xref: /linux/include/drm/gpu_scheduler.h (revision e9126d09eeb514d349771c3b90c94535396ffe0a)
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 
30 #define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
31 
32 struct drm_gpu_scheduler;
33 struct drm_sched_rq;
34 
35 enum drm_sched_priority {
36 	DRM_SCHED_PRIORITY_MIN,
37 	DRM_SCHED_PRIORITY_LOW = DRM_SCHED_PRIORITY_MIN,
38 	DRM_SCHED_PRIORITY_NORMAL,
39 	DRM_SCHED_PRIORITY_HIGH_SW,
40 	DRM_SCHED_PRIORITY_HIGH_HW,
41 	DRM_SCHED_PRIORITY_KERNEL,
42 	DRM_SCHED_PRIORITY_MAX,
43 	DRM_SCHED_PRIORITY_INVALID = -1,
44 	DRM_SCHED_PRIORITY_UNSET = -2
45 };
46 
47 /**
48  * struct drm_sched_entity - A wrapper around a job queue (typically
49  * attached to the DRM file_priv).
50  *
51  * @list: used to append this struct to the list of entities in the
52  *        runqueue.
53  * @rq: runqueue on which this entity is currently scheduled.
54  * @rq_list: a list of run queues on which jobs from this entity can
55  *           be scheduled
56  * @num_rq_list: number of run queues in the rq_list
57  * @rq_lock: lock to modify the runqueue to which this entity belongs.
58  * @job_queue: the list of jobs of this entity.
59  * @fence_seq: a linearly increasing seqno incremented with each
60  *             new &drm_sched_fence which is part of the entity.
61  * @fence_context: a unique context for all the fences which belong
62  *                 to this entity.
63  *                 The &drm_sched_fence.scheduled uses the
64  *                 fence_context but &drm_sched_fence.finished uses
65  *                 fence_context + 1.
66  * @dependency: the dependency fence of the job which is on the top
67  *              of the job queue.
68  * @cb: callback for the dependency fence above.
69  * @guilty: points to ctx's guilty.
70  * @fini_status: contains the exit status in case the process was signalled.
71  * @last_scheduled: points to the finished fence of the last scheduled job.
72  * @last_user: last group leader pushing a job into the entity.
73  *
74  * Entities will emit jobs in order to their corresponding hardware
75  * ring, and the scheduler will alternate between entities based on
76  * scheduling policy.
77  */
78 struct drm_sched_entity {
79 	struct list_head		list;
80 	struct drm_sched_rq		*rq;
81 	struct drm_sched_rq		**rq_list;
82 	unsigned int                    num_rq_list;
83 	spinlock_t			rq_lock;
84 
85 	struct spsc_queue		job_queue;
86 
87 	atomic_t			fence_seq;
88 	uint64_t			fence_context;
89 
90 	struct dma_fence		*dependency;
91 	struct dma_fence_cb		cb;
92 	atomic_t			*guilty;
93 	struct dma_fence                *last_scheduled;
94 	struct task_struct		*last_user;
95 };
96 
97 /**
98  * struct drm_sched_rq - queue of entities to be scheduled.
99  *
100  * @lock: to modify the entities list.
101  * @sched: the scheduler to which this rq belongs to.
102  * @entities: list of the entities to be scheduled.
103  * @current_entity: the entity which is to be scheduled.
104  *
105  * Run queue is a set of entities scheduling command submissions for
106  * one specific ring. It implements the scheduling policy that selects
107  * the next entity to emit commands from.
108  */
109 struct drm_sched_rq {
110 	spinlock_t			lock;
111 	struct drm_gpu_scheduler	*sched;
112 	struct list_head		entities;
113 	struct drm_sched_entity		*current_entity;
114 };
115 
116 /**
117  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
118  */
119 struct drm_sched_fence {
120         /**
121          * @scheduled: this fence is what will be signaled by the scheduler
122          * when the job is scheduled.
123          */
124 	struct dma_fence		scheduled;
125 
126         /**
127          * @finished: this fence is what will be signaled by the scheduler
128          * when the job is completed.
129          *
130          * When setting up an out fence for the job, you should use
131          * this, since it's available immediately upon
132          * drm_sched_job_init(), and the fence returned by the driver
133          * from run_job() won't be created until the dependencies have
134          * resolved.
135          */
136 	struct dma_fence		finished;
137 
138         /**
139          * @cb: the callback for the parent fence below.
140          */
141 	struct dma_fence_cb		cb;
142         /**
143          * @parent: the fence returned by &drm_sched_backend_ops.run_job
144          * when scheduling the job on hardware. We signal the
145          * &drm_sched_fence.finished fence once parent is signalled.
146          */
147 	struct dma_fence		*parent;
148         /**
149          * @sched: the scheduler instance to which the job having this struct
150          * belongs to.
151          */
152 	struct drm_gpu_scheduler	*sched;
153         /**
154          * @lock: the lock used by the scheduled and the finished fences.
155          */
156 	spinlock_t			lock;
157         /**
158          * @owner: job owner for debugging
159          */
160 	void				*owner;
161 };
162 
163 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
164 
165 /**
166  * struct drm_sched_job - A job to be run by an entity.
167  *
168  * @queue_node: used to append this struct to the queue of jobs in an entity.
169  * @sched: the scheduler instance on which this job is scheduled.
170  * @s_fence: contains the fences for the scheduling of job.
171  * @finish_cb: the callback for the finished fence.
172  * @finish_work: schedules the function @drm_sched_job_finish once the job has
173  *               finished to remove the job from the
174  *               @drm_gpu_scheduler.ring_mirror_list.
175  * @node: used to append this struct to the @drm_gpu_scheduler.ring_mirror_list.
176  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the timeout
177  *            interval is over.
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  *
185  * A job is created by the driver using drm_sched_job_init(), and
186  * should call drm_sched_entity_push_job() once it wants the scheduler
187  * to schedule the job.
188  */
189 struct drm_sched_job {
190 	struct spsc_node		queue_node;
191 	struct drm_gpu_scheduler	*sched;
192 	struct drm_sched_fence		*s_fence;
193 	struct dma_fence_cb		finish_cb;
194 	struct work_struct		finish_work;
195 	struct list_head		node;
196 	struct delayed_work		work_tdr;
197 	uint64_t			id;
198 	atomic_t			karma;
199 	enum drm_sched_priority		s_priority;
200 	struct drm_sched_entity  *entity;
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 /**
210  * struct drm_sched_backend_ops
211  *
212  * Define the backend operations called by the scheduler,
213  * these functions should be implemented in driver side.
214  */
215 struct drm_sched_backend_ops {
216 	/**
217          * @dependency: Called when the scheduler is considering scheduling
218          * this job next, to get another struct dma_fence for this job to
219 	 * block on.  Once it returns NULL, run_job() may be called.
220 	 */
221 	struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
222 					struct drm_sched_entity *s_entity);
223 
224 	/**
225          * @run_job: Called to execute the job once all of the dependencies
226          * have been resolved.  This may be called multiple times, if
227 	 * timedout_job() has happened and drm_sched_job_recovery()
228 	 * decides to try it again.
229 	 */
230 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
231 
232 	/**
233          * @timedout_job: Called when a job has taken too long to execute,
234          * to trigger GPU recovery.
235 	 */
236 	void (*timedout_job)(struct drm_sched_job *sched_job);
237 
238 	/**
239          * @free_job: Called once the job's finished fence has been signaled
240          * and it's time to clean it up.
241 	 */
242 	void (*free_job)(struct drm_sched_job *sched_job);
243 };
244 
245 /**
246  * struct drm_gpu_scheduler
247  *
248  * @ops: backend operations provided by the driver.
249  * @hw_submission_limit: the max size of the hardware queue.
250  * @timeout: the time after which a job is removed from the scheduler.
251  * @name: name of the ring for which this scheduler is being used.
252  * @sched_rq: priority wise array of run queues.
253  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
254  *                  is ready to be scheduled.
255  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
256  *                 waits on this wait queue until all the scheduled jobs are
257  *                 finished.
258  * @hw_rq_count: the number of jobs currently in the hardware queue.
259  * @job_id_count: used to assign unique id to the each job.
260  * @thread: the kthread on which the scheduler which run.
261  * @ring_mirror_list: the list of jobs which are currently in the job queue.
262  * @job_list_lock: lock to protect the ring_mirror_list.
263  * @hang_limit: once the hangs by a job crosses this limit then it is marked
264  *              guilty and it will be considered for scheduling further.
265  * @num_jobs: the number of jobs in queue in the scheduler
266  *
267  * One scheduler is implemented for each hardware ring.
268  */
269 struct drm_gpu_scheduler {
270 	const struct drm_sched_backend_ops	*ops;
271 	uint32_t			hw_submission_limit;
272 	long				timeout;
273 	const char			*name;
274 	struct drm_sched_rq		sched_rq[DRM_SCHED_PRIORITY_MAX];
275 	wait_queue_head_t		wake_up_worker;
276 	wait_queue_head_t		job_scheduled;
277 	atomic_t			hw_rq_count;
278 	atomic64_t			job_id_count;
279 	struct task_struct		*thread;
280 	struct list_head		ring_mirror_list;
281 	spinlock_t			job_list_lock;
282 	int				hang_limit;
283 	atomic_t                        num_jobs;
284 };
285 
286 int drm_sched_init(struct drm_gpu_scheduler *sched,
287 		   const struct drm_sched_backend_ops *ops,
288 		   uint32_t hw_submission, unsigned hang_limit, long timeout,
289 		   const char *name);
290 void drm_sched_fini(struct drm_gpu_scheduler *sched);
291 
292 int drm_sched_entity_init(struct drm_sched_entity *entity,
293 			  struct drm_sched_rq **rq_list,
294 			  unsigned int num_rq_list,
295 			  atomic_t *guilty);
296 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
297 void drm_sched_entity_fini(struct drm_sched_entity *entity);
298 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
299 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
300 			       struct drm_sched_entity *entity);
301 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
302 				   enum drm_sched_priority priority);
303 struct drm_sched_fence *drm_sched_fence_create(
304 	struct drm_sched_entity *s_entity, void *owner);
305 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
306 void drm_sched_fence_finished(struct drm_sched_fence *fence);
307 int drm_sched_job_init(struct drm_sched_job *job,
308 		       struct drm_sched_entity *entity,
309 		       void *owner);
310 void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched,
311 			    struct drm_sched_job *job);
312 void drm_sched_job_recovery(struct drm_gpu_scheduler *sched);
313 bool drm_sched_dependency_optimized(struct dma_fence* fence,
314 				    struct drm_sched_entity *entity);
315 void drm_sched_job_kickout(struct drm_sched_job *s_job);
316 
317 #endif
318