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         if (!(sched && entity && rq))
 113                 return -EINVAL;
 114
 115         memset(entity, 0, sizeof(struct amd_sched_entity));
 116         entity->belongto_rq = rq;
 117         entity->scheduler = sched;
 118         init_waitqueue_head(&entity->wait_queue);
 119         entity->fence_context = fence_context_alloc(1);
 120         if(kfifo_alloc(&entity->job_queue,
 121                        jobs * sizeof(void *),
 122                        GFP_KERNEL))
 123                 return -EINVAL;
 124
 125         spin_lock_init(&entity->queue_lock);
 126         atomic_set(&entity->fence_seq, 0);
 127
 128         /* Add the entity to the run queue */
 129         amd_sched_rq_add_entity(rq, entity);
 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 */
 141 static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
 142                                             struct amd_sched_entity *entity)
 143 {
 144         return entity->scheduler == sched &&
 145                 entity->belongto_rq != NULL;
 146 }
 147
 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)
 156 {
 157         rmb();
 158         if (kfifo_is_empty(&entity->job_queue))
 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  *
 170  * return 0 if succeed. negative error code on failure
 171  */
 172 int amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
 173                             struct amd_sched_entity *entity)
 174 {
 175         struct amd_sched_rq *rq = entity->belongto_rq;
 176         long r;
 177
 178         if (!amd_sched_entity_is_initialized(sched, entity))
 179                 return 0;
 180
 181         /**
 182          * The client will not queue more IBs during this fini, consume existing
 183          * queued IBs
 184         */
 185         r = wait_event_timeout(entity->wait_queue,
 186                 amd_sched_entity_is_idle(entity),
 187                 msecs_to_jiffies(AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS));
 188
 189         if (r <= 0)
 190                 DRM_INFO("Entity %p is in waiting state during fini\n",
 191                          entity);
 192
 193         amd_sched_rq_remove_entity(rq, entity);
 194         kfifo_free(&entity->job_queue);
 195         return r;
 196 }
 197
 198 /**
 199  * Helper to submit a job to the job queue
 200  *
 201  * @job         The pointer to job required to submit
 202  *
 203  * Returns true if we could submit the job.
 204  */
 205 static bool amd_sched_entity_in(struct amd_sched_job *job)
 206 {
 207         struct amd_sched_entity *entity = job->s_entity;
 208         bool added, first = false;
 209
 210         spin_lock(&entity->queue_lock);
 211         added = kfifo_in(&entity->job_queue, &job, sizeof(job)) == sizeof(job);
 212
 213         if (added && kfifo_len(&entity->job_queue) == sizeof(job))
 214                 first = true;
 215
 216         spin_unlock(&entity->queue_lock);
 217
 218         /* first job wakes up scheduler */
 219         if (first)
 220                 wake_up_interruptible(&job->sched->wait_queue);
 221
 222         return added;
 223 }
 224
 225 /**
 226  * Submit a job to the job queue
 227  *
 228  * @job         The pointer to job required to submit
 229  *
 230  * Returns 0 for success, negative error code otherwise.
 231  */
 232 int amd_sched_entity_push_job(struct amd_sched_job *sched_job)
 233 {
 234         struct amd_sched_entity *entity = sched_job->s_entity;
 235         struct amd_sched_fence *fence = amd_sched_fence_create(
 236                 entity, sched_job->owner);
 237         int r;
 238
 239         if (!fence)
 240                 return -ENOMEM;
 241
 242         fence_get(&fence->base);
 243         sched_job->s_fence = fence;
 244
 245         r = wait_event_interruptible(entity->wait_queue,
 246                                      amd_sched_entity_in(sched_job));
 247
 248         return r;
 249 }
 250
 251 /**
 252  * Return ture if we can push more jobs to the hw.
 253  */
 254 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
 255 {
 256         return atomic_read(&sched->hw_rq_count) <
 257                 sched->hw_submission_limit;
 258 }
 259
 260 /**
 261  * Select next entity containing real IB submissions
 262 */
 263 static struct amd_sched_entity *
 264 amd_sched_select_context(struct amd_gpu_scheduler *sched)
 265 {
 266         struct amd_sched_entity *tmp;
 267
 268         if (!amd_sched_ready(sched))
 269                 return NULL;
 270
 271         /* Kernel run queue has higher priority than normal run queue*/
 272         tmp = amd_sched_rq_select_entity(&sched->kernel_rq);
 273         if (tmp == NULL)
 274                 tmp = amd_sched_rq_select_entity(&sched->sched_rq);
 275
 276         return tmp;
 277 }
 278
 279 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 280 {
 281         struct amd_sched_job *sched_job =
 282                 container_of(cb, struct amd_sched_job, cb);
 283         struct amd_gpu_scheduler *sched;
 284
 285         sched = sched_job->sched;
 286         amd_sched_fence_signal(sched_job->s_fence);
 287         atomic_dec(&sched->hw_rq_count);
 288         fence_put(&sched_job->s_fence->base);
 289         sched->ops->process_job(sched, sched_job);
 290         wake_up_interruptible(&sched->wait_queue);
 291 }
 292
 293 static int amd_sched_main(void *param)
 294 {
 295         struct sched_param sparam = {.sched_priority = 1};
 296         struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
 297         int r;
 298
 299         sched_setscheduler(current, SCHED_FIFO, &sparam);
 300
 301         while (!kthread_should_stop()) {
 302                 struct amd_sched_entity *c_entity = NULL;
 303                 struct amd_sched_job *job;
 304                 struct fence *fence;
 305
 306                 wait_event_interruptible(sched->wait_queue,
 307                         kthread_should_stop() ||
 308                         (c_entity = amd_sched_select_context(sched)));
 309
 310                 if (!c_entity)
 311                         continue;
 312
 313                 r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *));
 314                 if (r != sizeof(void *))
 315                         continue;
 316                 atomic_inc(&sched->hw_rq_count);
 317
 318                 fence = sched->ops->run_job(sched, c_entity, job);
 319                 if (fence) {
 320                         r = fence_add_callback(fence, &job->cb,
 321                                                amd_sched_process_job);
 322                         if (r == -ENOENT)
 323                                 amd_sched_process_job(fence, &job->cb);
 324                         else if (r)
 325                                 DRM_ERROR("fence add callback failed (%d)\n", r);
 326                         fence_put(fence);
 327                 }
 328
 329                 wake_up(&c_entity->wait_queue);
 330         }
 331         return 0;
 332 }
 333
 334 /**
 335  * Create a gpu scheduler
 336  *
 337  * @ops                 The backend operations for this scheduler.
 338  * @ring                The the ring id for the scheduler.
 339  * @hw_submissions      Number of hw submissions to do.
 340  *
 341  * Return the pointer to scheduler for success, otherwise return NULL
 342 */
 343 struct amd_gpu_scheduler *amd_sched_create(struct amd_sched_backend_ops *ops,
 344                                            unsigned ring, unsigned hw_submission)
 345 {
 346         struct amd_gpu_scheduler *sched;
 347
 348         sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL);
 349         if (!sched)
 350                 return NULL;
 351
 352         sched->ops = ops;
 353         sched->ring_id = ring;
 354         sched->hw_submission_limit = hw_submission;
 355         snprintf(sched->name, sizeof(sched->name), "amdgpu[%d]", ring);
 356         amd_sched_rq_init(&sched->sched_rq);
 357         amd_sched_rq_init(&sched->kernel_rq);
 358
 359         init_waitqueue_head(&sched->wait_queue);
 360         atomic_set(&sched->hw_rq_count, 0);
 361         /* Each scheduler will run on a seperate kernel thread */
 362         sched->thread = kthread_run(amd_sched_main, sched, sched->name);
 363         if (IS_ERR(sched->thread)) {
 364                 DRM_ERROR("Failed to create scheduler for id %d.\n", ring);
 365                 kfree(sched);
 366                 return NULL;
 367         }
 368
 369         return sched;
 370 }
 371
 372 /**
 373  * Destroy a gpu scheduler
 374  *
 375  * @sched       The pointer to the scheduler
 376  *
 377  * return 0 if succeed. -1 if failed.
 378  */
 379 int amd_sched_destroy(struct amd_gpu_scheduler *sched)
 380 {
 381         kthread_stop(sched->thread);
 382         kfree(sched);
 383         return  0;
 384 }