xref: /openbsd-src/sys/dev/pci/drm/include/drm/gpu_scheduler.h (revision c1a45aed656e7d5627c30c92421893a76f370ccb)
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 #ifdef __linux__
102 	struct task_struct		*last_user;
103 #else
104 	struct process			*last_user;
105 #endif
106 	bool 				stopped;
107 	struct completion		entity_idle;
108 };
109 
110 /**
111  * struct drm_sched_rq - queue of entities to be scheduled.
112  *
113  * @lock: to modify the entities list.
114  * @sched: the scheduler to which this rq belongs to.
115  * @entities: list of the entities to be scheduled.
116  * @current_entity: the entity which is to be scheduled.
117  *
118  * Run queue is a set of entities scheduling command submissions for
119  * one specific ring. It implements the scheduling policy that selects
120  * the next entity to emit commands from.
121  */
122 struct drm_sched_rq {
123 	spinlock_t			lock;
124 	struct drm_gpu_scheduler	*sched;
125 	struct list_head		entities;
126 	struct drm_sched_entity		*current_entity;
127 };
128 
129 /**
130  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
131  */
132 struct drm_sched_fence {
133         /**
134          * @scheduled: this fence is what will be signaled by the scheduler
135          * when the job is scheduled.
136          */
137 	struct dma_fence		scheduled;
138 
139         /**
140          * @finished: this fence is what will be signaled by the scheduler
141          * when the job is completed.
142          *
143          * When setting up an out fence for the job, you should use
144          * this, since it's available immediately upon
145          * drm_sched_job_init(), and the fence returned by the driver
146          * from run_job() won't be created until the dependencies have
147          * resolved.
148          */
149 	struct dma_fence		finished;
150 
151         /**
152          * @parent: the fence returned by &drm_sched_backend_ops.run_job
153          * when scheduling the job on hardware. We signal the
154          * &drm_sched_fence.finished fence once parent is signalled.
155          */
156 	struct dma_fence		*parent;
157         /**
158          * @sched: the scheduler instance to which the job having this struct
159          * belongs to.
160          */
161 	struct drm_gpu_scheduler	*sched;
162         /**
163          * @lock: the lock used by the scheduled and the finished fences.
164          */
165 	spinlock_t			lock;
166         /**
167          * @owner: job owner for debugging
168          */
169 	void				*owner;
170 };
171 
172 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
173 
174 /**
175  * struct drm_sched_job - A job to be run by an entity.
176  *
177  * @queue_node: used to append this struct to the queue of jobs in an entity.
178  * @list: a job participates in a "pending" and "done" lists.
179  * @sched: the scheduler instance on which this job is scheduled.
180  * @s_fence: contains the fences for the scheduling of job.
181  * @finish_cb: the callback for the finished fence.
182  * @id: a unique id assigned to each job scheduled on the scheduler.
183  * @karma: increment on every hang caused by this job. If this exceeds the hang
184  *         limit of the scheduler then the job is marked guilty and will not
185  *         be scheduled further.
186  * @s_priority: the priority of the job.
187  * @entity: the entity to which this job belongs.
188  * @cb: the callback for the parent fence in s_fence.
189  *
190  * A job is created by the driver using drm_sched_job_init(), and
191  * should call drm_sched_entity_push_job() once it wants the scheduler
192  * to schedule the job.
193  */
194 struct drm_sched_job {
195 	struct spsc_node		queue_node;
196 	struct list_head		list;
197 	struct drm_gpu_scheduler	*sched;
198 	struct drm_sched_fence		*s_fence;
199 	struct dma_fence_cb		finish_cb;
200 	uint64_t			id;
201 	atomic_t			karma;
202 	enum drm_sched_priority		s_priority;
203 	struct drm_sched_entity         *entity;
204 	struct dma_fence_cb		cb;
205 };
206 
207 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
208 					    int threshold)
209 {
210 	return s_job && atomic_inc_return(&s_job->karma) > threshold;
211 }
212 
213 enum drm_gpu_sched_stat {
214 	DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
215 	DRM_GPU_SCHED_STAT_NOMINAL,
216 	DRM_GPU_SCHED_STAT_ENODEV,
217 };
218 
219 /**
220  * struct drm_sched_backend_ops
221  *
222  * Define the backend operations called by the scheduler,
223  * these functions should be implemented in driver side.
224  */
225 struct drm_sched_backend_ops {
226 	/**
227          * @dependency: Called when the scheduler is considering scheduling
228          * this job next, to get another struct dma_fence for this job to
229 	 * block on.  Once it returns NULL, run_job() may be called.
230 	 */
231 	struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
232 					struct drm_sched_entity *s_entity);
233 
234 	/**
235          * @run_job: Called to execute the job once all of the dependencies
236          * have been resolved.  This may be called multiple times, if
237 	 * timedout_job() has happened and drm_sched_job_recovery()
238 	 * decides to try it again.
239 	 */
240 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
241 
242 	/**
243 	 * @timedout_job: Called when a job has taken too long to execute,
244 	 * to trigger GPU recovery.
245 	 *
246 	 * This method is called in a workqueue context.
247 	 *
248 	 * Drivers typically issue a reset to recover from GPU hangs, and this
249 	 * procedure usually follows the following workflow:
250 	 *
251 	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
252 	 *    scheduler thread and cancel the timeout work, guaranteeing that
253 	 *    nothing is queued while we reset the hardware queue
254 	 * 2. Try to gracefully stop non-faulty jobs (optional)
255 	 * 3. Issue a GPU reset (driver-specific)
256 	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
257 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
258 	 *    jobs can be queued, and the scheduler thread is unblocked
259 	 *
260 	 * Note that some GPUs have distinct hardware queues but need to reset
261 	 * the GPU globally, which requires extra synchronization between the
262 	 * timeout handler of the different &drm_gpu_scheduler. One way to
263 	 * achieve this synchronization is to create an ordered workqueue
264 	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
265 	 * queue to drm_sched_init(), to guarantee that timeout handlers are
266 	 * executed sequentially. The above workflow needs to be slightly
267 	 * adjusted in that case:
268 	 *
269 	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
270 	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
271 	 *    the reset (optional)
272 	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
273 	 * 4. Re-submit jobs on all schedulers impacted by the reset using
274 	 *    drm_sched_resubmit_jobs()
275 	 * 5. Restart all schedulers that were stopped in step #1 using
276 	 *    drm_sched_start()
277 	 *
278 	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
279 	 * and the underlying driver has started or completed recovery.
280 	 *
281 	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
282 	 * available, i.e. has been unplugged.
283 	 */
284 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
285 
286 	/**
287          * @free_job: Called once the job's finished fence has been signaled
288          * and it's time to clean it up.
289 	 */
290 	void (*free_job)(struct drm_sched_job *sched_job);
291 };
292 
293 /**
294  * struct drm_gpu_scheduler
295  *
296  * @ops: backend operations provided by the driver.
297  * @hw_submission_limit: the max size of the hardware queue.
298  * @timeout: the time after which a job is removed from the scheduler.
299  * @name: name of the ring for which this scheduler is being used.
300  * @sched_rq: priority wise array of run queues.
301  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
302  *                  is ready to be scheduled.
303  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
304  *                 waits on this wait queue until all the scheduled jobs are
305  *                 finished.
306  * @hw_rq_count: the number of jobs currently in the hardware queue.
307  * @job_id_count: used to assign unique id to the each job.
308  * @timeout_wq: workqueue used to queue @work_tdr
309  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
310  *            timeout interval is over.
311  * @thread: the kthread on which the scheduler which run.
312  * @pending_list: the list of jobs which are currently in the job queue.
313  * @job_list_lock: lock to protect the pending_list.
314  * @hang_limit: once the hangs by a job crosses this limit then it is marked
315  *              guilty and it will no longer be considered for scheduling.
316  * @score: score to help loadbalancer pick a idle sched
317  * @_score: score used when the driver doesn't provide one
318  * @ready: marks if the underlying HW is ready to work
319  * @free_guilty: A hit to time out handler to free the guilty job.
320  *
321  * One scheduler is implemented for each hardware ring.
322  */
323 struct drm_gpu_scheduler {
324 	const struct drm_sched_backend_ops	*ops;
325 	uint32_t			hw_submission_limit;
326 	long				timeout;
327 	const char			*name;
328 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
329 	wait_queue_head_t		wake_up_worker;
330 	wait_queue_head_t		job_scheduled;
331 	atomic_t			hw_rq_count;
332 	atomic64_t			job_id_count;
333 	struct workqueue_struct		*timeout_wq;
334 	struct delayed_work		work_tdr;
335 #ifdef __linux__
336 	struct task_struct		*thread;
337 #else
338 	struct proc			*thread;
339 #endif
340 	struct list_head		pending_list;
341 	spinlock_t			job_list_lock;
342 	int				hang_limit;
343 	atomic_t                        *score;
344 	atomic_t                        _score;
345 	bool				ready;
346 	bool				free_guilty;
347 };
348 
349 int drm_sched_init(struct drm_gpu_scheduler *sched,
350 		   const struct drm_sched_backend_ops *ops,
351 		   uint32_t hw_submission, unsigned hang_limit,
352 		   long timeout, struct workqueue_struct *timeout_wq,
353 		   atomic_t *score, const char *name);
354 
355 void drm_sched_fini(struct drm_gpu_scheduler *sched);
356 int drm_sched_job_init(struct drm_sched_job *job,
357 		       struct drm_sched_entity *entity,
358 		       void *owner);
359 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
360 				    struct drm_gpu_scheduler **sched_list,
361                                    unsigned int num_sched_list);
362 
363 void drm_sched_job_cleanup(struct drm_sched_job *job);
364 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
365 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
366 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
367 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
368 void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max);
369 void drm_sched_increase_karma(struct drm_sched_job *bad);
370 void drm_sched_reset_karma(struct drm_sched_job *bad);
371 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
372 bool drm_sched_dependency_optimized(struct dma_fence* fence,
373 				    struct drm_sched_entity *entity);
374 void drm_sched_fault(struct drm_gpu_scheduler *sched);
375 void drm_sched_job_kickout(struct drm_sched_job *s_job);
376 
377 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
378 			     struct drm_sched_entity *entity);
379 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
380 				struct drm_sched_entity *entity);
381 
382 int drm_sched_entity_init(struct drm_sched_entity *entity,
383 			  enum drm_sched_priority priority,
384 			  struct drm_gpu_scheduler **sched_list,
385 			  unsigned int num_sched_list,
386 			  atomic_t *guilty);
387 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
388 void drm_sched_entity_fini(struct drm_sched_entity *entity);
389 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
390 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
391 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
392 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
393 			       struct drm_sched_entity *entity);
394 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
395 				   enum drm_sched_priority priority);
396 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
397 
398 struct drm_sched_fence *drm_sched_fence_create(
399 	struct drm_sched_entity *s_entity, void *owner);
400 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
401 void drm_sched_fence_finished(struct drm_sched_fence *fence);
402 
403 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
404 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
405 		                unsigned long remaining);
406 struct drm_gpu_scheduler *
407 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
408 		     unsigned int num_sched_list);
409 
410 #endif
411