Loading...
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#include <linux/xarray.h>
31#include <linux/workqueue.h>
32
33#define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
34
35/**
36 * DRM_SCHED_FENCE_DONT_PIPELINE - Prefent dependency pipelining
37 *
38 * Setting this flag on a scheduler fence prevents pipelining of jobs depending
39 * on this fence. In other words we always insert a full CPU round trip before
40 * dependen jobs are pushed to the hw queue.
41 */
42#define DRM_SCHED_FENCE_DONT_PIPELINE DMA_FENCE_FLAG_USER_BITS
43
44enum dma_resv_usage;
45struct dma_resv;
46struct drm_gem_object;
47
48struct drm_gpu_scheduler;
49struct drm_sched_rq;
50
51/* These are often used as an (initial) index
52 * to an array, and as such should start at 0.
53 */
54enum drm_sched_priority {
55 DRM_SCHED_PRIORITY_MIN,
56 DRM_SCHED_PRIORITY_NORMAL,
57 DRM_SCHED_PRIORITY_HIGH,
58 DRM_SCHED_PRIORITY_KERNEL,
59
60 DRM_SCHED_PRIORITY_COUNT,
61 DRM_SCHED_PRIORITY_UNSET = -2
62};
63
64/* Used to chose between FIFO and RR jobs scheduling */
65extern int drm_sched_policy;
66
67#define DRM_SCHED_POLICY_RR 0
68#define DRM_SCHED_POLICY_FIFO 1
69
70/**
71 * struct drm_sched_entity - A wrapper around a job queue (typically
72 * attached to the DRM file_priv).
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 */
78struct drm_sched_entity {
79 /**
80 * @list:
81 *
82 * Used to append this struct to the list of entities in the runqueue
83 * @rq under &drm_sched_rq.entities.
84 *
85 * Protected by &drm_sched_rq.lock of @rq.
86 */
87 struct list_head list;
88
89 /**
90 * @rq:
91 *
92 * Runqueue on which this entity is currently scheduled.
93 *
94 * FIXME: Locking is very unclear for this. Writers are protected by
95 * @rq_lock, but readers are generally lockless and seem to just race
96 * with not even a READ_ONCE.
97 */
98 struct drm_sched_rq *rq;
99
100 /**
101 * @sched_list:
102 *
103 * A list of schedulers (struct drm_gpu_scheduler). Jobs from this entity can
104 * be scheduled on any scheduler on this list.
105 *
106 * This can be modified by calling drm_sched_entity_modify_sched().
107 * Locking is entirely up to the driver, see the above function for more
108 * details.
109 *
110 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
111 * set already.
112 *
113 * FIXME: This means priority changes through
114 * drm_sched_entity_set_priority() will be lost henceforth in this case.
115 */
116 struct drm_gpu_scheduler **sched_list;
117
118 /**
119 * @num_sched_list:
120 *
121 * Number of drm_gpu_schedulers in the @sched_list.
122 */
123 unsigned int num_sched_list;
124
125 /**
126 * @priority:
127 *
128 * Priority of the entity. This can be modified by calling
129 * drm_sched_entity_set_priority(). Protected by &rq_lock.
130 */
131 enum drm_sched_priority priority;
132
133 /**
134 * @rq_lock:
135 *
136 * Lock to modify the runqueue to which this entity belongs.
137 */
138 spinlock_t rq_lock;
139
140 /**
141 * @job_queue: the list of jobs of this entity.
142 */
143 struct spsc_queue job_queue;
144
145 /**
146 * @fence_seq:
147 *
148 * A linearly increasing seqno incremented with each new
149 * &drm_sched_fence which is part of the entity.
150 *
151 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
152 * this doesn't need to be atomic.
153 */
154 atomic_t fence_seq;
155
156 /**
157 * @fence_context:
158 *
159 * A unique context for all the fences which belong to this entity. The
160 * &drm_sched_fence.scheduled uses the fence_context but
161 * &drm_sched_fence.finished uses fence_context + 1.
162 */
163 uint64_t fence_context;
164
165 /**
166 * @dependency:
167 *
168 * The dependency fence of the job which is on the top of the job queue.
169 */
170 struct dma_fence *dependency;
171
172 /**
173 * @cb:
174 *
175 * Callback for the dependency fence above.
176 */
177 struct dma_fence_cb cb;
178
179 /**
180 * @guilty:
181 *
182 * Points to entities' guilty.
183 */
184 atomic_t *guilty;
185
186 /**
187 * @last_scheduled:
188 *
189 * Points to the finished fence of the last scheduled job. Only written
190 * by the scheduler thread, can be accessed locklessly from
191 * drm_sched_job_arm() iff the queue is empty.
192 */
193 struct dma_fence *last_scheduled;
194
195 /**
196 * @last_user: last group leader pushing a job into the entity.
197 */
198 struct task_struct *last_user;
199
200 /**
201 * @stopped:
202 *
203 * Marks the enity as removed from rq and destined for
204 * termination. This is set by calling drm_sched_entity_flush() and by
205 * drm_sched_fini().
206 */
207 bool stopped;
208
209 /**
210 * @entity_idle:
211 *
212 * Signals when entity is not in use, used to sequence entity cleanup in
213 * drm_sched_entity_fini().
214 */
215 struct completion entity_idle;
216
217 /**
218 * @oldest_job_waiting:
219 *
220 * Marks earliest job waiting in SW queue
221 */
222 ktime_t oldest_job_waiting;
223
224 /**
225 * @rb_tree_node:
226 *
227 * The node used to insert this entity into time based priority queue
228 */
229 struct rb_node rb_tree_node;
230
231};
232
233/**
234 * struct drm_sched_rq - queue of entities to be scheduled.
235 *
236 * @lock: to modify the entities list.
237 * @sched: the scheduler to which this rq belongs to.
238 * @entities: list of the entities to be scheduled.
239 * @current_entity: the entity which is to be scheduled.
240 * @rb_tree_root: root of time based priory queue of entities for FIFO scheduling
241 *
242 * Run queue is a set of entities scheduling command submissions for
243 * one specific ring. It implements the scheduling policy that selects
244 * the next entity to emit commands from.
245 */
246struct drm_sched_rq {
247 spinlock_t lock;
248 struct drm_gpu_scheduler *sched;
249 struct list_head entities;
250 struct drm_sched_entity *current_entity;
251 struct rb_root_cached rb_tree_root;
252};
253
254/**
255 * struct drm_sched_fence - fences corresponding to the scheduling of a job.
256 */
257struct drm_sched_fence {
258 /**
259 * @scheduled: this fence is what will be signaled by the scheduler
260 * when the job is scheduled.
261 */
262 struct dma_fence scheduled;
263
264 /**
265 * @finished: this fence is what will be signaled by the scheduler
266 * when the job is completed.
267 *
268 * When setting up an out fence for the job, you should use
269 * this, since it's available immediately upon
270 * drm_sched_job_init(), and the fence returned by the driver
271 * from run_job() won't be created until the dependencies have
272 * resolved.
273 */
274 struct dma_fence finished;
275
276 /**
277 * @parent: the fence returned by &drm_sched_backend_ops.run_job
278 * when scheduling the job on hardware. We signal the
279 * &drm_sched_fence.finished fence once parent is signalled.
280 */
281 struct dma_fence *parent;
282 /**
283 * @sched: the scheduler instance to which the job having this struct
284 * belongs to.
285 */
286 struct drm_gpu_scheduler *sched;
287 /**
288 * @lock: the lock used by the scheduled and the finished fences.
289 */
290 spinlock_t lock;
291 /**
292 * @owner: job owner for debugging
293 */
294 void *owner;
295};
296
297struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
298
299/**
300 * struct drm_sched_job - A job to be run by an entity.
301 *
302 * @queue_node: used to append this struct to the queue of jobs in an entity.
303 * @list: a job participates in a "pending" and "done" lists.
304 * @sched: the scheduler instance on which this job is scheduled.
305 * @s_fence: contains the fences for the scheduling of job.
306 * @finish_cb: the callback for the finished fence.
307 * @work: Helper to reschdeule job kill to different context.
308 * @id: a unique id assigned to each job scheduled on the scheduler.
309 * @karma: increment on every hang caused by this job. If this exceeds the hang
310 * limit of the scheduler then the job is marked guilty and will not
311 * be scheduled further.
312 * @s_priority: the priority of the job.
313 * @entity: the entity to which this job belongs.
314 * @cb: the callback for the parent fence in s_fence.
315 *
316 * A job is created by the driver using drm_sched_job_init(), and
317 * should call drm_sched_entity_push_job() once it wants the scheduler
318 * to schedule the job.
319 */
320struct drm_sched_job {
321 struct spsc_node queue_node;
322 struct list_head list;
323 struct drm_gpu_scheduler *sched;
324 struct drm_sched_fence *s_fence;
325
326 /*
327 * work is used only after finish_cb has been used and will not be
328 * accessed anymore.
329 */
330 union {
331 struct dma_fence_cb finish_cb;
332 struct work_struct work;
333 };
334
335 uint64_t id;
336 atomic_t karma;
337 enum drm_sched_priority s_priority;
338 struct drm_sched_entity *entity;
339 struct dma_fence_cb cb;
340 /**
341 * @dependencies:
342 *
343 * Contains the dependencies as struct dma_fence for this job, see
344 * drm_sched_job_add_dependency() and
345 * drm_sched_job_add_implicit_dependencies().
346 */
347 struct xarray dependencies;
348
349 /** @last_dependency: tracks @dependencies as they signal */
350 unsigned long last_dependency;
351
352 /**
353 * @submit_ts:
354 *
355 * When the job was pushed into the entity queue.
356 */
357 ktime_t submit_ts;
358};
359
360static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
361 int threshold)
362{
363 return s_job && atomic_inc_return(&s_job->karma) > threshold;
364}
365
366enum drm_gpu_sched_stat {
367 DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
368 DRM_GPU_SCHED_STAT_NOMINAL,
369 DRM_GPU_SCHED_STAT_ENODEV,
370};
371
372/**
373 * struct drm_sched_backend_ops - Define the backend operations
374 * called by the scheduler
375 *
376 * These functions should be implemented in the driver side.
377 */
378struct drm_sched_backend_ops {
379 /**
380 * @prepare_job:
381 *
382 * Called when the scheduler is considering scheduling this job next, to
383 * get another struct dma_fence for this job to block on. Once it
384 * returns NULL, run_job() may be called.
385 *
386 * Can be NULL if no additional preparation to the dependencies are
387 * necessary. Skipped when jobs are killed instead of run.
388 */
389 struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job,
390 struct drm_sched_entity *s_entity);
391
392 /**
393 * @run_job: Called to execute the job once all of the dependencies
394 * have been resolved. This may be called multiple times, if
395 * timedout_job() has happened and drm_sched_job_recovery()
396 * decides to try it again.
397 */
398 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
399
400 /**
401 * @timedout_job: Called when a job has taken too long to execute,
402 * to trigger GPU recovery.
403 *
404 * This method is called in a workqueue context.
405 *
406 * Drivers typically issue a reset to recover from GPU hangs, and this
407 * procedure usually follows the following workflow:
408 *
409 * 1. Stop the scheduler using drm_sched_stop(). This will park the
410 * scheduler thread and cancel the timeout work, guaranteeing that
411 * nothing is queued while we reset the hardware queue
412 * 2. Try to gracefully stop non-faulty jobs (optional)
413 * 3. Issue a GPU reset (driver-specific)
414 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
415 * 5. Restart the scheduler using drm_sched_start(). At that point, new
416 * jobs can be queued, and the scheduler thread is unblocked
417 *
418 * Note that some GPUs have distinct hardware queues but need to reset
419 * the GPU globally, which requires extra synchronization between the
420 * timeout handler of the different &drm_gpu_scheduler. One way to
421 * achieve this synchronization is to create an ordered workqueue
422 * (using alloc_ordered_workqueue()) at the driver level, and pass this
423 * queue to drm_sched_init(), to guarantee that timeout handlers are
424 * executed sequentially. The above workflow needs to be slightly
425 * adjusted in that case:
426 *
427 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
428 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
429 * the reset (optional)
430 * 3. Issue a GPU reset on all faulty queues (driver-specific)
431 * 4. Re-submit jobs on all schedulers impacted by the reset using
432 * drm_sched_resubmit_jobs()
433 * 5. Restart all schedulers that were stopped in step #1 using
434 * drm_sched_start()
435 *
436 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
437 * and the underlying driver has started or completed recovery.
438 *
439 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
440 * available, i.e. has been unplugged.
441 */
442 enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
443
444 /**
445 * @free_job: Called once the job's finished fence has been signaled
446 * and it's time to clean it up.
447 */
448 void (*free_job)(struct drm_sched_job *sched_job);
449};
450
451/**
452 * struct drm_gpu_scheduler - scheduler instance-specific data
453 *
454 * @ops: backend operations provided by the driver.
455 * @hw_submission_limit: the max size of the hardware queue.
456 * @timeout: the time after which a job is removed from the scheduler.
457 * @name: name of the ring for which this scheduler is being used.
458 * @sched_rq: priority wise array of run queues.
459 * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
460 * is ready to be scheduled.
461 * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
462 * waits on this wait queue until all the scheduled jobs are
463 * finished.
464 * @hw_rq_count: the number of jobs currently in the hardware queue.
465 * @job_id_count: used to assign unique id to the each job.
466 * @timeout_wq: workqueue used to queue @work_tdr
467 * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
468 * timeout interval is over.
469 * @thread: the kthread on which the scheduler which run.
470 * @pending_list: the list of jobs which are currently in the job queue.
471 * @job_list_lock: lock to protect the pending_list.
472 * @hang_limit: once the hangs by a job crosses this limit then it is marked
473 * guilty and it will no longer be considered for scheduling.
474 * @score: score to help loadbalancer pick a idle sched
475 * @_score: score used when the driver doesn't provide one
476 * @ready: marks if the underlying HW is ready to work
477 * @free_guilty: A hit to time out handler to free the guilty job.
478 * @dev: system &struct device
479 *
480 * One scheduler is implemented for each hardware ring.
481 */
482struct drm_gpu_scheduler {
483 const struct drm_sched_backend_ops *ops;
484 uint32_t hw_submission_limit;
485 long timeout;
486 const char *name;
487 struct drm_sched_rq sched_rq[DRM_SCHED_PRIORITY_COUNT];
488 wait_queue_head_t wake_up_worker;
489 wait_queue_head_t job_scheduled;
490 atomic_t hw_rq_count;
491 atomic64_t job_id_count;
492 struct workqueue_struct *timeout_wq;
493 struct delayed_work work_tdr;
494 struct task_struct *thread;
495 struct list_head pending_list;
496 spinlock_t job_list_lock;
497 int hang_limit;
498 atomic_t *score;
499 atomic_t _score;
500 bool ready;
501 bool free_guilty;
502 struct device *dev;
503};
504
505int drm_sched_init(struct drm_gpu_scheduler *sched,
506 const struct drm_sched_backend_ops *ops,
507 uint32_t hw_submission, unsigned hang_limit,
508 long timeout, struct workqueue_struct *timeout_wq,
509 atomic_t *score, const char *name, struct device *dev);
510
511void drm_sched_fini(struct drm_gpu_scheduler *sched);
512int drm_sched_job_init(struct drm_sched_job *job,
513 struct drm_sched_entity *entity,
514 void *owner);
515void drm_sched_job_arm(struct drm_sched_job *job);
516int drm_sched_job_add_dependency(struct drm_sched_job *job,
517 struct dma_fence *fence);
518int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
519 struct dma_resv *resv,
520 enum dma_resv_usage usage);
521int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
522 struct drm_gem_object *obj,
523 bool write);
524
525
526void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
527 struct drm_gpu_scheduler **sched_list,
528 unsigned int num_sched_list);
529
530void drm_sched_job_cleanup(struct drm_sched_job *job);
531void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
532void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
533void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
534void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
535void drm_sched_increase_karma(struct drm_sched_job *bad);
536void drm_sched_reset_karma(struct drm_sched_job *bad);
537void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
538bool drm_sched_dependency_optimized(struct dma_fence* fence,
539 struct drm_sched_entity *entity);
540void drm_sched_fault(struct drm_gpu_scheduler *sched);
541void drm_sched_job_kickout(struct drm_sched_job *s_job);
542
543void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
544 struct drm_sched_entity *entity);
545void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
546 struct drm_sched_entity *entity);
547
548void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts);
549
550int drm_sched_entity_init(struct drm_sched_entity *entity,
551 enum drm_sched_priority priority,
552 struct drm_gpu_scheduler **sched_list,
553 unsigned int num_sched_list,
554 atomic_t *guilty);
555long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
556void drm_sched_entity_fini(struct drm_sched_entity *entity);
557void drm_sched_entity_destroy(struct drm_sched_entity *entity);
558void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
559struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
560void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
561void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
562 enum drm_sched_priority priority);
563bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
564
565struct drm_sched_fence *drm_sched_fence_alloc(
566 struct drm_sched_entity *s_entity, void *owner);
567void drm_sched_fence_init(struct drm_sched_fence *fence,
568 struct drm_sched_entity *entity);
569void drm_sched_fence_free(struct drm_sched_fence *fence);
570
571void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
572void drm_sched_fence_finished(struct drm_sched_fence *fence);
573
574unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
575void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
576 unsigned long remaining);
577struct drm_gpu_scheduler *
578drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
579 unsigned int num_sched_list);
580
581#endif
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#include <linux/xarray.h>
31#include <linux/workqueue.h>
32
33#define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
34
35/**
36 * DRM_SCHED_FENCE_DONT_PIPELINE - Prefent dependency pipelining
37 *
38 * Setting this flag on a scheduler fence prevents pipelining of jobs depending
39 * on this fence. In other words we always insert a full CPU round trip before
40 * dependen jobs are pushed to the hw queue.
41 */
42#define DRM_SCHED_FENCE_DONT_PIPELINE DMA_FENCE_FLAG_USER_BITS
43
44/**
45 * DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT - A fence deadline hint has been set
46 *
47 * Because we could have a deadline hint can be set before the backing hw
48 * fence is created, we need to keep track of whether a deadline has already
49 * been set.
50 */
51#define DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT (DMA_FENCE_FLAG_USER_BITS + 1)
52
53enum dma_resv_usage;
54struct dma_resv;
55struct drm_gem_object;
56
57struct drm_gpu_scheduler;
58struct drm_sched_rq;
59
60struct drm_file;
61
62/* These are often used as an (initial) index
63 * to an array, and as such should start at 0.
64 */
65enum drm_sched_priority {
66 DRM_SCHED_PRIORITY_KERNEL,
67 DRM_SCHED_PRIORITY_HIGH,
68 DRM_SCHED_PRIORITY_NORMAL,
69 DRM_SCHED_PRIORITY_LOW,
70
71 DRM_SCHED_PRIORITY_COUNT
72};
73
74/* Used to chose between FIFO and RR jobs scheduling */
75extern int drm_sched_policy;
76
77#define DRM_SCHED_POLICY_RR 0
78#define DRM_SCHED_POLICY_FIFO 1
79
80/**
81 * struct drm_sched_entity - A wrapper around a job queue (typically
82 * attached to the DRM file_priv).
83 *
84 * Entities will emit jobs in order to their corresponding hardware
85 * ring, and the scheduler will alternate between entities based on
86 * scheduling policy.
87 */
88struct drm_sched_entity {
89 /**
90 * @list:
91 *
92 * Used to append this struct to the list of entities in the runqueue
93 * @rq under &drm_sched_rq.entities.
94 *
95 * Protected by &drm_sched_rq.lock of @rq.
96 */
97 struct list_head list;
98
99 /**
100 * @rq:
101 *
102 * Runqueue on which this entity is currently scheduled.
103 *
104 * FIXME: Locking is very unclear for this. Writers are protected by
105 * @rq_lock, but readers are generally lockless and seem to just race
106 * with not even a READ_ONCE.
107 */
108 struct drm_sched_rq *rq;
109
110 /**
111 * @sched_list:
112 *
113 * A list of schedulers (struct drm_gpu_scheduler). Jobs from this entity can
114 * be scheduled on any scheduler on this list.
115 *
116 * This can be modified by calling drm_sched_entity_modify_sched().
117 * Locking is entirely up to the driver, see the above function for more
118 * details.
119 *
120 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
121 * set already.
122 *
123 * FIXME: This means priority changes through
124 * drm_sched_entity_set_priority() will be lost henceforth in this case.
125 */
126 struct drm_gpu_scheduler **sched_list;
127
128 /**
129 * @num_sched_list:
130 *
131 * Number of drm_gpu_schedulers in the @sched_list.
132 */
133 unsigned int num_sched_list;
134
135 /**
136 * @priority:
137 *
138 * Priority of the entity. This can be modified by calling
139 * drm_sched_entity_set_priority(). Protected by &rq_lock.
140 */
141 enum drm_sched_priority priority;
142
143 /**
144 * @rq_lock:
145 *
146 * Lock to modify the runqueue to which this entity belongs.
147 */
148 spinlock_t rq_lock;
149
150 /**
151 * @job_queue: the list of jobs of this entity.
152 */
153 struct spsc_queue job_queue;
154
155 /**
156 * @fence_seq:
157 *
158 * A linearly increasing seqno incremented with each new
159 * &drm_sched_fence which is part of the entity.
160 *
161 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
162 * this doesn't need to be atomic.
163 */
164 atomic_t fence_seq;
165
166 /**
167 * @fence_context:
168 *
169 * A unique context for all the fences which belong to this entity. The
170 * &drm_sched_fence.scheduled uses the fence_context but
171 * &drm_sched_fence.finished uses fence_context + 1.
172 */
173 uint64_t fence_context;
174
175 /**
176 * @dependency:
177 *
178 * The dependency fence of the job which is on the top of the job queue.
179 */
180 struct dma_fence *dependency;
181
182 /**
183 * @cb:
184 *
185 * Callback for the dependency fence above.
186 */
187 struct dma_fence_cb cb;
188
189 /**
190 * @guilty:
191 *
192 * Points to entities' guilty.
193 */
194 atomic_t *guilty;
195
196 /**
197 * @last_scheduled:
198 *
199 * Points to the finished fence of the last scheduled job. Only written
200 * by the scheduler thread, can be accessed locklessly from
201 * drm_sched_job_arm() iff the queue is empty.
202 */
203 struct dma_fence __rcu *last_scheduled;
204
205 /**
206 * @last_user: last group leader pushing a job into the entity.
207 */
208 struct task_struct *last_user;
209
210 /**
211 * @stopped:
212 *
213 * Marks the enity as removed from rq and destined for
214 * termination. This is set by calling drm_sched_entity_flush() and by
215 * drm_sched_fini().
216 */
217 bool stopped;
218
219 /**
220 * @entity_idle:
221 *
222 * Signals when entity is not in use, used to sequence entity cleanup in
223 * drm_sched_entity_fini().
224 */
225 struct completion entity_idle;
226
227 /**
228 * @oldest_job_waiting:
229 *
230 * Marks earliest job waiting in SW queue
231 */
232 ktime_t oldest_job_waiting;
233
234 /**
235 * @rb_tree_node:
236 *
237 * The node used to insert this entity into time based priority queue
238 */
239 struct rb_node rb_tree_node;
240
241};
242
243/**
244 * struct drm_sched_rq - queue of entities to be scheduled.
245 *
246 * @lock: to modify the entities list.
247 * @sched: the scheduler to which this rq belongs to.
248 * @entities: list of the entities to be scheduled.
249 * @current_entity: the entity which is to be scheduled.
250 * @rb_tree_root: root of time based priory queue of entities for FIFO scheduling
251 *
252 * Run queue is a set of entities scheduling command submissions for
253 * one specific ring. It implements the scheduling policy that selects
254 * the next entity to emit commands from.
255 */
256struct drm_sched_rq {
257 spinlock_t lock;
258 struct drm_gpu_scheduler *sched;
259 struct list_head entities;
260 struct drm_sched_entity *current_entity;
261 struct rb_root_cached rb_tree_root;
262};
263
264/**
265 * struct drm_sched_fence - fences corresponding to the scheduling of a job.
266 */
267struct drm_sched_fence {
268 /**
269 * @scheduled: this fence is what will be signaled by the scheduler
270 * when the job is scheduled.
271 */
272 struct dma_fence scheduled;
273
274 /**
275 * @finished: this fence is what will be signaled by the scheduler
276 * when the job is completed.
277 *
278 * When setting up an out fence for the job, you should use
279 * this, since it's available immediately upon
280 * drm_sched_job_init(), and the fence returned by the driver
281 * from run_job() won't be created until the dependencies have
282 * resolved.
283 */
284 struct dma_fence finished;
285
286 /**
287 * @deadline: deadline set on &drm_sched_fence.finished which
288 * potentially needs to be propagated to &drm_sched_fence.parent
289 */
290 ktime_t deadline;
291
292 /**
293 * @parent: the fence returned by &drm_sched_backend_ops.run_job
294 * when scheduling the job on hardware. We signal the
295 * &drm_sched_fence.finished fence once parent is signalled.
296 */
297 struct dma_fence *parent;
298 /**
299 * @sched: the scheduler instance to which the job having this struct
300 * belongs to.
301 */
302 struct drm_gpu_scheduler *sched;
303 /**
304 * @lock: the lock used by the scheduled and the finished fences.
305 */
306 spinlock_t lock;
307 /**
308 * @owner: job owner for debugging
309 */
310 void *owner;
311};
312
313struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
314
315/**
316 * struct drm_sched_job - A job to be run by an entity.
317 *
318 * @queue_node: used to append this struct to the queue of jobs in an entity.
319 * @list: a job participates in a "pending" and "done" lists.
320 * @sched: the scheduler instance on which this job is scheduled.
321 * @s_fence: contains the fences for the scheduling of job.
322 * @finish_cb: the callback for the finished fence.
323 * @credits: the number of credits this job contributes to the scheduler
324 * @work: Helper to reschdeule job kill to different context.
325 * @id: a unique id assigned to each job scheduled on the scheduler.
326 * @karma: increment on every hang caused by this job. If this exceeds the hang
327 * limit of the scheduler then the job is marked guilty and will not
328 * be scheduled further.
329 * @s_priority: the priority of the job.
330 * @entity: the entity to which this job belongs.
331 * @cb: the callback for the parent fence in s_fence.
332 *
333 * A job is created by the driver using drm_sched_job_init(), and
334 * should call drm_sched_entity_push_job() once it wants the scheduler
335 * to schedule the job.
336 */
337struct drm_sched_job {
338 struct spsc_node queue_node;
339 struct list_head list;
340 struct drm_gpu_scheduler *sched;
341 struct drm_sched_fence *s_fence;
342
343 u32 credits;
344
345 /*
346 * work is used only after finish_cb has been used and will not be
347 * accessed anymore.
348 */
349 union {
350 struct dma_fence_cb finish_cb;
351 struct work_struct work;
352 };
353
354 uint64_t id;
355 atomic_t karma;
356 enum drm_sched_priority s_priority;
357 struct drm_sched_entity *entity;
358 struct dma_fence_cb cb;
359 /**
360 * @dependencies:
361 *
362 * Contains the dependencies as struct dma_fence for this job, see
363 * drm_sched_job_add_dependency() and
364 * drm_sched_job_add_implicit_dependencies().
365 */
366 struct xarray dependencies;
367
368 /** @last_dependency: tracks @dependencies as they signal */
369 unsigned long last_dependency;
370
371 /**
372 * @submit_ts:
373 *
374 * When the job was pushed into the entity queue.
375 */
376 ktime_t submit_ts;
377};
378
379static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
380 int threshold)
381{
382 return s_job && atomic_inc_return(&s_job->karma) > threshold;
383}
384
385enum drm_gpu_sched_stat {
386 DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
387 DRM_GPU_SCHED_STAT_NOMINAL,
388 DRM_GPU_SCHED_STAT_ENODEV,
389};
390
391/**
392 * struct drm_sched_backend_ops - Define the backend operations
393 * called by the scheduler
394 *
395 * These functions should be implemented in the driver side.
396 */
397struct drm_sched_backend_ops {
398 /**
399 * @prepare_job:
400 *
401 * Called when the scheduler is considering scheduling this job next, to
402 * get another struct dma_fence for this job to block on. Once it
403 * returns NULL, run_job() may be called.
404 *
405 * Can be NULL if no additional preparation to the dependencies are
406 * necessary. Skipped when jobs are killed instead of run.
407 */
408 struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job,
409 struct drm_sched_entity *s_entity);
410
411 /**
412 * @run_job: Called to execute the job once all of the dependencies
413 * have been resolved. This may be called multiple times, if
414 * timedout_job() has happened and drm_sched_job_recovery()
415 * decides to try it again.
416 */
417 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
418
419 /**
420 * @timedout_job: Called when a job has taken too long to execute,
421 * to trigger GPU recovery.
422 *
423 * This method is called in a workqueue context.
424 *
425 * Drivers typically issue a reset to recover from GPU hangs, and this
426 * procedure usually follows the following workflow:
427 *
428 * 1. Stop the scheduler using drm_sched_stop(). This will park the
429 * scheduler thread and cancel the timeout work, guaranteeing that
430 * nothing is queued while we reset the hardware queue
431 * 2. Try to gracefully stop non-faulty jobs (optional)
432 * 3. Issue a GPU reset (driver-specific)
433 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
434 * 5. Restart the scheduler using drm_sched_start(). At that point, new
435 * jobs can be queued, and the scheduler thread is unblocked
436 *
437 * Note that some GPUs have distinct hardware queues but need to reset
438 * the GPU globally, which requires extra synchronization between the
439 * timeout handler of the different &drm_gpu_scheduler. One way to
440 * achieve this synchronization is to create an ordered workqueue
441 * (using alloc_ordered_workqueue()) at the driver level, and pass this
442 * queue to drm_sched_init(), to guarantee that timeout handlers are
443 * executed sequentially. The above workflow needs to be slightly
444 * adjusted in that case:
445 *
446 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
447 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
448 * the reset (optional)
449 * 3. Issue a GPU reset on all faulty queues (driver-specific)
450 * 4. Re-submit jobs on all schedulers impacted by the reset using
451 * drm_sched_resubmit_jobs()
452 * 5. Restart all schedulers that were stopped in step #1 using
453 * drm_sched_start()
454 *
455 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
456 * and the underlying driver has started or completed recovery.
457 *
458 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
459 * available, i.e. has been unplugged.
460 */
461 enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
462
463 /**
464 * @free_job: Called once the job's finished fence has been signaled
465 * and it's time to clean it up.
466 */
467 void (*free_job)(struct drm_sched_job *sched_job);
468
469 /**
470 * @update_job_credits: Called when the scheduler is considering this
471 * job for execution.
472 *
473 * This callback returns the number of credits the job would take if
474 * pushed to the hardware. Drivers may use this to dynamically update
475 * the job's credit count. For instance, deduct the number of credits
476 * for already signalled native fences.
477 *
478 * This callback is optional.
479 */
480 u32 (*update_job_credits)(struct drm_sched_job *sched_job);
481};
482
483/**
484 * struct drm_gpu_scheduler - scheduler instance-specific data
485 *
486 * @ops: backend operations provided by the driver.
487 * @credit_limit: the credit limit of this scheduler
488 * @credit_count: the current credit count of this scheduler
489 * @timeout: the time after which a job is removed from the scheduler.
490 * @name: name of the ring for which this scheduler is being used.
491 * @num_rqs: Number of run-queues. This is at most DRM_SCHED_PRIORITY_COUNT,
492 * as there's usually one run-queue per priority, but could be less.
493 * @sched_rq: An allocated array of run-queues of size @num_rqs;
494 * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
495 * waits on this wait queue until all the scheduled jobs are
496 * finished.
497 * @job_id_count: used to assign unique id to the each job.
498 * @submit_wq: workqueue used to queue @work_run_job and @work_free_job
499 * @timeout_wq: workqueue used to queue @work_tdr
500 * @work_run_job: work which calls run_job op of each scheduler.
501 * @work_free_job: work which calls free_job op of each scheduler.
502 * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
503 * timeout interval is over.
504 * @pending_list: the list of jobs which are currently in the job queue.
505 * @job_list_lock: lock to protect the pending_list.
506 * @hang_limit: once the hangs by a job crosses this limit then it is marked
507 * guilty and it will no longer be considered for scheduling.
508 * @score: score to help loadbalancer pick a idle sched
509 * @_score: score used when the driver doesn't provide one
510 * @ready: marks if the underlying HW is ready to work
511 * @free_guilty: A hit to time out handler to free the guilty job.
512 * @pause_submit: pause queuing of @work_run_job on @submit_wq
513 * @own_submit_wq: scheduler owns allocation of @submit_wq
514 * @dev: system &struct device
515 *
516 * One scheduler is implemented for each hardware ring.
517 */
518struct drm_gpu_scheduler {
519 const struct drm_sched_backend_ops *ops;
520 u32 credit_limit;
521 atomic_t credit_count;
522 long timeout;
523 const char *name;
524 u32 num_rqs;
525 struct drm_sched_rq **sched_rq;
526 wait_queue_head_t job_scheduled;
527 atomic64_t job_id_count;
528 struct workqueue_struct *submit_wq;
529 struct workqueue_struct *timeout_wq;
530 struct work_struct work_run_job;
531 struct work_struct work_free_job;
532 struct delayed_work work_tdr;
533 struct list_head pending_list;
534 spinlock_t job_list_lock;
535 int hang_limit;
536 atomic_t *score;
537 atomic_t _score;
538 bool ready;
539 bool free_guilty;
540 bool pause_submit;
541 bool own_submit_wq;
542 struct device *dev;
543};
544
545int drm_sched_init(struct drm_gpu_scheduler *sched,
546 const struct drm_sched_backend_ops *ops,
547 struct workqueue_struct *submit_wq,
548 u32 num_rqs, u32 credit_limit, unsigned int hang_limit,
549 long timeout, struct workqueue_struct *timeout_wq,
550 atomic_t *score, const char *name, struct device *dev);
551
552void drm_sched_fini(struct drm_gpu_scheduler *sched);
553int drm_sched_job_init(struct drm_sched_job *job,
554 struct drm_sched_entity *entity,
555 u32 credits, void *owner);
556void drm_sched_job_arm(struct drm_sched_job *job);
557int drm_sched_job_add_dependency(struct drm_sched_job *job,
558 struct dma_fence *fence);
559int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
560 struct drm_file *file,
561 u32 handle,
562 u32 point);
563int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
564 struct dma_resv *resv,
565 enum dma_resv_usage usage);
566int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
567 struct drm_gem_object *obj,
568 bool write);
569
570
571void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
572 struct drm_gpu_scheduler **sched_list,
573 unsigned int num_sched_list);
574
575void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched);
576void drm_sched_job_cleanup(struct drm_sched_job *job);
577void drm_sched_wakeup(struct drm_gpu_scheduler *sched, struct drm_sched_entity *entity);
578bool drm_sched_wqueue_ready(struct drm_gpu_scheduler *sched);
579void drm_sched_wqueue_stop(struct drm_gpu_scheduler *sched);
580void drm_sched_wqueue_start(struct drm_gpu_scheduler *sched);
581void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
582void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
583void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
584void drm_sched_increase_karma(struct drm_sched_job *bad);
585void drm_sched_reset_karma(struct drm_sched_job *bad);
586void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
587bool drm_sched_dependency_optimized(struct dma_fence* fence,
588 struct drm_sched_entity *entity);
589void drm_sched_fault(struct drm_gpu_scheduler *sched);
590
591void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
592 struct drm_sched_entity *entity);
593void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
594 struct drm_sched_entity *entity);
595
596void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts);
597
598int drm_sched_entity_init(struct drm_sched_entity *entity,
599 enum drm_sched_priority priority,
600 struct drm_gpu_scheduler **sched_list,
601 unsigned int num_sched_list,
602 atomic_t *guilty);
603long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
604void drm_sched_entity_fini(struct drm_sched_entity *entity);
605void drm_sched_entity_destroy(struct drm_sched_entity *entity);
606void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
607struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
608void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
609void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
610 enum drm_sched_priority priority);
611bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
612int drm_sched_entity_error(struct drm_sched_entity *entity);
613
614struct drm_sched_fence *drm_sched_fence_alloc(
615 struct drm_sched_entity *s_entity, void *owner);
616void drm_sched_fence_init(struct drm_sched_fence *fence,
617 struct drm_sched_entity *entity);
618void drm_sched_fence_free(struct drm_sched_fence *fence);
619
620void drm_sched_fence_scheduled(struct drm_sched_fence *fence,
621 struct dma_fence *parent);
622void drm_sched_fence_finished(struct drm_sched_fence *fence, int result);
623
624unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
625void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
626 unsigned long remaining);
627struct drm_gpu_scheduler *
628drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
629 unsigned int num_sched_list);
630
631#endif