xref: /openbmc/linux/include/drm/gpu_scheduler.h (revision 97e6ea6d)
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 
31 #define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
32 
33 struct drm_gpu_scheduler;
34 struct drm_sched_rq;
35 
36 /* These are often used as an (initial) index
37  * to an array, and as such should start at 0.
38  */
39 enum drm_sched_priority {
40 	DRM_SCHED_PRIORITY_MIN,
41 	DRM_SCHED_PRIORITY_NORMAL,
42 	DRM_SCHED_PRIORITY_HIGH,
43 	DRM_SCHED_PRIORITY_KERNEL,
44 
45 	DRM_SCHED_PRIORITY_COUNT,
46 	DRM_SCHED_PRIORITY_UNSET = -2
47 };
48 
49 /**
50  * struct drm_sched_entity - A wrapper around a job queue (typically
51  * attached to the DRM file_priv).
52  *
53  * @list: used to append this struct to the list of entities in the
54  *        runqueue.
55  * @rq: runqueue on which this entity is currently scheduled.
56  * @sched_list: A list of schedulers (drm_gpu_schedulers).
57  *              Jobs from this entity can be scheduled on any scheduler
58  *              on this list.
59  * @num_sched_list: number of drm_gpu_schedulers in the sched_list.
60  * @priority: priority of the entity
61  * @rq_lock: lock to modify the runqueue to which this entity belongs.
62  * @job_queue: the list of jobs of this entity.
63  * @fence_seq: a linearly increasing seqno incremented with each
64  *             new &drm_sched_fence which is part of the entity.
65  * @fence_context: a unique context for all the fences which belong
66  *                 to this entity.
67  *                 The &drm_sched_fence.scheduled uses the
68  *                 fence_context but &drm_sched_fence.finished uses
69  *                 fence_context + 1.
70  * @dependency: the dependency fence of the job which is on the top
71  *              of the job queue.
72  * @cb: callback for the dependency fence above.
73  * @guilty: points to ctx's guilty.
74  * @fini_status: contains the exit status in case the process was signalled.
75  * @last_scheduled: points to the finished fence of the last scheduled job.
76  * @last_user: last group leader pushing a job into the entity.
77  * @stopped: Marks the enity as removed from rq and destined for termination.
78  * @entity_idle: Signals when enityt is not in use
79  *
80  * Entities will emit jobs in order to their corresponding hardware
81  * ring, and the scheduler will alternate between entities based on
82  * scheduling policy.
83  */
84 struct drm_sched_entity {
85 	struct list_head		list;
86 	struct drm_sched_rq		*rq;
87 	struct drm_gpu_scheduler        **sched_list;
88 	unsigned int                    num_sched_list;
89 	enum drm_sched_priority         priority;
90 	spinlock_t			rq_lock;
91 
92 	struct spsc_queue		job_queue;
93 
94 	atomic_t			fence_seq;
95 	uint64_t			fence_context;
96 
97 	struct dma_fence		*dependency;
98 	struct dma_fence_cb		cb;
99 	atomic_t			*guilty;
100 	struct dma_fence                *last_scheduled;
101 	struct task_struct		*last_user;
102 	bool 				stopped;
103 	struct completion		entity_idle;
104 };
105 
106 /**
107  * struct drm_sched_rq - queue of entities to be scheduled.
108  *
109  * @lock: to modify the entities list.
110  * @sched: the scheduler to which this rq belongs to.
111  * @entities: list of the entities to be scheduled.
112  * @current_entity: the entity which is to be scheduled.
113  *
114  * Run queue is a set of entities scheduling command submissions for
115  * one specific ring. It implements the scheduling policy that selects
116  * the next entity to emit commands from.
117  */
118 struct drm_sched_rq {
119 	spinlock_t			lock;
120 	struct drm_gpu_scheduler	*sched;
121 	struct list_head		entities;
122 	struct drm_sched_entity		*current_entity;
123 };
124 
125 /**
126  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
127  */
128 struct drm_sched_fence {
129         /**
130          * @scheduled: this fence is what will be signaled by the scheduler
131          * when the job is scheduled.
132          */
133 	struct dma_fence		scheduled;
134 
135         /**
136          * @finished: this fence is what will be signaled by the scheduler
137          * when the job is completed.
138          *
139          * When setting up an out fence for the job, you should use
140          * this, since it's available immediately upon
141          * drm_sched_job_init(), and the fence returned by the driver
142          * from run_job() won't be created until the dependencies have
143          * resolved.
144          */
145 	struct dma_fence		finished;
146 
147         /**
148          * @parent: the fence returned by &drm_sched_backend_ops.run_job
149          * when scheduling the job on hardware. We signal the
150          * &drm_sched_fence.finished fence once parent is signalled.
151          */
152 	struct dma_fence		*parent;
153         /**
154          * @sched: the scheduler instance to which the job having this struct
155          * belongs to.
156          */
157 	struct drm_gpu_scheduler	*sched;
158         /**
159          * @lock: the lock used by the scheduled and the finished fences.
160          */
161 	spinlock_t			lock;
162         /**
163          * @owner: job owner for debugging
164          */
165 	void				*owner;
166 };
167 
168 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
169 
170 /**
171  * struct drm_sched_job - A job to be run by an entity.
172  *
173  * @queue_node: used to append this struct to the queue of jobs in an entity.
174  * @list: a job participates in a "pending" and "done" lists.
175  * @sched: the scheduler instance on which this job is scheduled.
176  * @s_fence: contains the fences for the scheduling of job.
177  * @finish_cb: the callback for the finished fence.
178  * @id: a unique id assigned to each job scheduled on the scheduler.
179  * @karma: increment on every hang caused by this job. If this exceeds the hang
180  *         limit of the scheduler then the job is marked guilty and will not
181  *         be scheduled further.
182  * @s_priority: the priority of the job.
183  * @entity: the entity to which this job belongs.
184  * @cb: the callback for the parent fence in s_fence.
185  *
186  * A job is created by the driver using drm_sched_job_init(), and
187  * should call drm_sched_entity_push_job() once it wants the scheduler
188  * to schedule the job.
189  */
190 struct drm_sched_job {
191 	struct spsc_node		queue_node;
192 	struct list_head		list;
193 	struct drm_gpu_scheduler	*sched;
194 	struct drm_sched_fence		*s_fence;
195 	struct dma_fence_cb		finish_cb;
196 	uint64_t			id;
197 	atomic_t			karma;
198 	enum drm_sched_priority		s_priority;
199 	struct drm_sched_entity         *entity;
200 	struct dma_fence_cb		cb;
201 };
202 
203 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
204 					    int threshold)
205 {
206 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
207 }
208 
209 enum drm_gpu_sched_stat {
210 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
211 	DRM_GPU_SCHED_STAT_NOMINAL,
212 	DRM_GPU_SCHED_STAT_ENODEV,
213 };
214 
215 /**
216  * struct drm_sched_backend_ops
217  *
218  * Define the backend operations called by the scheduler,
219  * these functions should be implemented in driver side.
220  */
221 struct drm_sched_backend_ops {
222 	/**
223          * @dependency: Called when the scheduler is considering scheduling
224          * this job next, to get another struct dma_fence for this job to
225 	 * block on.  Once it returns NULL, run_job() may be called.
226 	 */
227 	struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
228 					struct drm_sched_entity *s_entity);
229 
230 	/**
231          * @run_job: Called to execute the job once all of the dependencies
232          * have been resolved.  This may be called multiple times, if
233 	 * timedout_job() has happened and drm_sched_job_recovery()
234 	 * decides to try it again.
235 	 */
236 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
237 
238 	/**
239 	 * @timedout_job: Called when a job has taken too long to execute,
240 	 * to trigger GPU recovery.
241 	 *
242 	 * This method is called in a workqueue context.
243 	 *
244 	 * Drivers typically issue a reset to recover from GPU hangs, and this
245 	 * procedure usually follows the following workflow:
246 	 *
247 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
248 	 *    scheduler thread and cancel the timeout work, guaranteeing that
249 	 *    nothing is queued while we reset the hardware queue
250 	 * 2. Try to gracefully stop non-faulty jobs (optional)
251 	 * 3. Issue a GPU reset (driver-specific)
252 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
253 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
254 	 *    jobs can be queued, and the scheduler thread is unblocked
255 	 *
256 	 * Note that some GPUs have distinct hardware queues but need to reset
257 	 * the GPU globally, which requires extra synchronization between the
258 	 * timeout handler of the different &drm_gpu_scheduler. One way to
259 	 * achieve this synchronization is to create an ordered workqueue
260 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
261 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
262 	 * executed sequentially. The above workflow needs to be slightly
263 	 * adjusted in that case:
264 	 *
265 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
266 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
267 	 *    the reset (optional)
268 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
269 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
270 	 *    drm_sched_resubmit_jobs()
271 	 * 5. Restart all schedulers that were stopped in step #1 using
272 	 *    drm_sched_start()
273 	 *
274 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
275 	 * and the underlying driver has started or completed recovery.
276 	 *
277 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
278 	 * available, i.e. has been unplugged.
279 	 */
280 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
281 
282 	/**
283          * @free_job: Called once the job's finished fence has been signaled
284          * and it's time to clean it up.
285 	 */
286 	void (*free_job)(struct drm_sched_job *sched_job);
287 };
288 
289 /**
290  * struct drm_gpu_scheduler
291  *
292  * @ops: backend operations provided by the driver.
293  * @hw_submission_limit: the max size of the hardware queue.
294  * @timeout: the time after which a job is removed from the scheduler.
295  * @name: name of the ring for which this scheduler is being used.
296  * @sched_rq: priority wise array of run queues.
297  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
298  *                  is ready to be scheduled.
299  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
300  *                 waits on this wait queue until all the scheduled jobs are
301  *                 finished.
302  * @hw_rq_count: the number of jobs currently in the hardware queue.
303  * @job_id_count: used to assign unique id to the each job.
304  * @timeout_wq: workqueue used to queue @work_tdr
305  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
306  *            timeout interval is over.
307  * @thread: the kthread on which the scheduler which run.
308  * @pending_list: the list of jobs which are currently in the job queue.
309  * @job_list_lock: lock to protect the pending_list.
310  * @hang_limit: once the hangs by a job crosses this limit then it is marked
311  *              guilty and it will no longer be considered for scheduling.
312  * @score: score to help loadbalancer pick a idle sched
313  * @_score: score used when the driver doesn't provide one
314  * @ready: marks if the underlying HW is ready to work
315  * @free_guilty: A hit to time out handler to free the guilty job.
316  *
317  * One scheduler is implemented for each hardware ring.
318  */
319 struct drm_gpu_scheduler {
320 	const struct drm_sched_backend_ops	*ops;
321 	uint32_t			hw_submission_limit;
322 	long				timeout;
323 	const char			*name;
324 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
325 	wait_queue_head_t		wake_up_worker;
326 	wait_queue_head_t		job_scheduled;
327 	atomic_t			hw_rq_count;
328 	atomic64_t			job_id_count;
329 	struct workqueue_struct		*timeout_wq;
330 	struct delayed_work		work_tdr;
331 	struct task_struct		*thread;
332 	struct list_head		pending_list;
333 	spinlock_t			job_list_lock;
334 	int				hang_limit;
335 	atomic_t                        *score;
336 	atomic_t                        _score;
337 	bool				ready;
338 	bool				free_guilty;
339 };
340 
341 int drm_sched_init(struct drm_gpu_scheduler *sched,
342 		   const struct drm_sched_backend_ops *ops,
343 		   uint32_t hw_submission, unsigned hang_limit,
344 		   long timeout, struct workqueue_struct *timeout_wq,
345 		   atomic_t *score, const char *name);
346 
347 void drm_sched_fini(struct drm_gpu_scheduler *sched);
348 int drm_sched_job_init(struct drm_sched_job *job,
349 		       struct drm_sched_entity *entity,
350 		       void *owner);
351 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
352 				    struct drm_gpu_scheduler **sched_list,
353                                    unsigned int num_sched_list);
354 
355 void drm_sched_job_cleanup(struct drm_sched_job *job);
356 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
357 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
358 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
359 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
360 void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max);
361 void drm_sched_increase_karma(struct drm_sched_job *bad);
362 void drm_sched_reset_karma(struct drm_sched_job *bad);
363 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
364 bool drm_sched_dependency_optimized(struct dma_fence* fence,
365 				    struct drm_sched_entity *entity);
366 void drm_sched_fault(struct drm_gpu_scheduler *sched);
367 void drm_sched_job_kickout(struct drm_sched_job *s_job);
368 
369 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
370 			     struct drm_sched_entity *entity);
371 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
372 				struct drm_sched_entity *entity);
373 
374 int drm_sched_entity_init(struct drm_sched_entity *entity,
375 			  enum drm_sched_priority priority,
376 			  struct drm_gpu_scheduler **sched_list,
377 			  unsigned int num_sched_list,
378 			  atomic_t *guilty);
379 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
380 void drm_sched_entity_fini(struct drm_sched_entity *entity);
381 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
382 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
383 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
384 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
385 			       struct drm_sched_entity *entity);
386 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
387 				   enum drm_sched_priority priority);
388 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
389 
390 struct drm_sched_fence *drm_sched_fence_create(
391 	struct drm_sched_entity *s_entity, void *owner);
392 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
393 void drm_sched_fence_finished(struct drm_sched_fence *fence);
394 
395 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
396 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
397 		                unsigned long remaining);
398 struct drm_gpu_scheduler *
399 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
400 		     unsigned int num_sched_list);
401 
402 #endif
403