2 * Copyright 2015 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
24 #include <linux/kthread.h>
25 #include <linux/wait.h>
26 #include <linux/sched.h>
28 #include "gpu_scheduler.h"
30 /* Initialize a given run queue struct */
31 static void amd_sched_rq_init(struct amd_sched_rq *rq)
33 spin_lock_init(&rq->lock);
34 INIT_LIST_HEAD(&rq->entities);
35 rq->current_entity = NULL;
38 static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
39 struct amd_sched_entity *entity)
42 list_add_tail(&entity->list, &rq->entities);
43 spin_unlock(&rq->lock);
46 static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
47 struct amd_sched_entity *entity)
50 list_del_init(&entity->list);
51 if (rq->current_entity == entity)
52 rq->current_entity = NULL;
53 spin_unlock(&rq->lock);
57 * Select next entity from a specified run queue with round robin policy.
58 * It could return the same entity as current one if current is the only
59 * available one in the queue. Return NULL if nothing available.
61 static struct amd_sched_entity *
62 amd_sched_rq_select_entity(struct amd_sched_rq *rq)
64 struct amd_sched_entity *entity;
68 entity = rq->current_entity;
70 list_for_each_entry_continue(entity, &rq->entities, list) {
71 if (!kfifo_is_empty(&entity->job_queue)) {
72 rq->current_entity = entity;
73 spin_unlock(&rq->lock);
74 return rq->current_entity;
79 list_for_each_entry(entity, &rq->entities, list) {
81 if (!kfifo_is_empty(&entity->job_queue)) {
82 rq->current_entity = entity;
83 spin_unlock(&rq->lock);
84 return rq->current_entity;
87 if (entity == rq->current_entity)
91 spin_unlock(&rq->lock);
97 * Init a context entity used by scheduler when submit to HW ring.
99 * @sched The pointer to the scheduler
100 * @entity The pointer to a valid amd_sched_entity
101 * @rq The run queue this entity belongs
102 * @kernel If this is an entity for the kernel
103 * @jobs The max number of jobs in the job queue
105 * return 0 if succeed. negative error code on failure
107 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
108 struct amd_sched_entity *entity,
109 struct amd_sched_rq *rq,
114 if (!(sched && entity && rq))
117 memset(entity, 0, sizeof(struct amd_sched_entity));
118 entity->belongto_rq = rq;
119 entity->scheduler = sched;
120 init_waitqueue_head(&entity->wait_queue);
121 entity->fence_context = fence_context_alloc(1);
122 snprintf(name, sizeof(name), "c_entity[%llu]", entity->fence_context);
123 memcpy(entity->name, name, 20);
124 entity->need_wakeup = false;
125 if(kfifo_alloc(&entity->job_queue,
126 jobs * sizeof(void *),
130 spin_lock_init(&entity->queue_lock);
131 atomic_set(&entity->fence_seq, 0);
133 /* Add the entity to the run queue */
134 amd_sched_rq_add_entity(rq, entity);
139 * Query if entity is initialized
141 * @sched Pointer to scheduler instance
142 * @entity The pointer to a valid scheduler entity
144 * return true if entity is initialized, false otherwise
146 static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
147 struct amd_sched_entity *entity)
149 return entity->scheduler == sched &&
150 entity->belongto_rq != NULL;
154 * Check if entity is idle
156 * @entity The pointer to a valid scheduler entity
158 * Return true if entity don't has any unscheduled jobs.
160 static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
163 if (kfifo_is_empty(&entity->job_queue))
170 * Destroy a context entity
172 * @sched Pointer to scheduler instance
173 * @entity The pointer to a valid scheduler entity
175 * return 0 if succeed. negative error code on failure
177 int amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
178 struct amd_sched_entity *entity)
180 struct amd_sched_rq *rq = entity->belongto_rq;
183 if (!amd_sched_entity_is_initialized(sched, entity))
185 entity->need_wakeup = true;
187 * The client will not queue more IBs during this fini, consume existing
190 r = wait_event_timeout(entity->wait_queue,
191 amd_sched_entity_is_idle(entity),
192 msecs_to_jiffies(AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS));
195 DRM_INFO("Entity %p is in waiting state during fini\n",
198 amd_sched_rq_remove_entity(rq, entity);
199 kfifo_free(&entity->job_queue);
204 * Submit a normal job to the job queue
206 * @sched The pointer to the scheduler
207 * @c_entity The pointer to amd_sched_entity
208 * @job The pointer to job required to submit
209 * return 0 if succeed. -1 if failed.
210 * -2 indicate queue is full for this client, client should wait untill
211 * scheduler consum some queued command.
214 int amd_sched_push_job(struct amd_sched_job *sched_job)
216 struct amd_sched_fence *fence =
217 amd_sched_fence_create(sched_job->s_entity);
220 fence_get(&fence->base);
221 sched_job->s_fence = fence;
222 while (kfifo_in_spinlocked(&sched_job->s_entity->job_queue,
223 &sched_job, sizeof(void *),
224 &sched_job->s_entity->queue_lock) !=
227 * Current context used up all its IB slots
228 * wait here, or need to check whether GPU is hung
232 /* first job wake up scheduler */
233 if ((kfifo_len(&sched_job->s_entity->job_queue) / sizeof(void *)) == 1)
234 wake_up_interruptible(&sched_job->sched->wait_queue);
239 * Return ture if we can push more jobs to the hw.
241 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
243 return atomic_read(&sched->hw_rq_count) <
244 sched->hw_submission_limit;
248 * Select next entity containing real IB submissions
250 static struct amd_sched_entity *
251 amd_sched_select_context(struct amd_gpu_scheduler *sched)
253 struct amd_sched_entity *tmp;
255 if (!amd_sched_ready(sched))
258 /* Kernel run queue has higher priority than normal run queue*/
259 tmp = amd_sched_rq_select_entity(&sched->kernel_rq);
261 tmp = amd_sched_rq_select_entity(&sched->sched_rq);
266 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
268 struct amd_sched_job *sched_job =
269 container_of(cb, struct amd_sched_job, cb);
270 struct amd_gpu_scheduler *sched;
272 sched = sched_job->sched;
273 amd_sched_fence_signal(sched_job->s_fence);
274 atomic_dec(&sched->hw_rq_count);
275 fence_put(&sched_job->s_fence->base);
276 sched->ops->process_job(sched, sched_job);
277 wake_up_interruptible(&sched->wait_queue);
280 static int amd_sched_main(void *param)
282 struct sched_param sparam = {.sched_priority = 1};
283 struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
286 sched_setscheduler(current, SCHED_FIFO, &sparam);
288 while (!kthread_should_stop()) {
289 struct amd_sched_entity *c_entity = NULL;
290 struct amd_sched_job *job;
293 wait_event_interruptible(sched->wait_queue,
294 kthread_should_stop() ||
295 (c_entity = amd_sched_select_context(sched)));
300 r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *));
301 if (r != sizeof(void *))
304 if (sched->ops->prepare_job)
305 r = sched->ops->prepare_job(sched, c_entity, job);
307 atomic_inc(&sched->hw_rq_count);
309 mutex_lock(&sched->sched_lock);
310 fence = sched->ops->run_job(sched, c_entity, job);
312 r = fence_add_callback(fence, &job->cb,
313 amd_sched_process_job);
315 amd_sched_process_job(fence, &job->cb);
317 DRM_ERROR("fence add callback failed (%d)\n", r);
320 mutex_unlock(&sched->sched_lock);
322 if (c_entity->need_wakeup) {
323 c_entity->need_wakeup = false;
324 wake_up(&c_entity->wait_queue);
332 * Create a gpu scheduler
334 * @device The device context for this scheduler
335 * @ops The backend operations for this scheduler.
336 * @id The scheduler is per ring, here is ring id.
337 * @granularity The minumum ms unit the scheduler will scheduled.
338 * @preemption Indicate whether this ring support preemption, 0 is no.
340 * return the pointer to scheduler for success, otherwise return NULL
342 struct amd_gpu_scheduler *amd_sched_create(void *device,
343 struct amd_sched_backend_ops *ops,
345 unsigned granularity,
347 unsigned hw_submission)
349 struct amd_gpu_scheduler *sched;
352 sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL);
356 sched->device = device;
358 sched->granularity = granularity;
359 sched->ring_id = ring;
360 sched->preemption = preemption;
361 sched->hw_submission_limit = hw_submission;
362 snprintf(name, sizeof(name), "gpu_sched[%d]", ring);
363 mutex_init(&sched->sched_lock);
364 amd_sched_rq_init(&sched->sched_rq);
365 amd_sched_rq_init(&sched->kernel_rq);
367 init_waitqueue_head(&sched->wait_queue);
368 atomic_set(&sched->hw_rq_count, 0);
369 /* Each scheduler will run on a seperate kernel thread */
370 sched->thread = kthread_run(amd_sched_main, sched, name);
371 if (IS_ERR(sched->thread)) {
372 DRM_ERROR("Failed to create scheduler for id %d.\n", ring);
381 * Destroy a gpu scheduler
383 * @sched The pointer to the scheduler
385 * return 0 if succeed. -1 if failed.
387 int amd_sched_destroy(struct amd_gpu_scheduler *sched)
389 kthread_stop(sched->thread);