]>
Commit | Line | Data |
---|---|---|
a72ce6f8 JZ |
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 | #include <linux/kthread.h> | |
25 | #include <linux/wait.h> | |
26 | #include <linux/sched.h> | |
27 | #include <drm/drmP.h> | |
28 | #include "gpu_scheduler.h" | |
29 | ||
30 | /* Initialize a given run queue struct */ | |
432a4ff8 | 31 | static void amd_sched_rq_init(struct amd_sched_rq *rq) |
a72ce6f8 | 32 | { |
2b184d8d | 33 | spin_lock_init(&rq->lock); |
432a4ff8 | 34 | INIT_LIST_HEAD(&rq->entities); |
432a4ff8 | 35 | rq->current_entity = NULL; |
a72ce6f8 JZ |
36 | } |
37 | ||
432a4ff8 CK |
38 | static void amd_sched_rq_add_entity(struct amd_sched_rq *rq, |
39 | struct amd_sched_entity *entity) | |
a72ce6f8 | 40 | { |
2b184d8d | 41 | spin_lock(&rq->lock); |
432a4ff8 | 42 | list_add_tail(&entity->list, &rq->entities); |
2b184d8d | 43 | spin_unlock(&rq->lock); |
a72ce6f8 JZ |
44 | } |
45 | ||
432a4ff8 CK |
46 | static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq, |
47 | struct amd_sched_entity *entity) | |
a72ce6f8 | 48 | { |
2b184d8d | 49 | spin_lock(&rq->lock); |
432a4ff8 CK |
50 | list_del_init(&entity->list); |
51 | if (rq->current_entity == entity) | |
52 | rq->current_entity = NULL; | |
2b184d8d | 53 | spin_unlock(&rq->lock); |
a72ce6f8 JZ |
54 | } |
55 | ||
56 | /** | |
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. | |
60 | */ | |
432a4ff8 CK |
61 | static struct amd_sched_entity * |
62 | amd_sched_rq_select_entity(struct amd_sched_rq *rq) | |
a72ce6f8 | 63 | { |
2b184d8d | 64 | struct amd_sched_entity *entity; |
432a4ff8 | 65 | |
2b184d8d CK |
66 | spin_lock(&rq->lock); |
67 | ||
68 | entity = rq->current_entity; | |
432a4ff8 CK |
69 | if (entity) { |
70 | list_for_each_entry_continue(entity, &rq->entities, list) { | |
71 | if (!kfifo_is_empty(&entity->job_queue)) { | |
72 | rq->current_entity = entity; | |
2b184d8d | 73 | spin_unlock(&rq->lock); |
432a4ff8 CK |
74 | return rq->current_entity; |
75 | } | |
a72ce6f8 | 76 | } |
a72ce6f8 | 77 | } |
a72ce6f8 | 78 | |
432a4ff8 | 79 | list_for_each_entry(entity, &rq->entities, list) { |
a72ce6f8 | 80 | |
432a4ff8 CK |
81 | if (!kfifo_is_empty(&entity->job_queue)) { |
82 | rq->current_entity = entity; | |
2b184d8d | 83 | spin_unlock(&rq->lock); |
432a4ff8 CK |
84 | return rq->current_entity; |
85 | } | |
a72ce6f8 | 86 | |
432a4ff8 CK |
87 | if (entity == rq->current_entity) |
88 | break; | |
89 | } | |
a72ce6f8 | 90 | |
2b184d8d CK |
91 | spin_unlock(&rq->lock); |
92 | ||
432a4ff8 | 93 | return NULL; |
a72ce6f8 JZ |
94 | } |
95 | ||
a72ce6f8 JZ |
96 | /** |
97 | * Init a context entity used by scheduler when submit to HW ring. | |
98 | * | |
99 | * @sched The pointer to the scheduler | |
91404fb2 | 100 | * @entity The pointer to a valid amd_sched_entity |
a72ce6f8 | 101 | * @rq The run queue this entity belongs |
0e89d0c1 | 102 | * @kernel If this is an entity for the kernel |
1333f723 | 103 | * @jobs The max number of jobs in the job queue |
a72ce6f8 JZ |
104 | * |
105 | * return 0 if succeed. negative error code on failure | |
106 | */ | |
91404fb2 | 107 | int amd_sched_entity_init(struct amd_gpu_scheduler *sched, |
6f0e54a9 | 108 | struct amd_sched_entity *entity, |
432a4ff8 | 109 | struct amd_sched_rq *rq, |
6f0e54a9 | 110 | uint32_t jobs) |
a72ce6f8 | 111 | { |
a72ce6f8 JZ |
112 | if (!(sched && entity && rq)) |
113 | return -EINVAL; | |
114 | ||
91404fb2 | 115 | memset(entity, 0, sizeof(struct amd_sched_entity)); |
91404fb2 | 116 | entity->belongto_rq = rq; |
a72ce6f8 JZ |
117 | entity->scheduler = sched; |
118 | init_waitqueue_head(&entity->wait_queue); | |
f556cb0c | 119 | entity->fence_context = fence_context_alloc(1); |
a72ce6f8 | 120 | if(kfifo_alloc(&entity->job_queue, |
1333f723 | 121 | jobs * sizeof(void *), |
a72ce6f8 JZ |
122 | GFP_KERNEL)) |
123 | return -EINVAL; | |
124 | ||
125 | spin_lock_init(&entity->queue_lock); | |
ce882e6d | 126 | atomic_set(&entity->fence_seq, 0); |
a72ce6f8 JZ |
127 | |
128 | /* Add the entity to the run queue */ | |
432a4ff8 | 129 | amd_sched_rq_add_entity(rq, entity); |
a72ce6f8 JZ |
130 | return 0; |
131 | } | |
132 | ||
133 | /** | |
134 | * Query if entity is initialized | |
135 | * | |
136 | * @sched Pointer to scheduler instance | |
137 | * @entity The pointer to a valid scheduler entity | |
138 | * | |
139 | * return true if entity is initialized, false otherwise | |
140 | */ | |
d54fdb94 CK |
141 | static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, |
142 | struct amd_sched_entity *entity) | |
a72ce6f8 JZ |
143 | { |
144 | return entity->scheduler == sched && | |
91404fb2 | 145 | entity->belongto_rq != NULL; |
a72ce6f8 JZ |
146 | } |
147 | ||
aef4852e CK |
148 | /** |
149 | * Check if entity is idle | |
150 | * | |
151 | * @entity The pointer to a valid scheduler entity | |
152 | * | |
153 | * Return true if entity don't has any unscheduled jobs. | |
154 | */ | |
155 | static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) | |
a72ce6f8 | 156 | { |
aef4852e CK |
157 | rmb(); |
158 | if (kfifo_is_empty(&entity->job_queue)) | |
a72ce6f8 JZ |
159 | return true; |
160 | ||
161 | return false; | |
162 | } | |
163 | ||
164 | /** | |
165 | * Destroy a context entity | |
166 | * | |
167 | * @sched Pointer to scheduler instance | |
168 | * @entity The pointer to a valid scheduler entity | |
169 | * | |
062c7fb3 | 170 | * Cleanup and free the allocated resources. |
a72ce6f8 | 171 | */ |
062c7fb3 CK |
172 | void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, |
173 | struct amd_sched_entity *entity) | |
a72ce6f8 | 174 | { |
432a4ff8 | 175 | struct amd_sched_rq *rq = entity->belongto_rq; |
a72ce6f8 | 176 | |
d54fdb94 | 177 | if (!amd_sched_entity_is_initialized(sched, entity)) |
062c7fb3 | 178 | return; |
6c859274 | 179 | |
a72ce6f8 JZ |
180 | /** |
181 | * The client will not queue more IBs during this fini, consume existing | |
182 | * queued IBs | |
183 | */ | |
062c7fb3 | 184 | wait_event(entity->wait_queue, amd_sched_entity_is_idle(entity)); |
a72ce6f8 | 185 | |
432a4ff8 | 186 | amd_sched_rq_remove_entity(rq, entity); |
a72ce6f8 | 187 | kfifo_free(&entity->job_queue); |
a72ce6f8 JZ |
188 | } |
189 | ||
190 | /** | |
6c859274 | 191 | * Helper to submit a job to the job queue |
a72ce6f8 | 192 | * |
a72ce6f8 | 193 | * @job The pointer to job required to submit |
6c859274 CK |
194 | * |
195 | * Returns true if we could submit the job. | |
196 | */ | |
197 | static bool amd_sched_entity_in(struct amd_sched_job *job) | |
a72ce6f8 | 198 | { |
6c859274 CK |
199 | struct amd_sched_entity *entity = job->s_entity; |
200 | bool added, first = false; | |
201 | ||
202 | spin_lock(&entity->queue_lock); | |
203 | added = kfifo_in(&entity->job_queue, &job, sizeof(job)) == sizeof(job); | |
204 | ||
205 | if (added && kfifo_len(&entity->job_queue) == sizeof(job)) | |
206 | first = true; | |
207 | ||
208 | spin_unlock(&entity->queue_lock); | |
209 | ||
210 | /* first job wakes up scheduler */ | |
211 | if (first) | |
212 | wake_up_interruptible(&job->sched->wait_queue); | |
213 | ||
214 | return added; | |
215 | } | |
216 | ||
217 | /** | |
218 | * Submit a job to the job queue | |
219 | * | |
220 | * @job The pointer to job required to submit | |
221 | * | |
222 | * Returns 0 for success, negative error code otherwise. | |
223 | */ | |
224 | int amd_sched_entity_push_job(struct amd_sched_job *sched_job) | |
225 | { | |
226 | struct amd_sched_entity *entity = sched_job->s_entity; | |
84f76ea6 CZ |
227 | struct amd_sched_fence *fence = amd_sched_fence_create( |
228 | entity, sched_job->owner); | |
6c859274 CK |
229 | int r; |
230 | ||
f556cb0c | 231 | if (!fence) |
6c859274 CK |
232 | return -ENOMEM; |
233 | ||
bb977d37 CZ |
234 | fence_get(&fence->base); |
235 | sched_job->s_fence = fence; | |
6c859274 CK |
236 | |
237 | r = wait_event_interruptible(entity->wait_queue, | |
238 | amd_sched_entity_in(sched_job)); | |
239 | ||
240 | return r; | |
a72ce6f8 JZ |
241 | } |
242 | ||
e688b728 CK |
243 | /** |
244 | * Return ture if we can push more jobs to the hw. | |
245 | */ | |
246 | static bool amd_sched_ready(struct amd_gpu_scheduler *sched) | |
247 | { | |
248 | return atomic_read(&sched->hw_rq_count) < | |
249 | sched->hw_submission_limit; | |
250 | } | |
251 | ||
252 | /** | |
253 | * Select next entity containing real IB submissions | |
254 | */ | |
255 | static struct amd_sched_entity * | |
256 | amd_sched_select_context(struct amd_gpu_scheduler *sched) | |
257 | { | |
258 | struct amd_sched_entity *tmp; | |
259 | ||
260 | if (!amd_sched_ready(sched)) | |
261 | return NULL; | |
262 | ||
263 | /* Kernel run queue has higher priority than normal run queue*/ | |
264 | tmp = amd_sched_rq_select_entity(&sched->kernel_rq); | |
265 | if (tmp == NULL) | |
266 | tmp = amd_sched_rq_select_entity(&sched->sched_rq); | |
267 | ||
268 | return tmp; | |
269 | } | |
270 | ||
6f0e54a9 CK |
271 | static void amd_sched_process_job(struct fence *f, struct fence_cb *cb) |
272 | { | |
273 | struct amd_sched_job *sched_job = | |
274 | container_of(cb, struct amd_sched_job, cb); | |
275 | struct amd_gpu_scheduler *sched; | |
6f0e54a9 CK |
276 | |
277 | sched = sched_job->sched; | |
f556cb0c | 278 | amd_sched_fence_signal(sched_job->s_fence); |
c746ba22 | 279 | atomic_dec(&sched->hw_rq_count); |
f556cb0c | 280 | fence_put(&sched_job->s_fence->base); |
bb977d37 | 281 | sched->ops->process_job(sched, sched_job); |
6f0e54a9 CK |
282 | wake_up_interruptible(&sched->wait_queue); |
283 | } | |
284 | ||
a72ce6f8 JZ |
285 | static int amd_sched_main(void *param) |
286 | { | |
a72ce6f8 | 287 | struct sched_param sparam = {.sched_priority = 1}; |
a72ce6f8 | 288 | struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; |
f85a6dd9 | 289 | int r; |
a72ce6f8 JZ |
290 | |
291 | sched_setscheduler(current, SCHED_FIFO, &sparam); | |
292 | ||
293 | while (!kthread_should_stop()) { | |
f85a6dd9 CK |
294 | struct amd_sched_entity *c_entity = NULL; |
295 | struct amd_sched_job *job; | |
6f0e54a9 CK |
296 | struct fence *fence; |
297 | ||
a72ce6f8 | 298 | wait_event_interruptible(sched->wait_queue, |
f85a6dd9 CK |
299 | kthread_should_stop() || |
300 | (c_entity = amd_sched_select_context(sched))); | |
301 | ||
302 | if (!c_entity) | |
303 | continue; | |
304 | ||
a72ce6f8 JZ |
305 | r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *)); |
306 | if (r != sizeof(void *)) | |
307 | continue; | |
b034b572 CK |
308 | atomic_inc(&sched->hw_rq_count); |
309 | ||
953e8fd4 | 310 | fence = sched->ops->run_job(sched, c_entity, job); |
6f0e54a9 | 311 | if (fence) { |
953e8fd4 | 312 | r = fence_add_callback(fence, &job->cb, |
6f0e54a9 CK |
313 | amd_sched_process_job); |
314 | if (r == -ENOENT) | |
953e8fd4 | 315 | amd_sched_process_job(fence, &job->cb); |
6f0e54a9 CK |
316 | else if (r) |
317 | DRM_ERROR("fence add callback failed (%d)\n", r); | |
318 | fence_put(fence); | |
319 | } | |
aef4852e | 320 | |
6c859274 | 321 | wake_up(&c_entity->wait_queue); |
a72ce6f8 JZ |
322 | } |
323 | return 0; | |
324 | } | |
325 | ||
a72ce6f8 JZ |
326 | /** |
327 | * Create a gpu scheduler | |
328 | * | |
69f7dd65 CK |
329 | * @ops The backend operations for this scheduler. |
330 | * @ring The the ring id for the scheduler. | |
331 | * @hw_submissions Number of hw submissions to do. | |
a72ce6f8 | 332 | * |
69f7dd65 | 333 | * Return the pointer to scheduler for success, otherwise return NULL |
a72ce6f8 | 334 | */ |
69f7dd65 | 335 | struct amd_gpu_scheduler *amd_sched_create(struct amd_sched_backend_ops *ops, |
f38fdfdd CZ |
336 | unsigned ring, unsigned hw_submission, |
337 | void *priv) | |
a72ce6f8 JZ |
338 | { |
339 | struct amd_gpu_scheduler *sched; | |
a72ce6f8 JZ |
340 | |
341 | sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL); | |
342 | if (!sched) | |
343 | return NULL; | |
344 | ||
a72ce6f8 | 345 | sched->ops = ops; |
a72ce6f8 | 346 | sched->ring_id = ring; |
4cef9267 | 347 | sched->hw_submission_limit = hw_submission; |
f38fdfdd | 348 | sched->priv = priv; |
c14692f0 | 349 | snprintf(sched->name, sizeof(sched->name), "amdgpu[%d]", ring); |
432a4ff8 CK |
350 | amd_sched_rq_init(&sched->sched_rq); |
351 | amd_sched_rq_init(&sched->kernel_rq); | |
a72ce6f8 JZ |
352 | |
353 | init_waitqueue_head(&sched->wait_queue); | |
c746ba22 | 354 | atomic_set(&sched->hw_rq_count, 0); |
a72ce6f8 | 355 | /* Each scheduler will run on a seperate kernel thread */ |
c14692f0 | 356 | sched->thread = kthread_run(amd_sched_main, sched, sched->name); |
f4956598 CK |
357 | if (IS_ERR(sched->thread)) { |
358 | DRM_ERROR("Failed to create scheduler for id %d.\n", ring); | |
359 | kfree(sched); | |
360 | return NULL; | |
a72ce6f8 JZ |
361 | } |
362 | ||
f4956598 | 363 | return sched; |
a72ce6f8 JZ |
364 | } |
365 | ||
366 | /** | |
367 | * Destroy a gpu scheduler | |
368 | * | |
369 | * @sched The pointer to the scheduler | |
370 | * | |
371 | * return 0 if succeed. -1 if failed. | |
372 | */ | |
373 | int amd_sched_destroy(struct amd_gpu_scheduler *sched) | |
374 | { | |
375 | kthread_stop(sched->thread); | |
a72ce6f8 JZ |
376 | kfree(sched); |
377 | return 0; | |
378 | } |