xref: /openbsd-src/sys/dev/pci/drm/include/drm/gpu_scheduler.h (revision 4e1ee0786f11cc571bd0be17d38e46f635c719fc)
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  * @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  * @node: used to append this struct to the @drm_gpu_scheduler.ring_mirror_list.
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 drm_gpu_scheduler	*sched;
197 	struct drm_sched_fence		*s_fence;
198 	struct dma_fence_cb		finish_cb;
199 	struct list_head		node;
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 /**
214  * struct drm_sched_backend_ops
215  *
216  * Define the backend operations called by the scheduler,
217  * these functions should be implemented in driver side.
218  */
219 struct drm_sched_backend_ops {
220 	/**
221          * @dependency: Called when the scheduler is considering scheduling
222          * this job next, to get another struct dma_fence for this job to
223 	 * block on.  Once it returns NULL, run_job() may be called.
224 	 */
225 	struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
226 					struct drm_sched_entity *s_entity);
227 
228 	/**
229          * @run_job: Called to execute the job once all of the dependencies
230          * have been resolved.  This may be called multiple times, if
231 	 * timedout_job() has happened and drm_sched_job_recovery()
232 	 * decides to try it again.
233 	 */
234 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
235 
236 	/**
237          * @timedout_job: Called when a job has taken too long to execute,
238          * to trigger GPU recovery.
239 	 */
240 	void (*timedout_job)(struct drm_sched_job *sched_job);
241 
242 	/**
243          * @free_job: Called once the job's finished fence has been signaled
244          * and it's time to clean it up.
245 	 */
246 	void (*free_job)(struct drm_sched_job *sched_job);
247 };
248 
249 /**
250  * struct drm_gpu_scheduler
251  *
252  * @ops: backend operations provided by the driver.
253  * @hw_submission_limit: the max size of the hardware queue.
254  * @timeout: the time after which a job is removed from the scheduler.
255  * @name: name of the ring for which this scheduler is being used.
256  * @sched_rq: priority wise array of run queues.
257  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
258  *                  is ready to be scheduled.
259  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
260  *                 waits on this wait queue until all the scheduled jobs are
261  *                 finished.
262  * @hw_rq_count: the number of jobs currently in the hardware queue.
263  * @job_id_count: used to assign unique id to the each job.
264  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
265  *            timeout interval is over.
266  * @thread: the kthread on which the scheduler which run.
267  * @ring_mirror_list: the list of jobs which are currently in the job queue.
268  * @job_list_lock: lock to protect the ring_mirror_list.
269  * @hang_limit: once the hangs by a job crosses this limit then it is marked
270  *              guilty and it will be considered for scheduling further.
271  * @score: score to help loadbalancer pick a idle sched
272  * @ready: marks if the underlying HW is ready to work
273  * @free_guilty: A hit to time out handler to free the guilty job.
274  *
275  * One scheduler is implemented for each hardware ring.
276  */
277 struct drm_gpu_scheduler {
278 	const struct drm_sched_backend_ops	*ops;
279 	uint32_t			hw_submission_limit;
280 	long				timeout;
281 	const char			*name;
282 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_COUNT];
283 	wait_queue_head_t		wake_up_worker;
284 	wait_queue_head_t		job_scheduled;
285 	atomic_t			hw_rq_count;
286 	atomic64_t			job_id_count;
287 	struct delayed_work		work_tdr;
288 #ifdef __linux__
289 	struct task_struct		*thread;
290 #else
291 	struct proc			*thread;
292 #endif
293 	struct list_head		ring_mirror_list;
294 	spinlock_t			job_list_lock;
295 	int				hang_limit;
296 	atomic_t                        score;
297 	bool				ready;
298 	bool				free_guilty;
299 };
300 
301 int drm_sched_init(struct drm_gpu_scheduler *sched,
302 		   const struct drm_sched_backend_ops *ops,
303 		   uint32_t hw_submission, unsigned hang_limit, long timeout,
304 		   const char *name);
305 
306 void drm_sched_fini(struct drm_gpu_scheduler *sched);
307 int drm_sched_job_init(struct drm_sched_job *job,
308 		       struct drm_sched_entity *entity,
309 		       void *owner);
310 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
311 				    struct drm_gpu_scheduler **sched_list,
312                                    unsigned int num_sched_list);
313 
314 void drm_sched_job_cleanup(struct drm_sched_job *job);
315 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
316 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
317 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
318 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
319 void drm_sched_increase_karma(struct drm_sched_job *bad);
320 bool drm_sched_dependency_optimized(struct dma_fence* fence,
321 				    struct drm_sched_entity *entity);
322 void drm_sched_fault(struct drm_gpu_scheduler *sched);
323 void drm_sched_job_kickout(struct drm_sched_job *s_job);
324 
325 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
326 			     struct drm_sched_entity *entity);
327 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
328 				struct drm_sched_entity *entity);
329 
330 int drm_sched_entity_init(struct drm_sched_entity *entity,
331 			  enum drm_sched_priority priority,
332 			  struct drm_gpu_scheduler **sched_list,
333 			  unsigned int num_sched_list,
334 			  atomic_t *guilty);
335 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
336 void drm_sched_entity_fini(struct drm_sched_entity *entity);
337 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
338 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
339 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
340 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
341 			       struct drm_sched_entity *entity);
342 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
343 				   enum drm_sched_priority priority);
344 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
345 
346 struct drm_sched_fence *drm_sched_fence_create(
347 	struct drm_sched_entity *s_entity, void *owner);
348 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
349 void drm_sched_fence_finished(struct drm_sched_fence *fence);
350 
351 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
352 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
353 		                unsigned long remaining);
354 struct drm_gpu_scheduler *
355 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
356 		     unsigned int num_sched_list);
357 
358 #endif
359