drivers/gpu/drm/amd/scheduler/gpu_scheduler.c

   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 */
  31 static void amd_sched_rq_init(struct amd_sched_rq *rq)
  32 {
  33         spin_lock_init(&rq->lock);
  34         INIT_LIST_HEAD(&rq->entities);
  35         rq->current_entity = NULL;
  36 }
  37
  38 static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
  39                                     struct amd_sched_entity *entity)
  40 {
  41         spin_lock(&rq->lock);
  42         list_add_tail(&entity->list, &rq->entities);
  43         spin_unlock(&rq->lock);
  44 }
  45
  46 static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
  47                                        struct amd_sched_entity *entity)
  48 {
  49         spin_lock(&rq->lock);
  50         list_del_init(&entity->list);
  51         if (rq->current_entity == entity)
  52                 rq->current_entity = NULL;
  53         spin_unlock(&rq->lock);
  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  */
  61 static struct amd_sched_entity *
  62 amd_sched_rq_select_entity(struct amd_sched_rq *rq)
  63 {
  64         struct amd_sched_entity *entity;
  65
  66         spin_lock(&rq->lock);
  67
  68         entity = rq->current_entity;
  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;
  73                                 spin_unlock(&rq->lock);
  74                                 return rq->current_entity;
  75                         }
  76                 }
  77         }
  78
  79         list_for_each_entry(entity, &rq->entities, list) {
  80
  81                 if (!kfifo_is_empty(&entity->job_queue)) {
  82                         rq->current_entity = entity;
  83                         spin_unlock(&rq->lock);
  84                         return rq->current_entity;
  85                 }
  86
  87                 if (entity == rq->current_entity)
  88                         break;
  89         }
  90
  91         spin_unlock(&rq->lock);
  92
  93         return NULL;
  94 }
  95
  96 /**
  97  * Init a context entity used by scheduler when submit to HW ring.
  98  *
  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
 104  *
 105  * return 0 if succeed. negative error code on failure
 106 */
 107 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 108                           struct amd_sched_entity *entity,
 109                           struct amd_sched_rq *rq,
 110                           uint32_t jobs)
 111 {
 112         char name[20];
 113
 114         if (!(sched && entity && rq))
 115                 return -EINVAL;
 116
 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         if(kfifo_alloc(&entity->job_queue,
 125                        jobs * sizeof(void *),
 126                        GFP_KERNEL))
 127                 return -EINVAL;
 128
 129         spin_lock_init(&entity->queue_lock);
 130         atomic_set(&entity->fence_seq, 0);
 131
 132         /* Add the entity to the run queue */
 133         amd_sched_rq_add_entity(rq, entity);
 134         return 0;
 135 }
 136
 137 /**
 138  * Query if entity is initialized
 139  *
 140  * @sched       Pointer to scheduler instance
 141  * @entity      The pointer to a valid scheduler entity
 142  *
 143  * return true if entity is initialized, false otherwise
 144 */
 145 static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
 146                                             struct amd_sched_entity *entity)
 147 {
 148         return entity->scheduler == sched &&
 149                 entity->belongto_rq != NULL;
 150 }
 151
 152 /**
 153  * Check if entity is idle
 154  *
 155  * @entity      The pointer to a valid scheduler entity
 156  *
 157  * Return true if entity don't has any unscheduled jobs.
 158  */
 159 static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
 160 {
 161         rmb();
 162         if (kfifo_is_empty(&entity->job_queue))
 163                 return true;
 164
 165         return false;
 166 }
 167
 168 /**
 169  * Destroy a context entity
 170  *
 171  * @sched       Pointer to scheduler instance
 172  * @entity      The pointer to a valid scheduler entity
 173  *
 174  * return 0 if succeed. negative error code on failure
 175  */
 176 int amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
 177                             struct amd_sched_entity *entity)
 178 {
 179         struct amd_sched_rq *rq = entity->belongto_rq;
 180         long r;
 181
 182         if (!amd_sched_entity_is_initialized(sched, entity))
 183                 return 0;
 184
 185         /**
 186          * The client will not queue more IBs during this fini, consume existing
 187          * queued IBs
 188         */
 189         r = wait_event_timeout(entity->wait_queue,
 190                 amd_sched_entity_is_idle(entity),
 191                 msecs_to_jiffies(AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS));
 192
 193         if (r <= 0)
 194                 DRM_INFO("Entity %p is in waiting state during fini\n",
 195                          entity);
 196
 197         amd_sched_rq_remove_entity(rq, entity);
 198         kfifo_free(&entity->job_queue);
 199         return r;
 200 }
 201
 202 /**
 203  * Helper to submit a job to the job queue
 204  *
 205  * @job         The pointer to job required to submit
 206  *
 207  * Returns true if we could submit the job.
 208  */
 209 static bool amd_sched_entity_in(struct amd_sched_job *job)
 210 {
 211         struct amd_sched_entity *entity = job->s_entity;
 212         bool added, first = false;
 213
 214         spin_lock(&entity->queue_lock);
 215         added = kfifo_in(&entity->job_queue, &job, sizeof(job)) == sizeof(job);
 216
 217         if (added && kfifo_len(&entity->job_queue) == sizeof(job))
 218                 first = true;
 219
 220         spin_unlock(&entity->queue_lock);
 221
 222         /* first job wakes up scheduler */
 223         if (first)
 224                 wake_up_interruptible(&job->sched->wait_queue);
 225
 226         return added;
 227 }
 228
 229 /**
 230  * Submit a job to the job queue
 231  *
 232  * @job         The pointer to job required to submit
 233  *
 234  * Returns 0 for success, negative error code otherwise.
 235  */
 236 int amd_sched_entity_push_job(struct amd_sched_job *sched_job)
 237 {
 238         struct amd_sched_entity *entity = sched_job->s_entity;
 239         struct amd_sched_fence *fence = amd_sched_fence_create(entity);
 240         int r;
 241
 242         if (!fence)
 243                 return -ENOMEM;
 244
 245         fence_get(&fence->base);
 246         sched_job->s_fence = fence;
 247
 248         r = wait_event_interruptible(entity->wait_queue,
 249                                      amd_sched_entity_in(sched_job));
 250
 251         return r;
 252 }
 253
 254 /**
 255  * Return ture if we can push more jobs to the hw.
 256  */
 257 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
 258 {
 259         return atomic_read(&sched->hw_rq_count) <
 260                 sched->hw_submission_limit;
 261 }
 262
 263 /**
 264  * Select next entity containing real IB submissions
 265 */
 266 static struct amd_sched_entity *
 267 amd_sched_select_context(struct amd_gpu_scheduler *sched)
 268 {
 269         struct amd_sched_entity *tmp;
 270
 271         if (!amd_sched_ready(sched))
 272                 return NULL;
 273
 274         /* Kernel run queue has higher priority than normal run queue*/
 275         tmp = amd_sched_rq_select_entity(&sched->kernel_rq);
 276         if (tmp == NULL)
 277                 tmp = amd_sched_rq_select_entity(&sched->sched_rq);
 278
 279         return tmp;
 280 }
 281
 282 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 283 {
 284         struct amd_sched_job *sched_job =
 285                 container_of(cb, struct amd_sched_job, cb);
 286         struct amd_gpu_scheduler *sched;
 287
 288         sched = sched_job->sched;
 289         amd_sched_fence_signal(sched_job->s_fence);
 290         atomic_dec(&sched->hw_rq_count);
 291         fence_put(&sched_job->s_fence->base);
 292         sched->ops->process_job(sched, sched_job);
 293         wake_up_interruptible(&sched->wait_queue);
 294 }
 295
 296 static int amd_sched_main(void *param)
 297 {
 298         struct sched_param sparam = {.sched_priority = 1};
 299         struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
 300         int r;
 301
 302         sched_setscheduler(current, SCHED_FIFO, &sparam);
 303
 304         while (!kthread_should_stop()) {
 305                 struct amd_sched_entity *c_entity = NULL;
 306                 struct amd_sched_job *job;
 307                 struct fence *fence;
 308
 309                 wait_event_interruptible(sched->wait_queue,
 310                         kthread_should_stop() ||
 311                         (c_entity = amd_sched_select_context(sched)));
 312
 313                 if (!c_entity)
 314                         continue;
 315
 316                 r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *));
 317                 if (r != sizeof(void *))
 318                         continue;
 319                 atomic_inc(&sched->hw_rq_count);
 320
 321                 fence = sched->ops->run_job(sched, c_entity, job);
 322                 if (fence) {
 323                         r = fence_add_callback(fence, &job->cb,
 324                                                amd_sched_process_job);
 325                         if (r == -ENOENT)
 326                                 amd_sched_process_job(fence, &job->cb);
 327                         else if (r)
 328                                 DRM_ERROR("fence add callback failed (%d)\n", r);
 329                         fence_put(fence);
 330                 }
 331
 332                 wake_up(&c_entity->wait_queue);
 333         }
 334         return 0;
 335 }
 336
 337 /**
 338  * Create a gpu scheduler
 339  *
 340  * @ops                 The backend operations for this scheduler.
 341  * @ring                The the ring id for the scheduler.
 342  * @hw_submissions      Number of hw submissions to do.
 343  *
 344  * Return the pointer to scheduler for success, otherwise return NULL
 345 */
 346 struct amd_gpu_scheduler *amd_sched_create(struct amd_sched_backend_ops *ops,
 347                                            unsigned ring, unsigned hw_submission)
 348 {
 349         struct amd_gpu_scheduler *sched;
 350         char name[20];
 351
 352         sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL);
 353         if (!sched)
 354                 return NULL;
 355
 356         sched->ops = ops;
 357         sched->ring_id = ring;
 358         sched->hw_submission_limit = hw_submission;
 359         snprintf(name, sizeof(name), "gpu_sched[%d]", ring);
 360         amd_sched_rq_init(&sched->sched_rq);
 361         amd_sched_rq_init(&sched->kernel_rq);
 362
 363         init_waitqueue_head(&sched->wait_queue);
 364         atomic_set(&sched->hw_rq_count, 0);
 365         /* Each scheduler will run on a seperate kernel thread */
 366         sched->thread = kthread_run(amd_sched_main, sched, name);
 367         if (IS_ERR(sched->thread)) {
 368                 DRM_ERROR("Failed to create scheduler for id %d.\n", ring);
 369                 kfree(sched);
 370                 return NULL;
 371         }
 372
 373         return sched;
 374 }
 375
 376 /**
 377  * Destroy a gpu scheduler
 378  *
 379  * @sched       The pointer to the scheduler
 380  *
 381  * return 0 if succeed. -1 if failed.
 382  */
 383 int amd_sched_destroy(struct amd_gpu_scheduler *sched)
 384 {
 385         kthread_stop(sched->thread);
 386         kfree(sched);
 387         return  0;
 388 }