512 void drm_sched_entity_fini(struct drm_sched_entity *entity);
513 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
514 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
515 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
516 void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
517 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
518 				   enum drm_sched_priority priority);
519 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
520 
521 struct drm_sched_fence *drm_sched_fence_alloc(
522 	struct drm_sched_entity *s_entity, void *owner);
523 void drm_sched_fence_init(struct drm_sched_fence *fence,
524 			  struct drm_sched_entity *entity);
525 void drm_sched_fence_free(struct drm_sched_fence *fence);
526 
527 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
528 void drm_sched_fence_finished(struct drm_sched_fence *fence);
529 
530 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
531 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
532 		                unsigned long remaining);
533 struct drm_gpu_scheduler *
534 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
535 		     unsigned int num_sched_list);
536 
537 #endif
538