xref: /openbsd-src/sys/dev/pci/drm/include/drm/gpu_scheduler.h (revision 1ad61ae0a79a724d2d3ec69e69c8e1d1ff6b53a0)
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 struct drm_gem_object;
45 
46 struct drm_gpu_scheduler;
47 struct drm_sched_rq;
48 
49 /* These are often used as an (initial) index
50  * to an array, and as such should start at 0.
51  */
52 enum drm_sched_priority {
53 	DRM_SCHED_PRIORITY_MIN,
54 	DRM_SCHED_PRIORITY_NORMAL,
55 	DRM_SCHED_PRIORITY_HIGH,
56 	DRM_SCHED_PRIORITY_KERNEL,
57 
58 	DRM_SCHED_PRIORITY_COUNT
59 };
60 
61 /**
62  * struct drm_sched_entity - A wrapper around a job queue (typically
63  * attached to the DRM file_priv).
64  *
65  * Entities will emit jobs in order to their corresponding hardware
66  * ring, and the scheduler will alternate between entities based on
67  * scheduling policy.
68  */
69 struct drm_sched_entity {
70 	/**
71 	 * @list:
72 	 *
73 	 * Used to append this struct to the list of entities in the runqueue
74 	 * @rq under &drm_sched_rq.entities.
75 	 *
76 	 * Protected by &drm_sched_rq.lock of @rq.
77 	 */
78 	struct list_head		list;
79 
80 	/**
81 	 * @rq:
82 	 *
83 	 * Runqueue on which this entity is currently scheduled.
84 	 *
85 	 * FIXME: Locking is very unclear for this. Writers are protected by
86 	 * @rq_lock, but readers are generally lockless and seem to just race
87 	 * with not even a READ_ONCE.
88 	 */
89 	struct drm_sched_rq		*rq;
90 
91 	/**
92 	 * @sched_list:
93 	 *
94 	 * A list of schedulers (struct drm_gpu_scheduler).  Jobs from this entity can
95 	 * be scheduled on any scheduler on this list.
96 	 *
97 	 * This can be modified by calling drm_sched_entity_modify_sched().
98 	 * Locking is entirely up to the driver, see the above function for more
99 	 * details.
100 	 *
101 	 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
102 	 * set already.
103 	 *
104 	 * FIXME: This means priority changes through
105 	 * drm_sched_entity_set_priority() will be lost henceforth in this case.
106 	 */
107 	struct drm_gpu_scheduler        **sched_list;
108 
109 	/**
110 	 * @num_sched_list:
111 	 *
112 	 * Number of drm_gpu_schedulers in the @sched_list.
113 	 */
114 	unsigned int                    num_sched_list;
115 
116 	/**
117 	 * @priority:
118 	 *
119 	 * Priority of the entity. This can be modified by calling
120 	 * drm_sched_entity_set_priority(). Protected by &rq_lock.
121 	 */
122 	enum drm_sched_priority         priority;
123 
124 	/**
125 	 * @rq_lock:
126 	 *
127 	 * Lock to modify the runqueue to which this entity belongs.
128 	 */
129 	spinlock_t			rq_lock;
130 
131 	/**
132 	 * @job_queue: the list of jobs of this entity.
133 	 */
134 	struct spsc_queue		job_queue;
135 
136 	/**
137 	 * @fence_seq:
138 	 *
139 	 * A linearly increasing seqno incremented with each new
140 	 * &drm_sched_fence which is part of the entity.
141 	 *
142 	 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
143 	 * this doesn't need to be atomic.
144 	 */
145 	atomic_t			fence_seq;
146 
147 	/**
148 	 * @fence_context:
149 	 *
150 	 * A unique context for all the fences which belong to this entity.  The
151 	 * &drm_sched_fence.scheduled uses the fence_context but
152 	 * &drm_sched_fence.finished uses fence_context + 1.
153 	 */
154 	uint64_t			fence_context;
155 
156 	/**
157 	 * @dependency:
158 	 *
159 	 * The dependency fence of the job which is on the top of the job queue.
160 	 */
161 	struct dma_fence		*dependency;
162 
163 	/**
164 	 * @cb:
165 	 *
166 	 * Callback for the dependency fence above.
167 	 */
168 	struct dma_fence_cb		cb;
169 
170 	/**
171 	 * @guilty:
172 	 *
173 	 * Points to entities' guilty.
174 	 */
175 	atomic_t			*guilty;
176 
177 	/**
178 	 * @last_scheduled:
179 	 *
180 	 * Points to the finished fence of the last scheduled job. Only written
181 	 * by the scheduler thread, can be accessed locklessly from
182 	 * drm_sched_job_arm() iff the queue is empty.
183 	 */
184 	struct dma_fence                *last_scheduled;
185 
186 	/**
187 	 * @last_user: last group leader pushing a job into the entity.
188 	 */
189 #ifdef __linux__
190 	struct task_struct		*last_user;
191 #else
192 	struct process			*last_user;
193 #endif
194 
195 	/**
196 	 * @stopped:
197 	 *
198 	 * Marks the enity as removed from rq and destined for
199 	 * termination. This is set by calling drm_sched_entity_flush() and by
200 	 * drm_sched_fini().
201 	 */
202 	bool 				stopped;
203 
204 	/**
205 	 * @entity_idle:
206 	 *
207 	 * Signals when entity is not in use, used to sequence entity cleanup in
208 	 * drm_sched_entity_fini().
209 	 */
210 	struct completion		entity_idle;
211 };
212 
213 /**
214  * struct drm_sched_rq - queue of entities to be scheduled.
215  *
216  * @lock: to modify the entities list.
217  * @sched: the scheduler to which this rq belongs to.
218  * @entities: list of the entities to be scheduled.
219  * @current_entity: the entity which is to be scheduled.
220  *
221  * Run queue is a set of entities scheduling command submissions for
222  * one specific ring. It implements the scheduling policy that selects
223  * the next entity to emit commands from.
224  */
225 struct drm_sched_rq {
226 	spinlock_t			lock;
227 	struct drm_gpu_scheduler	*sched;
228 	struct list_head		entities;
229 	struct drm_sched_entity		*current_entity;
230 };
231 
232 /**
233  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
234  */
235 struct drm_sched_fence {
236         /**
237          * @scheduled: this fence is what will be signaled by the scheduler
238          * when the job is scheduled.
239          */
240 	struct dma_fence		scheduled;
241 
242         /**
243          * @finished: this fence is what will be signaled by the scheduler
244          * when the job is completed.
245          *
246          * When setting up an out fence for the job, you should use
247          * this, since it's available immediately upon
248          * drm_sched_job_init(), and the fence returned by the driver
249          * from run_job() won't be created until the dependencies have
250          * resolved.
251          */
252 	struct dma_fence		finished;
253 
254         /**
255          * @parent: the fence returned by &drm_sched_backend_ops.run_job
256          * when scheduling the job on hardware. We signal the
257          * &drm_sched_fence.finished fence once parent is signalled.
258          */
259 	struct dma_fence		*parent;
260         /**
261          * @sched: the scheduler instance to which the job having this struct
262          * belongs to.
263          */
264 	struct drm_gpu_scheduler	*sched;
265         /**
266          * @lock: the lock used by the scheduled and the finished fences.
267          */
268 	spinlock_t			lock;
269         /**
270          * @owner: job owner for debugging
271          */
272 	void				*owner;
273 };
274 
275 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
276 
277 /**
278  * struct drm_sched_job - A job to be run by an entity.
279  *
280  * @queue_node: used to append this struct to the queue of jobs in an entity.
281  * @list: a job participates in a "pending" and "done" lists.
282  * @sched: the scheduler instance on which this job is scheduled.
283  * @s_fence: contains the fences for the scheduling of job.
284  * @finish_cb: the callback for the finished fence.
285  * @work: Helper to reschdeule job kill to different context.
286  * @id: a unique id assigned to each job scheduled on the scheduler.
287  * @karma: increment on every hang caused by this job. If this exceeds the hang
288  *         limit of the scheduler then the job is marked guilty and will not
289  *         be scheduled further.
290  * @s_priority: the priority of the job.
291  * @entity: the entity to which this job belongs.
292  * @cb: the callback for the parent fence in s_fence.
293  *
294  * A job is created by the driver using drm_sched_job_init(), and
295  * should call drm_sched_entity_push_job() once it wants the scheduler
296  * to schedule the job.
297  */
298 struct drm_sched_job {
299 	struct spsc_node		queue_node;
300 	struct list_head		list;
301 	struct drm_gpu_scheduler	*sched;
302 	struct drm_sched_fence		*s_fence;
303 
304 	/*
305 	 * work is used only after finish_cb has been used and will not be
306 	 * accessed anymore.
307 	 */
308 	union {
309 		struct dma_fence_cb		finish_cb;
310 		struct work_struct 		work;
311 	};
312 
313 	uint64_t			id;
314 	atomic_t			karma;
315 	enum drm_sched_priority		s_priority;
316 	struct drm_sched_entity         *entity;
317 	struct dma_fence_cb		cb;
318 	/**
319 	 * @dependencies:
320 	 *
321 	 * Contains the dependencies as struct dma_fence for this job, see
322 	 * drm_sched_job_add_dependency() and
323 	 * drm_sched_job_add_implicit_dependencies().
324 	 */
325 	struct xarray			dependencies;
326 
327 	/** @last_dependency: tracks @dependencies as they signal */
328 	unsigned long			last_dependency;
329 };
330 
331 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
332 					    int threshold)
333 {
334 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
335 }
336 
337 enum drm_gpu_sched_stat {
338 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
339 	DRM_GPU_SCHED_STAT_NOMINAL,
340 	DRM_GPU_SCHED_STAT_ENODEV,
341 };
342 
343 /**
344  * struct drm_sched_backend_ops - Define the backend operations
345  *	called by the scheduler
346  *
347  * These functions should be implemented in the driver side.
348  */
349 struct drm_sched_backend_ops {
350 	/**
351 	 * @dependency:
352 	 *
353 	 * Called when the scheduler is considering scheduling this job next, to
354 	 * get another struct dma_fence for this job to block on.  Once it
355 	 * returns NULL, run_job() may be called.
356 	 *
357 	 * If a driver exclusively uses drm_sched_job_add_dependency() and
358 	 * drm_sched_job_add_implicit_dependencies() this can be ommitted and
359 	 * left as NULL.
360 	 */
361 	struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
362 					struct drm_sched_entity *s_entity);
363 
364 	/**
365          * @run_job: Called to execute the job once all of the dependencies
366          * have been resolved.  This may be called multiple times, if
367 	 * timedout_job() has happened and drm_sched_job_recovery()
368 	 * decides to try it again.
369 	 */
370 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
371 
372 	/**
373 	 * @timedout_job: Called when a job has taken too long to execute,
374 	 * to trigger GPU recovery.
375 	 *
376 	 * This method is called in a workqueue context.
377 	 *
378 	 * Drivers typically issue a reset to recover from GPU hangs, and this
379 	 * procedure usually follows the following workflow:
380 	 *
381 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
382 	 *    scheduler thread and cancel the timeout work, guaranteeing that
383 	 *    nothing is queued while we reset the hardware queue
384 	 * 2. Try to gracefully stop non-faulty jobs (optional)
385 	 * 3. Issue a GPU reset (driver-specific)
386 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
387 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
388 	 *    jobs can be queued, and the scheduler thread is unblocked
389 	 *
390 	 * Note that some GPUs have distinct hardware queues but need to reset
391 	 * the GPU globally, which requires extra synchronization between the
392 	 * timeout handler of the different &drm_gpu_scheduler. One way to
393 	 * achieve this synchronization is to create an ordered workqueue
394 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
395 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
396 	 * executed sequentially. The above workflow needs to be slightly
397 	 * adjusted in that case:
398 	 *
399 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
400 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
401 	 *    the reset (optional)
402 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
403 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
404 	 *    drm_sched_resubmit_jobs()
405 	 * 5. Restart all schedulers that were stopped in step #1 using
406 	 *    drm_sched_start()
407 	 *
408 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
409 	 * and the underlying driver has started or completed recovery.
410 	 *
411 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
412 	 * available, i.e. has been unplugged.
413 	 */
414 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
415 
416 	/**
417          * @free_job: Called once the job's finished fence has been signaled
418          * and it's time to clean it up.
419 	 */
420 	void (*free_job)(struct drm_sched_job *sched_job);
421 };
422 
423 /**
424  * struct drm_gpu_scheduler - scheduler instance-specific data
425  *
426  * @ops: backend operations provided by the driver.
427  * @hw_submission_limit: the max size of the hardware queue.
428  * @timeout: the time after which a job is removed from the scheduler.
429  * @name: name of the ring for which this scheduler is being used.
430  * @sched_rq: priority wise array of run queues.
431  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
432  *                  is ready to be scheduled.
433  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
434  *                 waits on this wait queue until all the scheduled jobs are
435  *                 finished.
436  * @hw_rq_count: the number of jobs currently in the hardware queue.
437  * @job_id_count: used to assign unique id to the each job.
438  * @timeout_wq: workqueue used to queue @work_tdr
439  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
440  *            timeout interval is over.
441  * @thread: the kthread on which the scheduler which run.
442  * @pending_list: the list of jobs which are currently in the job queue.
443  * @job_list_lock: lock to protect the pending_list.
444  * @hang_limit: once the hangs by a job crosses this limit then it is marked
445  *              guilty and it will no longer be considered for scheduling.
446  * @score: score to help loadbalancer pick a idle sched
447  * @_score: score used when the driver doesn't provide one
448  * @ready: marks if the underlying HW is ready to work
449  * @free_guilty: A hit to time out handler to free the guilty job.
450  * @dev: system &struct device
451  *
452  * One scheduler is implemented for each hardware ring.
453  */
454 struct drm_gpu_scheduler {
455 	const struct drm_sched_backend_ops	*ops;
456 	uint32_t			hw_submission_limit;
457 	long				timeout;
458 	const char			*name;
459 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
460 	wait_queue_head_t		wake_up_worker;
461 	wait_queue_head_t		job_scheduled;
462 	atomic_t			hw_rq_count;
463 	atomic64_t			job_id_count;
464 	struct workqueue_struct		*timeout_wq;
465 	struct delayed_work		work_tdr;
466 #ifdef __linux__
467 	struct task_struct		*thread;
468 #else
469 	struct proc			*thread;
470 #endif
471 	struct list_head		pending_list;
472 	spinlock_t			job_list_lock;
473 	int				hang_limit;
474 	atomic_t                        *score;
475 	atomic_t                        _score;
476 	bool				ready;
477 	bool				free_guilty;
478 	struct device			*dev;
479 };
480 
481 int drm_sched_init(struct drm_gpu_scheduler *sched,
482 		   const struct drm_sched_backend_ops *ops,
483 		   uint32_t hw_submission, unsigned hang_limit,
484 		   long timeout, struct workqueue_struct *timeout_wq,
485 		   atomic_t *score, const char *name, struct device *dev);
486 
487 void drm_sched_fini(struct drm_gpu_scheduler *sched);
488 int drm_sched_job_init(struct drm_sched_job *job,
489 		       struct drm_sched_entity *entity,
490 		       void *owner);
491 void drm_sched_job_arm(struct drm_sched_job *job);
492 int drm_sched_job_add_dependency(struct drm_sched_job *job,
493 				 struct dma_fence *fence);
494 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
495 					    struct drm_gem_object *obj,
496 					    bool write);
497 
498 
499 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
500 				    struct drm_gpu_scheduler **sched_list,
501                                    unsigned int num_sched_list);
502 
503 void drm_sched_job_cleanup(struct drm_sched_job *job);
504 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
505 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
506 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
507 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
508 void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max);
509 void drm_sched_increase_karma(struct drm_sched_job *bad);
510 void drm_sched_reset_karma(struct drm_sched_job *bad);
511 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
512 bool drm_sched_dependency_optimized(struct dma_fence* fence,
513 				    struct drm_sched_entity *entity);
514 void drm_sched_fault(struct drm_gpu_scheduler *sched);
515 void drm_sched_job_kickout(struct drm_sched_job *s_job);
516 
517 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
518 			     struct drm_sched_entity *entity);
519 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
520 				struct drm_sched_entity *entity);
521 
522 int drm_sched_entity_init(struct drm_sched_entity *entity,
523 			  enum drm_sched_priority priority,
524 			  struct drm_gpu_scheduler **sched_list,
525 			  unsigned int num_sched_list,
526 			  atomic_t *guilty);
527 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
528 void drm_sched_entity_fini(struct drm_sched_entity *entity);
529 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
530 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
531 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
532 void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
533 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
534 				   enum drm_sched_priority priority);
535 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
536 
537 struct drm_sched_fence *drm_sched_fence_alloc(
538 	struct drm_sched_entity *s_entity, void *owner);
539 void drm_sched_fence_init(struct drm_sched_fence *fence,
540 			  struct drm_sched_entity *entity);
541 void drm_sched_fence_free(struct drm_sched_fence *fence);
542 
543 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
544 void drm_sched_fence_finished(struct drm_sched_fence *fence);
545 
546 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
547 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
548 		                unsigned long remaining);
549 struct drm_gpu_scheduler *
550 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
551 		     unsigned int num_sched_list);
552 
553 #endif
554