]>
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 | { |
f556cb0c | 112 | char name[20]; |
a72ce6f8 JZ |
113 | |
114 | if (!(sched && entity && rq)) | |
115 | return -EINVAL; | |
116 | ||
91404fb2 | 117 | memset(entity, 0, sizeof(struct amd_sched_entity)); |
91404fb2 | 118 | entity->belongto_rq = rq; |
a72ce6f8 JZ |
119 | entity->scheduler = sched; |
120 | init_waitqueue_head(&entity->wait_queue); | |
f556cb0c CZ |
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); | |
1c8f805a | 124 | entity->need_wakeup = false; |
a72ce6f8 | 125 | if(kfifo_alloc(&entity->job_queue, |
1333f723 | 126 | jobs * sizeof(void *), |
a72ce6f8 JZ |
127 | GFP_KERNEL)) |
128 | return -EINVAL; | |
129 | ||
130 | spin_lock_init(&entity->queue_lock); | |
ce882e6d | 131 | atomic_set(&entity->fence_seq, 0); |
a72ce6f8 JZ |
132 | |
133 | /* Add the entity to the run queue */ | |
432a4ff8 | 134 | amd_sched_rq_add_entity(rq, entity); |
a72ce6f8 JZ |
135 | return 0; |
136 | } | |
137 | ||
138 | /** | |
139 | * Query if entity is initialized | |
140 | * | |
141 | * @sched Pointer to scheduler instance | |
142 | * @entity The pointer to a valid scheduler entity | |
143 | * | |
144 | * return true if entity is initialized, false otherwise | |
145 | */ | |
d54fdb94 CK |
146 | static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, |
147 | struct amd_sched_entity *entity) | |
a72ce6f8 JZ |
148 | { |
149 | return entity->scheduler == sched && | |
91404fb2 | 150 | entity->belongto_rq != NULL; |
a72ce6f8 JZ |
151 | } |
152 | ||
aef4852e CK |
153 | /** |
154 | * Check if entity is idle | |
155 | * | |
156 | * @entity The pointer to a valid scheduler entity | |
157 | * | |
158 | * Return true if entity don't has any unscheduled jobs. | |
159 | */ | |
160 | static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) | |
a72ce6f8 | 161 | { |
aef4852e CK |
162 | rmb(); |
163 | if (kfifo_is_empty(&entity->job_queue)) | |
a72ce6f8 JZ |
164 | return true; |
165 | ||
166 | return false; | |
167 | } | |
168 | ||
169 | /** | |
170 | * Destroy a context entity | |
171 | * | |
172 | * @sched Pointer to scheduler instance | |
173 | * @entity The pointer to a valid scheduler entity | |
174 | * | |
175 | * return 0 if succeed. negative error code on failure | |
176 | */ | |
91404fb2 CK |
177 | int amd_sched_entity_fini(struct amd_gpu_scheduler *sched, |
178 | struct amd_sched_entity *entity) | |
a72ce6f8 | 179 | { |
432a4ff8 | 180 | struct amd_sched_rq *rq = entity->belongto_rq; |
aef4852e | 181 | long r; |
a72ce6f8 | 182 | |
d54fdb94 | 183 | if (!amd_sched_entity_is_initialized(sched, entity)) |
a72ce6f8 | 184 | return 0; |
1c8f805a | 185 | entity->need_wakeup = true; |
a72ce6f8 JZ |
186 | /** |
187 | * The client will not queue more IBs during this fini, consume existing | |
188 | * queued IBs | |
189 | */ | |
aef4852e CK |
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)); | |
a72ce6f8 | 193 | |
aef4852e | 194 | if (r <= 0) |
9788ec40 CK |
195 | DRM_INFO("Entity %p is in waiting state during fini\n", |
196 | entity); | |
a72ce6f8 | 197 | |
432a4ff8 | 198 | amd_sched_rq_remove_entity(rq, entity); |
a72ce6f8 JZ |
199 | kfifo_free(&entity->job_queue); |
200 | return r; | |
201 | } | |
202 | ||
203 | /** | |
204 | * Submit a normal job to the job queue | |
205 | * | |
206 | * @sched The pointer to the scheduler | |
91404fb2 | 207 | * @c_entity The pointer to amd_sched_entity |
a72ce6f8 | 208 | * @job The pointer to job required to submit |
80de5913 CZ |
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. | |
212 | * -1 other fail. | |
a72ce6f8 | 213 | */ |
bb977d37 | 214 | int amd_sched_push_job(struct amd_sched_job *sched_job) |
a72ce6f8 | 215 | { |
bb977d37 CZ |
216 | struct amd_sched_fence *fence = |
217 | amd_sched_fence_create(sched_job->s_entity); | |
f556cb0c CZ |
218 | if (!fence) |
219 | return -EINVAL; | |
bb977d37 CZ |
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) != | |
225 | sizeof(void *)) { | |
a72ce6f8 JZ |
226 | /** |
227 | * Current context used up all its IB slots | |
228 | * wait here, or need to check whether GPU is hung | |
229 | */ | |
230 | schedule(); | |
231 | } | |
1c8f805a | 232 | /* first job wake up scheduler */ |
bb977d37 CZ |
233 | if ((kfifo_len(&sched_job->s_entity->job_queue) / sizeof(void *)) == 1) |
234 | wake_up_interruptible(&sched_job->sched->wait_queue); | |
80de5913 | 235 | return 0; |
a72ce6f8 JZ |
236 | } |
237 | ||
e688b728 CK |
238 | /** |
239 | * Return ture if we can push more jobs to the hw. | |
240 | */ | |
241 | static bool amd_sched_ready(struct amd_gpu_scheduler *sched) | |
242 | { | |
243 | return atomic_read(&sched->hw_rq_count) < | |
244 | sched->hw_submission_limit; | |
245 | } | |
246 | ||
247 | /** | |
248 | * Select next entity containing real IB submissions | |
249 | */ | |
250 | static struct amd_sched_entity * | |
251 | amd_sched_select_context(struct amd_gpu_scheduler *sched) | |
252 | { | |
253 | struct amd_sched_entity *tmp; | |
254 | ||
255 | if (!amd_sched_ready(sched)) | |
256 | return NULL; | |
257 | ||
258 | /* Kernel run queue has higher priority than normal run queue*/ | |
259 | tmp = amd_sched_rq_select_entity(&sched->kernel_rq); | |
260 | if (tmp == NULL) | |
261 | tmp = amd_sched_rq_select_entity(&sched->sched_rq); | |
262 | ||
263 | return tmp; | |
264 | } | |
265 | ||
6f0e54a9 CK |
266 | static void amd_sched_process_job(struct fence *f, struct fence_cb *cb) |
267 | { | |
268 | struct amd_sched_job *sched_job = | |
269 | container_of(cb, struct amd_sched_job, cb); | |
270 | struct amd_gpu_scheduler *sched; | |
6f0e54a9 CK |
271 | |
272 | sched = sched_job->sched; | |
f556cb0c | 273 | amd_sched_fence_signal(sched_job->s_fence); |
c746ba22 | 274 | atomic_dec(&sched->hw_rq_count); |
f556cb0c | 275 | fence_put(&sched_job->s_fence->base); |
bb977d37 | 276 | sched->ops->process_job(sched, sched_job); |
6f0e54a9 CK |
277 | wake_up_interruptible(&sched->wait_queue); |
278 | } | |
279 | ||
a72ce6f8 JZ |
280 | static int amd_sched_main(void *param) |
281 | { | |
a72ce6f8 | 282 | struct sched_param sparam = {.sched_priority = 1}; |
a72ce6f8 | 283 | struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; |
f85a6dd9 | 284 | int r; |
a72ce6f8 JZ |
285 | |
286 | sched_setscheduler(current, SCHED_FIFO, &sparam); | |
287 | ||
288 | while (!kthread_should_stop()) { | |
f85a6dd9 CK |
289 | struct amd_sched_entity *c_entity = NULL; |
290 | struct amd_sched_job *job; | |
6f0e54a9 CK |
291 | struct fence *fence; |
292 | ||
a72ce6f8 | 293 | wait_event_interruptible(sched->wait_queue, |
f85a6dd9 CK |
294 | kthread_should_stop() || |
295 | (c_entity = amd_sched_select_context(sched))); | |
296 | ||
297 | if (!c_entity) | |
298 | continue; | |
299 | ||
a72ce6f8 JZ |
300 | r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *)); |
301 | if (r != sizeof(void *)) | |
302 | continue; | |
b034b572 CK |
303 | atomic_inc(&sched->hw_rq_count); |
304 | ||
953e8fd4 | 305 | fence = sched->ops->run_job(sched, c_entity, job); |
6f0e54a9 | 306 | if (fence) { |
953e8fd4 | 307 | r = fence_add_callback(fence, &job->cb, |
6f0e54a9 CK |
308 | amd_sched_process_job); |
309 | if (r == -ENOENT) | |
953e8fd4 | 310 | amd_sched_process_job(fence, &job->cb); |
6f0e54a9 CK |
311 | else if (r) |
312 | DRM_ERROR("fence add callback failed (%d)\n", r); | |
313 | fence_put(fence); | |
314 | } | |
aef4852e CK |
315 | |
316 | if (c_entity->need_wakeup) { | |
317 | c_entity->need_wakeup = false; | |
318 | wake_up(&c_entity->wait_queue); | |
319 | } | |
320 | ||
a72ce6f8 JZ |
321 | } |
322 | return 0; | |
323 | } | |
324 | ||
a72ce6f8 JZ |
325 | /** |
326 | * Create a gpu scheduler | |
327 | * | |
69f7dd65 CK |
328 | * @ops The backend operations for this scheduler. |
329 | * @ring The the ring id for the scheduler. | |
330 | * @hw_submissions Number of hw submissions to do. | |
a72ce6f8 | 331 | * |
69f7dd65 | 332 | * Return the pointer to scheduler for success, otherwise return NULL |
a72ce6f8 | 333 | */ |
69f7dd65 CK |
334 | struct amd_gpu_scheduler *amd_sched_create(struct amd_sched_backend_ops *ops, |
335 | unsigned ring, unsigned hw_submission) | |
a72ce6f8 JZ |
336 | { |
337 | struct amd_gpu_scheduler *sched; | |
4cd7f42c | 338 | char name[20]; |
a72ce6f8 JZ |
339 | |
340 | sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL); | |
341 | if (!sched) | |
342 | return NULL; | |
343 | ||
a72ce6f8 | 344 | sched->ops = ops; |
a72ce6f8 | 345 | sched->ring_id = ring; |
4cef9267 | 346 | sched->hw_submission_limit = hw_submission; |
a72ce6f8 | 347 | snprintf(name, sizeof(name), "gpu_sched[%d]", ring); |
432a4ff8 CK |
348 | amd_sched_rq_init(&sched->sched_rq); |
349 | amd_sched_rq_init(&sched->kernel_rq); | |
a72ce6f8 JZ |
350 | |
351 | init_waitqueue_head(&sched->wait_queue); | |
c746ba22 | 352 | atomic_set(&sched->hw_rq_count, 0); |
a72ce6f8 | 353 | /* Each scheduler will run on a seperate kernel thread */ |
f4956598 CK |
354 | sched->thread = kthread_run(amd_sched_main, sched, name); |
355 | if (IS_ERR(sched->thread)) { | |
356 | DRM_ERROR("Failed to create scheduler for id %d.\n", ring); | |
357 | kfree(sched); | |
358 | return NULL; | |
a72ce6f8 JZ |
359 | } |
360 | ||
f4956598 | 361 | return sched; |
a72ce6f8 JZ |
362 | } |
363 | ||
364 | /** | |
365 | * Destroy a gpu scheduler | |
366 | * | |
367 | * @sched The pointer to the scheduler | |
368 | * | |
369 | * return 0 if succeed. -1 if failed. | |
370 | */ | |
371 | int amd_sched_destroy(struct amd_gpu_scheduler *sched) | |
372 | { | |
373 | kthread_stop(sched->thread); | |
a72ce6f8 JZ |
374 | kfree(sched); |
375 | return 0; | |
376 | } |