drivers/gpu/drm/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 <uapi/linux/sched/types.h>
  28 #include <drm/drmP.h>
  29 #include <drm/gpu_scheduler.h>
  30 #include <drm/spsc_queue.h>
  31
  32 #define CREATE_TRACE_POINTS
  33 #include "gpu_scheduler_trace.h"
  34
  35 #define to_drm_sched_job(sched_job)             \
  36                 container_of((sched_job), struct drm_sched_job, queue_node)
  37
  38 static bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
  39 static void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
  40 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
  41
  42 /* Initialize a given run queue struct */
  43 static void drm_sched_rq_init(struct drm_sched_rq *rq)
  44 {
  45         spin_lock_init(&rq->lock);
  46         INIT_LIST_HEAD(&rq->entities);
  47         rq->current_entity = NULL;
  48 }
  49
  50 static void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
  51                                     struct drm_sched_entity *entity)
  52 {
  53         if (!list_empty(&entity->list))
  54                 return;
  55         spin_lock(&rq->lock);
  56         list_add_tail(&entity->list, &rq->entities);
  57         spin_unlock(&rq->lock);
  58 }
  59
  60 static void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
  61                                        struct drm_sched_entity *entity)
  62 {
  63         if (list_empty(&entity->list))
  64                 return;
  65         spin_lock(&rq->lock);
  66         list_del_init(&entity->list);
  67         if (rq->current_entity == entity)
  68                 rq->current_entity = NULL;
  69         spin_unlock(&rq->lock);
  70 }
  71
  72 /**
  73  * Select an entity which could provide a job to run
  74  *
  75  * @rq          The run queue to check.
  76  *
  77  * Try to find a ready entity, returns NULL if none found.
  78  */
  79 static struct drm_sched_entity *
  80 drm_sched_rq_select_entity(struct drm_sched_rq *rq)
  81 {
  82         struct drm_sched_entity *entity;
  83
  84         spin_lock(&rq->lock);
  85
  86         entity = rq->current_entity;
  87         if (entity) {
  88                 list_for_each_entry_continue(entity, &rq->entities, list) {
  89                         if (drm_sched_entity_is_ready(entity)) {
  90                                 rq->current_entity = entity;
  91                                 spin_unlock(&rq->lock);
  92                                 return entity;
  93                         }
  94                 }
  95         }
  96
  97         list_for_each_entry(entity, &rq->entities, list) {
  98
  99                 if (drm_sched_entity_is_ready(entity)) {
 100                         rq->current_entity = entity;
 101                         spin_unlock(&rq->lock);
 102                         return entity;
 103                 }
 104
 105                 if (entity == rq->current_entity)
 106                         break;
 107         }
 108
 109         spin_unlock(&rq->lock);
 110
 111         return NULL;
 112 }
 113
 114 /**
 115  * Init a context entity used by scheduler when submit to HW ring.
 116  *
 117  * @sched       The pointer to the scheduler
 118  * @entity      The pointer to a valid drm_sched_entity
 119  * @rq          The run queue this entity belongs
 120  * @guilty      atomic_t set to 1 when a job on this queue
 121  *              is found to be guilty causing a timeout
 122  *
 123  * return 0 if succeed. negative error code on failure
 124 */
 125 int drm_sched_entity_init(struct drm_gpu_scheduler *sched,
 126                           struct drm_sched_entity *entity,
 127                           struct drm_sched_rq *rq,
 128                           atomic_t *guilty)
 129 {
 130         if (!(sched && entity && rq))
 131                 return -EINVAL;
 132
 133         memset(entity, 0, sizeof(struct drm_sched_entity));
 134         INIT_LIST_HEAD(&entity->list);
 135         entity->rq = rq;
 136         entity->sched = sched;
 137         entity->guilty = guilty;
 138         entity->fini_status = 0;
 139         entity->last_scheduled = NULL;
 140
 141         spin_lock_init(&entity->rq_lock);
 142         spsc_queue_init(&entity->job_queue);
 143
 144         atomic_set(&entity->fence_seq, 0);
 145         entity->fence_context = dma_fence_context_alloc(2);
 146
 147         return 0;
 148 }
 149 EXPORT_SYMBOL(drm_sched_entity_init);
 150
 151 /**
 152  * Query if entity is initialized
 153  *
 154  * @sched       Pointer to scheduler instance
 155  * @entity      The pointer to a valid scheduler entity
 156  *
 157  * return true if entity is initialized, false otherwise
 158 */
 159 static bool drm_sched_entity_is_initialized(struct drm_gpu_scheduler *sched,
 160                                             struct drm_sched_entity *entity)
 161 {
 162         return entity->sched == sched &&
 163                 entity->rq != NULL;
 164 }
 165
 166 /**
 167  * Check if entity is idle
 168  *
 169  * @entity      The pointer to a valid scheduler entity
 170  *
 171  * Return true if entity don't has any unscheduled jobs.
 172  */
 173 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
 174 {
 175         rmb();
 176         if (spsc_queue_peek(&entity->job_queue) == NULL)
 177                 return true;
 178
 179         return false;
 180 }
 181
 182 /**
 183  * Check if entity is ready
 184  *
 185  * @entity      The pointer to a valid scheduler entity
 186  *
 187  * Return true if entity could provide a job.
 188  */
 189 static bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 190 {
 191         if (spsc_queue_peek(&entity->job_queue) == NULL)
 192                 return false;
 193
 194         if (READ_ONCE(entity->dependency))
 195                 return false;
 196
 197         return true;
 198 }
 199
 200 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 201                                     struct dma_fence_cb *cb)
 202 {
 203         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 204                                                  finish_cb);
 205         drm_sched_fence_finished(job->s_fence);
 206         WARN_ON(job->s_fence->parent);
 207         dma_fence_put(&job->s_fence->finished);
 208         job->sched->ops->free_job(job);
 209 }
 210
 211
 212 /**
 213  * Destroy a context entity
 214  *
 215  * @sched       Pointer to scheduler instance
 216  * @entity      The pointer to a valid scheduler entity
 217  *
 218  * Splitting drm_sched_entity_fini() into two functions, The first one is does the waiting,
 219  * removes the entity from the runqueue and returns an error when the process was killed.
 220  */
 221 void drm_sched_entity_do_release(struct drm_gpu_scheduler *sched,
 222                            struct drm_sched_entity *entity)
 223 {
 224         if (!drm_sched_entity_is_initialized(sched, entity))
 225                 return;
 226         /**
 227          * The client will not queue more IBs during this fini, consume existing
 228          * queued IBs or discard them on SIGKILL
 229         */
 230         if ((current->flags & PF_SIGNALED) && current->exit_code == SIGKILL)
 231                 entity->fini_status = -ERESTARTSYS;
 232         else
 233                 entity->fini_status = wait_event_killable(sched->job_scheduled,
 234                                         drm_sched_entity_is_idle(entity));
 235         drm_sched_entity_set_rq(entity, NULL);
 236 }
 237 EXPORT_SYMBOL(drm_sched_entity_do_release);
 238
 239 /**
 240  * Destroy a context entity
 241  *
 242  * @sched       Pointer to scheduler instance
 243  * @entity      The pointer to a valid scheduler entity
 244  *
 245  * The second one then goes over the entity and signals all jobs with an error code.
 246  */
 247 void drm_sched_entity_cleanup(struct drm_gpu_scheduler *sched,
 248                            struct drm_sched_entity *entity)
 249 {
 250         if (entity->fini_status) {
 251                 struct drm_sched_job *job;
 252                 int r;
 253
 254                 /* Park the kernel for a moment to make sure it isn't processing
 255                  * our enity.
 256                  */
 257                 kthread_park(sched->thread);
 258                 kthread_unpark(sched->thread);
 259                 if (entity->dependency) {
 260                         dma_fence_remove_callback(entity->dependency,
 261                                                   &entity->cb);
 262                         dma_fence_put(entity->dependency);
 263                         entity->dependency = NULL;
 264                 }
 265
 266                 while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 267                         struct drm_sched_fence *s_fence = job->s_fence;
 268                         drm_sched_fence_scheduled(s_fence);
 269                         dma_fence_set_error(&s_fence->finished, -ESRCH);
 270                         r = dma_fence_add_callback(entity->last_scheduled, &job->finish_cb,
 271                                                         drm_sched_entity_kill_jobs_cb);
 272                         if (r == -ENOENT)
 273                                 drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 274                         else if (r)
 275                                 DRM_ERROR("fence add callback failed (%d)\n", r);
 276                 }
 277         }
 278
 279         dma_fence_put(entity->last_scheduled);
 280         entity->last_scheduled = NULL;
 281 }
 282 EXPORT_SYMBOL(drm_sched_entity_cleanup);
 283
 284 void drm_sched_entity_fini(struct drm_gpu_scheduler *sched,
 285                                 struct drm_sched_entity *entity)
 286 {
 287         drm_sched_entity_do_release(sched, entity);
 288         drm_sched_entity_cleanup(sched, entity);
 289 }
 290 EXPORT_SYMBOL(drm_sched_entity_fini);
 291
 292 static void drm_sched_entity_wakeup(struct dma_fence *f, struct dma_fence_cb *cb)
 293 {
 294         struct drm_sched_entity *entity =
 295                 container_of(cb, struct drm_sched_entity, cb);
 296         entity->dependency = NULL;
 297         dma_fence_put(f);
 298         drm_sched_wakeup(entity->sched);
 299 }
 300
 301 static void drm_sched_entity_clear_dep(struct dma_fence *f, struct dma_fence_cb *cb)
 302 {
 303         struct drm_sched_entity *entity =
 304                 container_of(cb, struct drm_sched_entity, cb);
 305         entity->dependency = NULL;
 306         dma_fence_put(f);
 307 }
 308
 309 void drm_sched_entity_set_rq(struct drm_sched_entity *entity,
 310                              struct drm_sched_rq *rq)
 311 {
 312         if (entity->rq == rq)
 313                 return;
 314
 315         spin_lock(&entity->rq_lock);
 316
 317         if (entity->rq)
 318                 drm_sched_rq_remove_entity(entity->rq, entity);
 319
 320         entity->rq = rq;
 321         if (rq)
 322                 drm_sched_rq_add_entity(rq, entity);
 323
 324         spin_unlock(&entity->rq_lock);
 325 }
 326 EXPORT_SYMBOL(drm_sched_entity_set_rq);
 327
 328 bool drm_sched_dependency_optimized(struct dma_fence* fence,
 329                                     struct drm_sched_entity *entity)
 330 {
 331         struct drm_gpu_scheduler *sched = entity->sched;
 332         struct drm_sched_fence *s_fence;
 333
 334         if (!fence || dma_fence_is_signaled(fence))
 335                 return false;
 336         if (fence->context == entity->fence_context)
 337                 return true;
 338         s_fence = to_drm_sched_fence(fence);
 339         if (s_fence && s_fence->sched == sched)
 340                 return true;
 341
 342         return false;
 343 }
 344 EXPORT_SYMBOL(drm_sched_dependency_optimized);
 345
 346 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 347 {
 348         struct drm_gpu_scheduler *sched = entity->sched;
 349         struct dma_fence * fence = entity->dependency;
 350         struct drm_sched_fence *s_fence;
 351
 352         if (fence->context == entity->fence_context ||
 353             fence->context == entity->fence_context + 1) {
 354                 /*
 355                  * Fence is a scheduled/finished fence from a job
 356                  * which belongs to the same entity, we can ignore
 357                  * fences from ourself
 358                  */
 359                 dma_fence_put(entity->dependency);
 360                 return false;
 361         }
 362
 363         s_fence = to_drm_sched_fence(fence);
 364         if (s_fence && s_fence->sched == sched) {
 365
 366                 /*
 367                  * Fence is from the same scheduler, only need to wait for
 368                  * it to be scheduled
 369                  */
 370                 fence = dma_fence_get(&s_fence->scheduled);
 371                 dma_fence_put(entity->dependency);
 372                 entity->dependency = fence;
 373                 if (!dma_fence_add_callback(fence, &entity->cb,
 374                                             drm_sched_entity_clear_dep))
 375                         return true;
 376
 377                 /* Ignore it when it is already scheduled */
 378                 dma_fence_put(fence);
 379                 return false;
 380         }
 381
 382         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 383                                     drm_sched_entity_wakeup))
 384                 return true;
 385
 386         dma_fence_put(entity->dependency);
 387         return false;
 388 }
 389
 390 static struct drm_sched_job *
 391 drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 392 {
 393         struct drm_gpu_scheduler *sched = entity->sched;
 394         struct drm_sched_job *sched_job = to_drm_sched_job(
 395                                                 spsc_queue_peek(&entity->job_queue));
 396
 397         if (!sched_job)
 398                 return NULL;
 399
 400         while ((entity->dependency = sched->ops->dependency(sched_job, entity)))
 401                 if (drm_sched_entity_add_dependency_cb(entity))
 402                         return NULL;
 403
 404         /* skip jobs from entity that marked guilty */
 405         if (entity->guilty && atomic_read(entity->guilty))
 406                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 407
 408         dma_fence_put(entity->last_scheduled);
 409         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 410
 411         spsc_queue_pop(&entity->job_queue);
 412         return sched_job;
 413 }
 414
 415 /**
 416  * Submit a job to the job queue
 417  *
 418  * @sched_job           The pointer to job required to submit
 419  *
 420  * Note: To guarantee that the order of insertion to queue matches
 421  * the job's fence sequence number this function should be
 422  * called with drm_sched_job_init under common lock.
 423  *
 424  * Returns 0 for success, negative error code otherwise.
 425  */
 426 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 427                                struct drm_sched_entity *entity)
 428 {
 429         struct drm_gpu_scheduler *sched = sched_job->sched;
 430         bool first = false;
 431
 432         trace_drm_sched_job(sched_job, entity);
 433
 434         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 435
 436         /* first job wakes up scheduler */
 437         if (first) {
 438                 /* Add the entity to the run queue */
 439                 spin_lock(&entity->rq_lock);
 440                 drm_sched_rq_add_entity(entity->rq, entity);
 441                 spin_unlock(&entity->rq_lock);
 442                 drm_sched_wakeup(sched);
 443         }
 444 }
 445 EXPORT_SYMBOL(drm_sched_entity_push_job);
 446
 447 /* job_finish is called after hw fence signaled
 448  */
 449 static void drm_sched_job_finish(struct work_struct *work)
 450 {
 451         struct drm_sched_job *s_job = container_of(work, struct drm_sched_job,
 452                                                    finish_work);
 453         struct drm_gpu_scheduler *sched = s_job->sched;
 454
 455         /* remove job from ring_mirror_list */
 456         spin_lock(&sched->job_list_lock);
 457         list_del_init(&s_job->node);
 458         if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
 459                 struct drm_sched_job *next;
 460
 461                 spin_unlock(&sched->job_list_lock);
 462                 cancel_delayed_work_sync(&s_job->work_tdr);
 463                 spin_lock(&sched->job_list_lock);
 464
 465                 /* queue TDR for next job */
 466                 next = list_first_entry_or_null(&sched->ring_mirror_list,
 467                                                 struct drm_sched_job, node);
 468
 469                 if (next)
 470                         schedule_delayed_work(&next->work_tdr, sched->timeout);
 471         }
 472         spin_unlock(&sched->job_list_lock);
 473         dma_fence_put(&s_job->s_fence->finished);
 474         sched->ops->free_job(s_job);
 475 }
 476
 477 static void drm_sched_job_finish_cb(struct dma_fence *f,
 478                                     struct dma_fence_cb *cb)
 479 {
 480         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 481                                                  finish_cb);
 482         schedule_work(&job->finish_work);
 483 }
 484
 485 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 486 {
 487         struct drm_gpu_scheduler *sched = s_job->sched;
 488
 489         dma_fence_add_callback(&s_job->s_fence->finished, &s_job->finish_cb,
 490                                drm_sched_job_finish_cb);
 491
 492         spin_lock(&sched->job_list_lock);
 493         list_add_tail(&s_job->node, &sched->ring_mirror_list);
 494         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 495             list_first_entry_or_null(&sched->ring_mirror_list,
 496                                      struct drm_sched_job, node) == s_job)
 497                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
 498         spin_unlock(&sched->job_list_lock);
 499 }
 500
 501 static void drm_sched_job_timedout(struct work_struct *work)
 502 {
 503         struct drm_sched_job *job = container_of(work, struct drm_sched_job,
 504                                                  work_tdr.work);
 505
 506         job->sched->ops->timedout_job(job);
 507 }
 508
 509 void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 510 {
 511         struct drm_sched_job *s_job;
 512         struct drm_sched_entity *entity, *tmp;
 513         int i;
 514
 515         spin_lock(&sched->job_list_lock);
 516         list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
 517                 if (s_job->s_fence->parent &&
 518                     dma_fence_remove_callback(s_job->s_fence->parent,
 519                                               &s_job->s_fence->cb)) {
 520                         dma_fence_put(s_job->s_fence->parent);
 521                         s_job->s_fence->parent = NULL;
 522                         atomic_dec(&sched->hw_rq_count);
 523                 }
 524         }
 525         spin_unlock(&sched->job_list_lock);
 526
 527         if (bad && bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 528                 atomic_inc(&bad->karma);
 529                 /* don't increase @bad's karma if it's from KERNEL RQ,
 530                  * becuase sometimes GPU hang would cause kernel jobs (like VM updating jobs)
 531                  * corrupt but keep in mind that kernel jobs always considered good.
 532                  */
 533                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL; i++ ) {
 534                         struct drm_sched_rq *rq = &sched->sched_rq[i];
 535
 536                         spin_lock(&rq->lock);
 537                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
 538                                 if (bad->s_fence->scheduled.context == entity->fence_context) {
 539                                     if (atomic_read(&bad->karma) > bad->sched->hang_limit)
 540                                                 if (entity->guilty)
 541                                                         atomic_set(entity->guilty, 1);
 542                                         break;
 543                                 }
 544                         }
 545                         spin_unlock(&rq->lock);
 546                         if (&entity->list != &rq->entities)
 547                                 break;
 548                 }
 549         }
 550 }
 551 EXPORT_SYMBOL(drm_sched_hw_job_reset);
 552
 553 void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
 554 {
 555         struct drm_sched_job *s_job, *tmp;
 556         bool found_guilty = false;
 557         int r;
 558
 559         spin_lock(&sched->job_list_lock);
 560         s_job = list_first_entry_or_null(&sched->ring_mirror_list,
 561                                          struct drm_sched_job, node);
 562         if (s_job && sched->timeout != MAX_SCHEDULE_TIMEOUT)
 563                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
 564
 565         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 566                 struct drm_sched_fence *s_fence = s_job->s_fence;
 567                 struct dma_fence *fence;
 568                 uint64_t guilty_context;
 569
 570                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 571                         found_guilty = true;
 572                         guilty_context = s_job->s_fence->scheduled.context;
 573                 }
 574
 575                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 576                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 577
 578                 spin_unlock(&sched->job_list_lock);
 579                 fence = sched->ops->run_job(s_job);
 580                 atomic_inc(&sched->hw_rq_count);
 581
 582                 if (fence) {
 583                         s_fence->parent = dma_fence_get(fence);
 584                         r = dma_fence_add_callback(fence, &s_fence->cb,
 585                                                    drm_sched_process_job);
 586                         if (r == -ENOENT)
 587                                 drm_sched_process_job(fence, &s_fence->cb);
 588                         else if (r)
 589                                 DRM_ERROR("fence add callback failed (%d)\n",
 590                                           r);
 591                         dma_fence_put(fence);
 592                 } else {
 593                         drm_sched_process_job(NULL, &s_fence->cb);
 594                 }
 595                 spin_lock(&sched->job_list_lock);
 596         }
 597         spin_unlock(&sched->job_list_lock);
 598 }
 599 EXPORT_SYMBOL(drm_sched_job_recovery);
 600
 601 /**
 602  * Init a sched_job with basic field
 603  *
 604  * Note: Refer to drm_sched_entity_push_job documentation
 605  * for locking considerations.
 606  */
 607 int drm_sched_job_init(struct drm_sched_job *job,
 608                        struct drm_gpu_scheduler *sched,
 609                        struct drm_sched_entity *entity,
 610                        void *owner)
 611 {
 612         job->sched = sched;
 613         job->entity = entity;
 614         job->s_priority = entity->rq - sched->sched_rq;
 615         job->s_fence = drm_sched_fence_create(entity, owner);
 616         if (!job->s_fence)
 617                 return -ENOMEM;
 618         job->id = atomic64_inc_return(&sched->job_id_count);
 619
 620         INIT_WORK(&job->finish_work, drm_sched_job_finish);
 621         INIT_LIST_HEAD(&job->node);
 622         INIT_DELAYED_WORK(&job->work_tdr, drm_sched_job_timedout);
 623
 624         return 0;
 625 }
 626 EXPORT_SYMBOL(drm_sched_job_init);
 627
 628 /**
 629  * Return ture if we can push more jobs to the hw.
 630  */
 631 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 632 {
 633         return atomic_read(&sched->hw_rq_count) <
 634                 sched->hw_submission_limit;
 635 }
 636
 637 /**
 638  * Wake up the scheduler when it is ready
 639  */
 640 static void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 641 {
 642         if (drm_sched_ready(sched))
 643                 wake_up_interruptible(&sched->wake_up_worker);
 644 }
 645
 646 /**
 647  * Select next entity to process
 648 */
 649 static struct drm_sched_entity *
 650 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 651 {
 652         struct drm_sched_entity *entity;
 653         int i;
 654
 655         if (!drm_sched_ready(sched))
 656                 return NULL;
 657
 658         /* Kernel run queue has higher priority than normal run queue*/
 659         for (i = DRM_SCHED_PRIORITY_MAX - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 660                 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
 661                 if (entity)
 662                         break;
 663         }
 664
 665         return entity;
 666 }
 667
 668 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
 669 {
 670         struct drm_sched_fence *s_fence =
 671                 container_of(cb, struct drm_sched_fence, cb);
 672         struct drm_gpu_scheduler *sched = s_fence->sched;
 673
 674         dma_fence_get(&s_fence->finished);
 675         atomic_dec(&sched->hw_rq_count);
 676         drm_sched_fence_finished(s_fence);
 677
 678         trace_drm_sched_process_job(s_fence);
 679         dma_fence_put(&s_fence->finished);
 680         wake_up_interruptible(&sched->wake_up_worker);
 681 }
 682
 683 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 684 {
 685         if (kthread_should_park()) {
 686                 kthread_parkme();
 687                 return true;
 688         }
 689
 690         return false;
 691 }
 692
 693 static int drm_sched_main(void *param)
 694 {
 695         struct sched_param sparam = {.sched_priority = 1};
 696         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
 697         int r;
 698
 699         sched_setscheduler(current, SCHED_FIFO, &sparam);
 700
 701         while (!kthread_should_stop()) {
 702                 struct drm_sched_entity *entity = NULL;
 703                 struct drm_sched_fence *s_fence;
 704                 struct drm_sched_job *sched_job;
 705                 struct dma_fence *fence;
 706
 707                 wait_event_interruptible(sched->wake_up_worker,
 708                                          (!drm_sched_blocked(sched) &&
 709                                           (entity = drm_sched_select_entity(sched))) ||
 710                                          kthread_should_stop());
 711
 712                 if (!entity)
 713                         continue;
 714
 715                 sched_job = drm_sched_entity_pop_job(entity);
 716                 if (!sched_job)
 717                         continue;
 718
 719                 s_fence = sched_job->s_fence;
 720
 721                 atomic_inc(&sched->hw_rq_count);
 722                 drm_sched_job_begin(sched_job);
 723
 724                 fence = sched->ops->run_job(sched_job);
 725                 drm_sched_fence_scheduled(s_fence);
 726
 727                 if (fence) {
 728                         s_fence->parent = dma_fence_get(fence);
 729                         r = dma_fence_add_callback(fence, &s_fence->cb,
 730                                                    drm_sched_process_job);
 731                         if (r == -ENOENT)
 732                                 drm_sched_process_job(fence, &s_fence->cb);
 733                         else if (r)
 734                                 DRM_ERROR("fence add callback failed (%d)\n",
 735                                           r);
 736                         dma_fence_put(fence);
 737                 } else {
 738                         drm_sched_process_job(NULL, &s_fence->cb);
 739                 }
 740
 741                 wake_up(&sched->job_scheduled);
 742         }
 743         return 0;
 744 }
 745
 746 /**
 747  * Init a gpu scheduler instance
 748  *
 749  * @sched               The pointer to the scheduler
 750  * @ops                 The backend operations for this scheduler.
 751  * @hw_submissions      Number of hw submissions to do.
 752  * @name                Name used for debugging
 753  *
 754  * Return 0 on success, otherwise error code.
 755 */
 756 int drm_sched_init(struct drm_gpu_scheduler *sched,
 757                    const struct drm_sched_backend_ops *ops,
 758                    unsigned hw_submission,
 759                    unsigned hang_limit,
 760                    long timeout,
 761                    const char *name)
 762 {
 763         int i;
 764         sched->ops = ops;
 765         sched->hw_submission_limit = hw_submission;
 766         sched->name = name;
 767         sched->timeout = timeout;
 768         sched->hang_limit = hang_limit;
 769         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++)
 770                 drm_sched_rq_init(&sched->sched_rq[i]);
 771
 772         init_waitqueue_head(&sched->wake_up_worker);
 773         init_waitqueue_head(&sched->job_scheduled);
 774         INIT_LIST_HEAD(&sched->ring_mirror_list);
 775         spin_lock_init(&sched->job_list_lock);
 776         atomic_set(&sched->hw_rq_count, 0);
 777         atomic64_set(&sched->job_id_count, 0);
 778
 779         /* Each scheduler will run on a seperate kernel thread */
 780         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
 781         if (IS_ERR(sched->thread)) {
 782                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 783                 return PTR_ERR(sched->thread);
 784         }
 785
 786         return 0;
 787 }
 788 EXPORT_SYMBOL(drm_sched_init);
 789
 790 /**
 791  * Destroy a gpu scheduler
 792  *
 793  * @sched       The pointer to the scheduler
 794  */
 795 void drm_sched_fini(struct drm_gpu_scheduler *sched)
 796 {
 797         if (sched->thread)
 798                 kthread_stop(sched->thread);
 799 }
 800 EXPORT_SYMBOL(drm_sched_fini);