xref: /openbmc/linux/include/drm/gpu_scheduler.h (revision ebd5f74255b9f5f8a154ba5535f83387ae599d46)
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 
32 #define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
33 
34 struct drm_gem_object;
35 
36 struct drm_gpu_scheduler;
37 struct drm_sched_rq;
38 
39 /* These are often used as an (initial) index
40  * to an array, and as such should start at 0.
41  */
42 enum drm_sched_priority {
43 	DRM_SCHED_PRIORITY_MIN,
44 	DRM_SCHED_PRIORITY_NORMAL,
45 	DRM_SCHED_PRIORITY_HIGH,
46 	DRM_SCHED_PRIORITY_KERNEL,
47 
48 	DRM_SCHED_PRIORITY_COUNT,
49 	DRM_SCHED_PRIORITY_UNSET = -2
50 };
51 
52 /**
53  * struct drm_sched_entity - A wrapper around a job queue (typically
54  * attached to the DRM file_priv).
55  *
56  * @list: used to append this struct to the list of entities in the
57  *        runqueue.
58  * @rq: runqueue on which this entity is currently scheduled.
59  * @sched_list: A list of schedulers (drm_gpu_schedulers).
60  *              Jobs from this entity can be scheduled on any scheduler
61  *              on this list.
62  * @num_sched_list: number of drm_gpu_schedulers in the sched_list.
63  * @priority: priority of the entity
64  * @rq_lock: lock to modify the runqueue to which this entity belongs.
65  * @job_queue: the list of jobs of this entity.
66  * @fence_seq: a linearly increasing seqno incremented with each
67  *             new &drm_sched_fence which is part of the entity.
68  * @fence_context: a unique context for all the fences which belong
69  *                 to this entity.
70  *                 The &drm_sched_fence.scheduled uses the
71  *                 fence_context but &drm_sched_fence.finished uses
72  *                 fence_context + 1.
73  * @dependency: the dependency fence of the job which is on the top
74  *              of the job queue.
75  * @cb: callback for the dependency fence above.
76  * @guilty: points to ctx's guilty.
77  * @fini_status: contains the exit status in case the process was signalled.
78  * @last_scheduled: points to the finished fence of the last scheduled job.
79  * @last_user: last group leader pushing a job into the entity.
80  * @stopped: Marks the enity as removed from rq and destined for termination.
81  * @entity_idle: Signals when enityt is not in use
82  *
83  * Entities will emit jobs in order to their corresponding hardware
84  * ring, and the scheduler will alternate between entities based on
85  * scheduling policy.
86  */
87 struct drm_sched_entity {
88 	struct list_head		list;
89 	struct drm_sched_rq		*rq;
90 	struct drm_gpu_scheduler        **sched_list;
91 	unsigned int                    num_sched_list;
92 	enum drm_sched_priority         priority;
93 	spinlock_t			rq_lock;
94 
95 	struct spsc_queue		job_queue;
96 
97 	atomic_t			fence_seq;
98 	uint64_t			fence_context;
99 
100 	struct dma_fence		*dependency;
101 	struct dma_fence_cb		cb;
102 	atomic_t			*guilty;
103 	struct dma_fence                *last_scheduled;
104 	struct task_struct		*last_user;
105 	bool 				stopped;
106 	struct completion		entity_idle;
107 };
108 
109 /**
110  * struct drm_sched_rq - queue of entities to be scheduled.
111  *
112  * @lock: to modify the entities list.
113  * @sched: the scheduler to which this rq belongs to.
114  * @entities: list of the entities to be scheduled.
115  * @current_entity: the entity which is to be scheduled.
116  *
117  * Run queue is a set of entities scheduling command submissions for
118  * one specific ring. It implements the scheduling policy that selects
119  * the next entity to emit commands from.
120  */
121 struct drm_sched_rq {
122 	spinlock_t			lock;
123 	struct drm_gpu_scheduler	*sched;
124 	struct list_head		entities;
125 	struct drm_sched_entity		*current_entity;
126 };
127 
128 /**
129  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
130  */
131 struct drm_sched_fence {
132         /**
133          * @scheduled: this fence is what will be signaled by the scheduler
134          * when the job is scheduled.
135          */
136 	struct dma_fence		scheduled;
137 
138         /**
139          * @finished: this fence is what will be signaled by the scheduler
140          * when the job is completed.
141          *
142          * When setting up an out fence for the job, you should use
143          * this, since it's available immediately upon
144          * drm_sched_job_init(), and the fence returned by the driver
145          * from run_job() won't be created until the dependencies have
146          * resolved.
147          */
148 	struct dma_fence		finished;
149 
150         /**
151          * @parent: the fence returned by &drm_sched_backend_ops.run_job
152          * when scheduling the job on hardware. We signal the
153          * &drm_sched_fence.finished fence once parent is signalled.
154          */
155 	struct dma_fence		*parent;
156         /**
157          * @sched: the scheduler instance to which the job having this struct
158          * belongs to.
159          */
160 	struct drm_gpu_scheduler	*sched;
161         /**
162          * @lock: the lock used by the scheduled and the finished fences.
163          */
164 	spinlock_t			lock;
165         /**
166          * @owner: job owner for debugging
167          */
168 	void				*owner;
169 };
170 
171 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
172 
173 /**
174  * struct drm_sched_job - A job to be run by an entity.
175  *
176  * @queue_node: used to append this struct to the queue of jobs in an entity.
177  * @list: a job participates in a "pending" and "done" lists.
178  * @sched: the scheduler instance on which this job is scheduled.
179  * @s_fence: contains the fences for the scheduling of job.
180  * @finish_cb: the callback for the finished fence.
181  * @id: a unique id assigned to each job scheduled on the scheduler.
182  * @karma: increment on every hang caused by this job. If this exceeds the hang
183  *         limit of the scheduler then the job is marked guilty and will not
184  *         be scheduled further.
185  * @s_priority: the priority of the job.
186  * @entity: the entity to which this job belongs.
187  * @cb: the callback for the parent fence in s_fence.
188  *
189  * A job is created by the driver using drm_sched_job_init(), and
190  * should call drm_sched_entity_push_job() once it wants the scheduler
191  * to schedule the job.
192  */
193 struct drm_sched_job {
194 	struct spsc_node		queue_node;
195 	struct list_head		list;
196 	struct drm_gpu_scheduler	*sched;
197 	struct drm_sched_fence		*s_fence;
198 	struct dma_fence_cb		finish_cb;
199 	uint64_t			id;
200 	atomic_t			karma;
201 	enum drm_sched_priority		s_priority;
202 	struct drm_sched_entity         *entity;
203 	struct dma_fence_cb		cb;
204 	/**
205 	 * @dependencies:
206 	 *
207 	 * Contains the dependencies as struct dma_fence for this job, see
208 	 * drm_sched_job_add_dependency() and
209 	 * drm_sched_job_add_implicit_dependencies().
210 	 */
211 	struct xarray			dependencies;
212 
213 	/** @last_dependency: tracks @dependencies as they signal */
214 	unsigned long			last_dependency;
215 };
216 
217 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
218 					    int threshold)
219 {
220 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
221 }
222 
223 enum drm_gpu_sched_stat {
224 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
225 	DRM_GPU_SCHED_STAT_NOMINAL,
226 	DRM_GPU_SCHED_STAT_ENODEV,
227 };
228 
229 /**
230  * struct drm_sched_backend_ops
231  *
232  * Define the backend operations called by the scheduler,
233  * these functions should be implemented in driver side.
234  */
235 struct drm_sched_backend_ops {
236 	/**
237 	 * @dependency:
238 	 *
239 	 * Called when the scheduler is considering scheduling this job next, to
240 	 * get another struct dma_fence for this job to block on.  Once it
241 	 * returns NULL, run_job() may be called.
242 	 *
243 	 * If a driver exclusively uses drm_sched_job_add_dependency() and
244 	 * drm_sched_job_add_implicit_dependencies() this can be ommitted and
245 	 * left as NULL.
246 	 */
247 	struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
248 					struct drm_sched_entity *s_entity);
249 
250 	/**
251          * @run_job: Called to execute the job once all of the dependencies
252          * have been resolved.  This may be called multiple times, if
253 	 * timedout_job() has happened and drm_sched_job_recovery()
254 	 * decides to try it again.
255 	 */
256 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
257 
258 	/**
259 	 * @timedout_job: Called when a job has taken too long to execute,
260 	 * to trigger GPU recovery.
261 	 *
262 	 * This method is called in a workqueue context.
263 	 *
264 	 * Drivers typically issue a reset to recover from GPU hangs, and this
265 	 * procedure usually follows the following workflow:
266 	 *
267 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
268 	 *    scheduler thread and cancel the timeout work, guaranteeing that
269 	 *    nothing is queued while we reset the hardware queue
270 	 * 2. Try to gracefully stop non-faulty jobs (optional)
271 	 * 3. Issue a GPU reset (driver-specific)
272 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
273 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
274 	 *    jobs can be queued, and the scheduler thread is unblocked
275 	 *
276 	 * Note that some GPUs have distinct hardware queues but need to reset
277 	 * the GPU globally, which requires extra synchronization between the
278 	 * timeout handler of the different &drm_gpu_scheduler. One way to
279 	 * achieve this synchronization is to create an ordered workqueue
280 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
281 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
282 	 * executed sequentially. The above workflow needs to be slightly
283 	 * adjusted in that case:
284 	 *
285 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
286 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
287 	 *    the reset (optional)
288 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
289 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
290 	 *    drm_sched_resubmit_jobs()
291 	 * 5. Restart all schedulers that were stopped in step #1 using
292 	 *    drm_sched_start()
293 	 *
294 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
295 	 * and the underlying driver has started or completed recovery.
296 	 *
297 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
298 	 * available, i.e. has been unplugged.
299 	 */
300 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
301 
302 	/**
303          * @free_job: Called once the job's finished fence has been signaled
304          * and it's time to clean it up.
305 	 */
306 	void (*free_job)(struct drm_sched_job *sched_job);
307 };
308 
309 /**
310  * struct drm_gpu_scheduler
311  *
312  * @ops: backend operations provided by the driver.
313  * @hw_submission_limit: the max size of the hardware queue.
314  * @timeout: the time after which a job is removed from the scheduler.
315  * @name: name of the ring for which this scheduler is being used.
316  * @sched_rq: priority wise array of run queues.
317  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
318  *                  is ready to be scheduled.
319  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
320  *                 waits on this wait queue until all the scheduled jobs are
321  *                 finished.
322  * @hw_rq_count: the number of jobs currently in the hardware queue.
323  * @job_id_count: used to assign unique id to the each job.
324  * @timeout_wq: workqueue used to queue @work_tdr
325  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
326  *            timeout interval is over.
327  * @thread: the kthread on which the scheduler which run.
328  * @pending_list: the list of jobs which are currently in the job queue.
329  * @job_list_lock: lock to protect the pending_list.
330  * @hang_limit: once the hangs by a job crosses this limit then it is marked
331  *              guilty and it will no longer be considered for scheduling.
332  * @score: score to help loadbalancer pick a idle sched
333  * @_score: score used when the driver doesn't provide one
334  * @ready: marks if the underlying HW is ready to work
335  * @free_guilty: A hit to time out handler to free the guilty job.
336  *
337  * One scheduler is implemented for each hardware ring.
338  */
339 struct drm_gpu_scheduler {
340 	const struct drm_sched_backend_ops	*ops;
341 	uint32_t			hw_submission_limit;
342 	long				timeout;
343 	const char			*name;
344 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
345 	wait_queue_head_t		wake_up_worker;
346 	wait_queue_head_t		job_scheduled;
347 	atomic_t			hw_rq_count;
348 	atomic64_t			job_id_count;
349 	struct workqueue_struct		*timeout_wq;
350 	struct delayed_work		work_tdr;
351 	struct task_struct		*thread;
352 	struct list_head		pending_list;
353 	spinlock_t			job_list_lock;
354 	int				hang_limit;
355 	atomic_t                        *score;
356 	atomic_t                        _score;
357 	bool				ready;
358 	bool				free_guilty;
359 };
360 
361 int drm_sched_init(struct drm_gpu_scheduler *sched,
362 		   const struct drm_sched_backend_ops *ops,
363 		   uint32_t hw_submission, unsigned hang_limit,
364 		   long timeout, struct workqueue_struct *timeout_wq,
365 		   atomic_t *score, const char *name);
366 
367 void drm_sched_fini(struct drm_gpu_scheduler *sched);
368 int drm_sched_job_init(struct drm_sched_job *job,
369 		       struct drm_sched_entity *entity,
370 		       void *owner);
371 void drm_sched_job_arm(struct drm_sched_job *job);
372 int drm_sched_job_add_dependency(struct drm_sched_job *job,
373 				 struct dma_fence *fence);
374 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
375 					    struct drm_gem_object *obj,
376 					    bool write);
377 
378 
379 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
380 				    struct drm_gpu_scheduler **sched_list,
381                                    unsigned int num_sched_list);
382 
383 void drm_sched_job_cleanup(struct drm_sched_job *job);
384 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
385 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
386 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
387 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
388 void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max);
389 void drm_sched_increase_karma(struct drm_sched_job *bad);
390 void drm_sched_reset_karma(struct drm_sched_job *bad);
391 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
392 bool drm_sched_dependency_optimized(struct dma_fence* fence,
393 				    struct drm_sched_entity *entity);
394 void drm_sched_fault(struct drm_gpu_scheduler *sched);
395 void drm_sched_job_kickout(struct drm_sched_job *s_job);
396 
397 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
398 			     struct drm_sched_entity *entity);
399 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
400 				struct drm_sched_entity *entity);
401 
402 int drm_sched_entity_init(struct drm_sched_entity *entity,
403 			  enum drm_sched_priority priority,
404 			  struct drm_gpu_scheduler **sched_list,
405 			  unsigned int num_sched_list,
406 			  atomic_t *guilty);
407 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
408 void drm_sched_entity_fini(struct drm_sched_entity *entity);
409 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
410 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
411 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
412 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
413 			       struct drm_sched_entity *entity);
414 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
415 				   enum drm_sched_priority priority);
416 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
417 
418 struct drm_sched_fence *drm_sched_fence_alloc(
419 	struct drm_sched_entity *s_entity, void *owner);
420 void drm_sched_fence_init(struct drm_sched_fence *fence,
421 			  struct drm_sched_entity *entity);
422 void drm_sched_fence_free(struct rcu_head *rcu);
423 
424 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
425 void drm_sched_fence_finished(struct drm_sched_fence *fence);
426 
427 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
428 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
429 		                unsigned long remaining);
430 struct drm_gpu_scheduler *
431 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
432 		     unsigned int num_sched_list);
433 
434 #endif
435