Merge tag 'for-linus-20190620' of git://git.kernel.dk/linux-block
[linux-2.6-block.git] / include / drm / gpu_scheduler.h
CommitLineData
1b1f42d8
LS
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
741f01e6
AG
30#define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
31
1b1f42d8
LS
32struct drm_gpu_scheduler;
33struct drm_sched_rq;
34
35enum 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/**
2d33948e
ND
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.
ac0a6cf1
ND
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
2d33948e 57 * @rq_lock: lock to modify the runqueue to which this entity belongs.
2d33948e
ND
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.
43bce41c 72 * @last_user: last group leader pushing a job into the entity.
62347a33 73 * @stopped: Marks the enity as removed from rq and destined for termination.
1a61ee07
EA
74 *
75 * Entities will emit jobs in order to their corresponding hardware
76 * ring, and the scheduler will alternate between entities based on
77 * scheduling policy.
2d33948e 78 */
1b1f42d8
LS
79struct drm_sched_entity {
80 struct list_head list;
81 struct drm_sched_rq *rq;
ac0a6cf1
ND
82 struct drm_sched_rq **rq_list;
83 unsigned int num_rq_list;
1b1f42d8 84 spinlock_t rq_lock;
1b1f42d8 85
1b1f42d8
LS
86 struct spsc_queue job_queue;
87
88 atomic_t fence_seq;
89 uint64_t fence_context;
90
91 struct dma_fence *dependency;
92 struct dma_fence_cb cb;
2d33948e 93 atomic_t *guilty;
2d33948e 94 struct dma_fence *last_scheduled;
43bce41c 95 struct task_struct *last_user;
62347a33 96 bool stopped;
1b1f42d8
LS
97};
98
99/**
2d33948e
ND
100 * struct drm_sched_rq - queue of entities to be scheduled.
101 *
102 * @lock: to modify the entities list.
8dc9fbbf 103 * @sched: the scheduler to which this rq belongs to.
2d33948e
ND
104 * @entities: list of the entities to be scheduled.
105 * @current_entity: the entity which is to be scheduled.
106 *
1b1f42d8
LS
107 * Run queue is a set of entities scheduling command submissions for
108 * one specific ring. It implements the scheduling policy that selects
109 * the next entity to emit commands from.
2d33948e 110 */
1b1f42d8
LS
111struct drm_sched_rq {
112 spinlock_t lock;
8dc9fbbf 113 struct drm_gpu_scheduler *sched;
1b1f42d8
LS
114 struct list_head entities;
115 struct drm_sched_entity *current_entity;
116};
117
2d33948e
ND
118/**
119 * struct drm_sched_fence - fences corresponding to the scheduling of a job.
120 */
1b1f42d8 121struct drm_sched_fence {
2d33948e
ND
122 /**
123 * @scheduled: this fence is what will be signaled by the scheduler
124 * when the job is scheduled.
125 */
1b1f42d8 126 struct dma_fence scheduled;
1a61ee07 127
2d33948e
ND
128 /**
129 * @finished: this fence is what will be signaled by the scheduler
130 * when the job is completed.
131 *
132 * When setting up an out fence for the job, you should use
133 * this, since it's available immediately upon
134 * drm_sched_job_init(), and the fence returned by the driver
135 * from run_job() won't be created until the dependencies have
136 * resolved.
137 */
1b1f42d8 138 struct dma_fence finished;
1a61ee07 139
2d33948e
ND
140 /**
141 * @parent: the fence returned by &drm_sched_backend_ops.run_job
142 * when scheduling the job on hardware. We signal the
143 * &drm_sched_fence.finished fence once parent is signalled.
144 */
1b1f42d8 145 struct dma_fence *parent;
2d33948e
ND
146 /**
147 * @sched: the scheduler instance to which the job having this struct
148 * belongs to.
149 */
1b1f42d8 150 struct drm_gpu_scheduler *sched;
2d33948e
ND
151 /**
152 * @lock: the lock used by the scheduled and the finished fences.
153 */
1b1f42d8 154 spinlock_t lock;
2d33948e
ND
155 /**
156 * @owner: job owner for debugging
157 */
1b1f42d8
LS
158 void *owner;
159};
160
161struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
162
1a61ee07 163/**
2d33948e
ND
164 * struct drm_sched_job - A job to be run by an entity.
165 *
166 * @queue_node: used to append this struct to the queue of jobs in an entity.
167 * @sched: the scheduler instance on which this job is scheduled.
168 * @s_fence: contains the fences for the scheduling of job.
169 * @finish_cb: the callback for the finished fence.
170 * @finish_work: schedules the function @drm_sched_job_finish once the job has
171 * finished to remove the job from the
172 * @drm_gpu_scheduler.ring_mirror_list.
173 * @node: used to append this struct to the @drm_gpu_scheduler.ring_mirror_list.
2d33948e
ND
174 * @id: a unique id assigned to each job scheduled on the scheduler.
175 * @karma: increment on every hang caused by this job. If this exceeds the hang
176 * limit of the scheduler then the job is marked guilty and will not
177 * be scheduled further.
178 * @s_priority: the priority of the job.
179 * @entity: the entity to which this job belongs.
3741540e 180 * @cb: the callback for the parent fence in s_fence.
1a61ee07
EA
181 *
182 * A job is created by the driver using drm_sched_job_init(), and
183 * should call drm_sched_entity_push_job() once it wants the scheduler
184 * to schedule the job.
185 */
1b1f42d8
LS
186struct drm_sched_job {
187 struct spsc_node queue_node;
188 struct drm_gpu_scheduler *sched;
189 struct drm_sched_fence *s_fence;
190 struct dma_fence_cb finish_cb;
191 struct work_struct finish_work;
192 struct list_head node;
1b1f42d8
LS
193 uint64_t id;
194 atomic_t karma;
195 enum drm_sched_priority s_priority;
8ee3a52e 196 struct drm_sched_entity *entity;
3741540e 197 struct dma_fence_cb cb;
1b1f42d8
LS
198};
199
200static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
201 int threshold)
202{
203 return (s_job && atomic_inc_return(&s_job->karma) > threshold);
204}
205
206/**
2d33948e
ND
207 * struct drm_sched_backend_ops
208 *
1b1f42d8 209 * Define the backend operations called by the scheduler,
2d33948e
ND
210 * these functions should be implemented in driver side.
211 */
1b1f42d8 212struct drm_sched_backend_ops {
2d33948e
ND
213 /**
214 * @dependency: Called when the scheduler is considering scheduling
215 * this job next, to get another struct dma_fence for this job to
1a61ee07
EA
216 * block on. Once it returns NULL, run_job() may be called.
217 */
1b1f42d8
LS
218 struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
219 struct drm_sched_entity *s_entity);
1a61ee07 220
2d33948e
ND
221 /**
222 * @run_job: Called to execute the job once all of the dependencies
223 * have been resolved. This may be called multiple times, if
1a61ee07
EA
224 * timedout_job() has happened and drm_sched_job_recovery()
225 * decides to try it again.
226 */
1b1f42d8 227 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
1a61ee07 228
2d33948e
ND
229 /**
230 * @timedout_job: Called when a job has taken too long to execute,
231 * to trigger GPU recovery.
1a61ee07 232 */
1b1f42d8 233 void (*timedout_job)(struct drm_sched_job *sched_job);
1a61ee07 234
2d33948e
ND
235 /**
236 * @free_job: Called once the job's finished fence has been signaled
237 * and it's time to clean it up.
1a61ee07 238 */
1b1f42d8
LS
239 void (*free_job)(struct drm_sched_job *sched_job);
240};
241
242/**
2d33948e
ND
243 * struct drm_gpu_scheduler
244 *
245 * @ops: backend operations provided by the driver.
246 * @hw_submission_limit: the max size of the hardware queue.
247 * @timeout: the time after which a job is removed from the scheduler.
248 * @name: name of the ring for which this scheduler is being used.
249 * @sched_rq: priority wise array of run queues.
250 * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
251 * is ready to be scheduled.
252 * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
253 * waits on this wait queue until all the scheduled jobs are
254 * finished.
255 * @hw_rq_count: the number of jobs currently in the hardware queue.
256 * @job_id_count: used to assign unique id to the each job.
6a962430
ND
257 * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
258 * timeout interval is over.
2d33948e
ND
259 * @thread: the kthread on which the scheduler which run.
260 * @ring_mirror_list: the list of jobs which are currently in the job queue.
261 * @job_list_lock: lock to protect the ring_mirror_list.
262 * @hang_limit: once the hangs by a job crosses this limit then it is marked
263 * guilty and it will be considered for scheduling further.
249a07c0 264 * @num_jobs: the number of jobs in queue in the scheduler
faf6e1a8 265 * @ready: marks if the underlying HW is ready to work
2d33948e
ND
266 *
267 * One scheduler is implemented for each hardware ring.
268 */
1b1f42d8
LS
269struct 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;
6a962430 279 struct delayed_work work_tdr;
1b1f42d8
LS
280 struct task_struct *thread;
281 struct list_head ring_mirror_list;
282 spinlock_t job_list_lock;
283 int hang_limit;
249a07c0 284 atomic_t num_jobs;
faf6e1a8 285 bool ready;
1b1f42d8
LS
286};
287
288int drm_sched_init(struct drm_gpu_scheduler *sched,
289 const struct drm_sched_backend_ops *ops,
290 uint32_t hw_submission, unsigned hang_limit, long timeout,
291 const char *name);
faf6e1a8 292
1b1f42d8 293void drm_sched_fini(struct drm_gpu_scheduler *sched);
620e762f
CK
294int drm_sched_job_init(struct drm_sched_job *job,
295 struct drm_sched_entity *entity,
296 void *owner);
26efecf9 297void drm_sched_job_cleanup(struct drm_sched_job *job);
620e762f 298void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
222b5f04
AG
299void drm_sched_stop(struct drm_gpu_scheduler *sched);
300void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
301void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
302void drm_sched_increase_karma(struct drm_sched_job *bad);
620e762f
CK
303bool drm_sched_dependency_optimized(struct dma_fence* fence,
304 struct drm_sched_entity *entity);
8fe159b0 305void drm_sched_fault(struct drm_gpu_scheduler *sched);
620e762f
CK
306void drm_sched_job_kickout(struct drm_sched_job *s_job);
307
308void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
309 struct drm_sched_entity *entity);
310void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
311 struct drm_sched_entity *entity);
1b1f42d8 312
aa16b6c6
ND
313int drm_sched_entity_init(struct drm_sched_entity *entity,
314 struct drm_sched_rq **rq_list,
315 unsigned int num_rq_list,
8344c53f 316 atomic_t *guilty);
cdc50176
ND
317long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
318void drm_sched_entity_fini(struct drm_sched_entity *entity);
319void drm_sched_entity_destroy(struct drm_sched_entity *entity);
620e762f
CK
320void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
321struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
1b1f42d8
LS
322void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
323 struct drm_sched_entity *entity);
7febe4bf
CK
324void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
325 enum drm_sched_priority priority);
620e762f
CK
326bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
327
1b1f42d8
LS
328struct drm_sched_fence *drm_sched_fence_create(
329 struct drm_sched_entity *s_entity, void *owner);
330void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
331void drm_sched_fence_finished(struct drm_sched_fence *fence);
1b1f42d8 332
1db8c142
SM
333unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
334void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
335 unsigned long remaining);
336
1b1f42d8 337#endif