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                 /* We can ignore fences from ourself */
 354                 dma_fence_put(entity->dependency);
 355                 return false;
 356         }
 357
 358         s_fence = to_drm_sched_fence(fence);
 359         if (s_fence && s_fence->sched == sched) {
 360
 361                 /*
 362                  * Fence is from the same scheduler, only need to wait for
 363                  * it to be scheduled
 364                  */
 365                 fence = dma_fence_get(&s_fence->scheduled);
 366                 dma_fence_put(entity->dependency);
 367                 entity->dependency = fence;
 368                 if (!dma_fence_add_callback(fence, &entity->cb,
 369                                             drm_sched_entity_clear_dep))
 370                         return true;
 371
 372                 /* Ignore it when it is already scheduled */
 373                 dma_fence_put(fence);
 374                 return false;
 375         }
 376
 377         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 378                                     drm_sched_entity_wakeup))
 379                 return true;
 380
 381         dma_fence_put(entity->dependency);
 382         return false;
 383 }
 384
 385 static struct drm_sched_job *
 386 drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 387 {
 388         struct drm_gpu_scheduler *sched = entity->sched;
 389         struct drm_sched_job *sched_job = to_drm_sched_job(
 390                                                 spsc_queue_peek(&entity->job_queue));
 391
 392         if (!sched_job)
 393                 return NULL;
 394
 395         while ((entity->dependency = sched->ops->dependency(sched_job, entity)))
 396                 if (drm_sched_entity_add_dependency_cb(entity))
 397                         return NULL;
 398
 399         /* skip jobs from entity that marked guilty */
 400         if (entity->guilty && atomic_read(entity->guilty))
 401                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 402
 403         dma_fence_put(entity->last_scheduled);
 404         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 405
 406         spsc_queue_pop(&entity->job_queue);
 407         return sched_job;
 408 }
 409
 410 /**
 411  * Submit a job to the job queue
 412  *
 413  * @sched_job           The pointer to job required to submit
 414  *
 415  * Note: To guarantee that the order of insertion to queue matches
 416  * the job's fence sequence number this function should be
 417  * called with drm_sched_job_init under common lock.
 418  *
 419  * Returns 0 for success, negative error code otherwise.
 420  */
 421 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 422                                struct drm_sched_entity *entity)
 423 {
 424         struct drm_gpu_scheduler *sched = sched_job->sched;
 425         bool first = false;
 426
 427         trace_drm_sched_job(sched_job, entity);
 428
 429         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 430
 431         /* first job wakes up scheduler */
 432         if (first) {
 433                 /* Add the entity to the run queue */
 434                 spin_lock(&entity->rq_lock);
 435                 drm_sched_rq_add_entity(entity->rq, entity);
 436                 spin_unlock(&entity->rq_lock);
 437                 drm_sched_wakeup(sched);
 438         }
 439 }
 440 EXPORT_SYMBOL(drm_sched_entity_push_job);
 441
 442 /* job_finish is called after hw fence signaled
 443  */
 444 static void drm_sched_job_finish(struct work_struct *work)
 445 {
 446         struct drm_sched_job *s_job = container_of(work, struct drm_sched_job,
 447                                                    finish_work);
 448         struct drm_gpu_scheduler *sched = s_job->sched;
 449
 450         /* remove job from ring_mirror_list */
 451         spin_lock(&sched->job_list_lock);
 452         list_del_init(&s_job->node);
 453         if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
 454                 struct drm_sched_job *next;
 455
 456                 spin_unlock(&sched->job_list_lock);
 457                 cancel_delayed_work_sync(&s_job->work_tdr);
 458                 spin_lock(&sched->job_list_lock);
 459
 460                 /* queue TDR for next job */
 461                 next = list_first_entry_or_null(&sched->ring_mirror_list,
 462                                                 struct drm_sched_job, node);
 463
 464                 if (next)
 465                         schedule_delayed_work(&next->work_tdr, sched->timeout);
 466         }
 467         spin_unlock(&sched->job_list_lock);
 468         dma_fence_put(&s_job->s_fence->finished);
 469         sched->ops->free_job(s_job);
 470 }
 471
 472 static void drm_sched_job_finish_cb(struct dma_fence *f,
 473                                     struct dma_fence_cb *cb)
 474 {
 475         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 476                                                  finish_cb);
 477         schedule_work(&job->finish_work);
 478 }
 479
 480 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 481 {
 482         struct drm_gpu_scheduler *sched = s_job->sched;
 483
 484         dma_fence_add_callback(&s_job->s_fence->finished, &s_job->finish_cb,
 485                                drm_sched_job_finish_cb);
 486
 487         spin_lock(&sched->job_list_lock);
 488         list_add_tail(&s_job->node, &sched->ring_mirror_list);
 489         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 490             list_first_entry_or_null(&sched->ring_mirror_list,
 491                                      struct drm_sched_job, node) == s_job)
 492                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
 493         spin_unlock(&sched->job_list_lock);
 494 }
 495
 496 static void drm_sched_job_timedout(struct work_struct *work)
 497 {
 498         struct drm_sched_job *job = container_of(work, struct drm_sched_job,
 499                                                  work_tdr.work);
 500
 501         job->sched->ops->timedout_job(job);
 502 }
 503
 504 void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 505 {
 506         struct drm_sched_job *s_job;
 507         struct drm_sched_entity *entity, *tmp;
 508         int i;
 509
 510         spin_lock(&sched->job_list_lock);
 511         list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
 512                 if (s_job->s_fence->parent &&
 513                     dma_fence_remove_callback(s_job->s_fence->parent,
 514                                               &s_job->s_fence->cb)) {
 515                         dma_fence_put(s_job->s_fence->parent);
 516                         s_job->s_fence->parent = NULL;
 517                         atomic_dec(&sched->hw_rq_count);
 518                 }
 519         }
 520         spin_unlock(&sched->job_list_lock);
 521
 522         if (bad && bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 523                 atomic_inc(&bad->karma);
 524                 /* don't increase @bad's karma if it's from KERNEL RQ,
 525                  * becuase sometimes GPU hang would cause kernel jobs (like VM updating jobs)
 526                  * corrupt but keep in mind that kernel jobs always considered good.
 527                  */
 528                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL; i++ ) {
 529                         struct drm_sched_rq *rq = &sched->sched_rq[i];
 530
 531                         spin_lock(&rq->lock);
 532                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
 533                                 if (bad->s_fence->scheduled.context == entity->fence_context) {
 534                                     if (atomic_read(&bad->karma) > bad->sched->hang_limit)
 535                                                 if (entity->guilty)
 536                                                         atomic_set(entity->guilty, 1);
 537                                         break;
 538                                 }
 539                         }
 540                         spin_unlock(&rq->lock);
 541                         if (&entity->list != &rq->entities)
 542                                 break;
 543                 }
 544         }
 545 }
 546 EXPORT_SYMBOL(drm_sched_hw_job_reset);
 547
 548 void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
 549 {
 550         struct drm_sched_job *s_job, *tmp;
 551         bool found_guilty = false;
 552         int r;
 553
 554         spin_lock(&sched->job_list_lock);
 555         s_job = list_first_entry_or_null(&sched->ring_mirror_list,
 556                                          struct drm_sched_job, node);
 557         if (s_job && sched->timeout != MAX_SCHEDULE_TIMEOUT)
 558                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
 559
 560         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 561                 struct drm_sched_fence *s_fence = s_job->s_fence;
 562                 struct dma_fence *fence;
 563                 uint64_t guilty_context;
 564
 565                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 566                         found_guilty = true;
 567                         guilty_context = s_job->s_fence->scheduled.context;
 568                 }
 569
 570                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 571                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 572
 573                 spin_unlock(&sched->job_list_lock);
 574                 fence = sched->ops->run_job(s_job);
 575                 atomic_inc(&sched->hw_rq_count);
 576
 577                 if (fence) {
 578                         s_fence->parent = dma_fence_get(fence);
 579                         r = dma_fence_add_callback(fence, &s_fence->cb,
 580                                                    drm_sched_process_job);
 581                         if (r == -ENOENT)
 582                                 drm_sched_process_job(fence, &s_fence->cb);
 583                         else if (r)
 584                                 DRM_ERROR("fence add callback failed (%d)\n",
 585                                           r);
 586                         dma_fence_put(fence);
 587                 } else {
 588                         drm_sched_process_job(NULL, &s_fence->cb);
 589                 }
 590                 spin_lock(&sched->job_list_lock);
 591         }
 592         spin_unlock(&sched->job_list_lock);
 593 }
 594 EXPORT_SYMBOL(drm_sched_job_recovery);
 595
 596 /**
 597  * Init a sched_job with basic field
 598  *
 599  * Note: Refer to drm_sched_entity_push_job documentation
 600  * for locking considerations.
 601  */
 602 int drm_sched_job_init(struct drm_sched_job *job,
 603                        struct drm_gpu_scheduler *sched,
 604                        struct drm_sched_entity *entity,
 605                        void *owner)
 606 {
 607         job->sched = sched;
 608         job->entity = entity;
 609         job->s_priority = entity->rq - sched->sched_rq;
 610         job->s_fence = drm_sched_fence_create(entity, owner);
 611         if (!job->s_fence)
 612                 return -ENOMEM;
 613         job->id = atomic64_inc_return(&sched->job_id_count);
 614
 615         INIT_WORK(&job->finish_work, drm_sched_job_finish);
 616         INIT_LIST_HEAD(&job->node);
 617         INIT_DELAYED_WORK(&job->work_tdr, drm_sched_job_timedout);
 618
 619         return 0;
 620 }
 621 EXPORT_SYMBOL(drm_sched_job_init);
 622
 623 /**
 624  * Return ture if we can push more jobs to the hw.
 625  */
 626 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 627 {
 628         return atomic_read(&sched->hw_rq_count) <
 629                 sched->hw_submission_limit;
 630 }
 631
 632 /**
 633  * Wake up the scheduler when it is ready
 634  */
 635 static void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 636 {
 637         if (drm_sched_ready(sched))
 638                 wake_up_interruptible(&sched->wake_up_worker);
 639 }
 640
 641 /**
 642  * Select next entity to process
 643 */
 644 static struct drm_sched_entity *
 645 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 646 {
 647         struct drm_sched_entity *entity;
 648         int i;
 649
 650         if (!drm_sched_ready(sched))
 651                 return NULL;
 652
 653         /* Kernel run queue has higher priority than normal run queue*/
 654         for (i = DRM_SCHED_PRIORITY_MAX - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 655                 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
 656                 if (entity)
 657                         break;
 658         }
 659
 660         return entity;
 661 }
 662
 663 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
 664 {
 665         struct drm_sched_fence *s_fence =
 666                 container_of(cb, struct drm_sched_fence, cb);
 667         struct drm_gpu_scheduler *sched = s_fence->sched;
 668
 669         dma_fence_get(&s_fence->finished);
 670         atomic_dec(&sched->hw_rq_count);
 671         drm_sched_fence_finished(s_fence);
 672
 673         trace_drm_sched_process_job(s_fence);
 674         dma_fence_put(&s_fence->finished);
 675         wake_up_interruptible(&sched->wake_up_worker);
 676 }
 677
 678 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 679 {
 680         if (kthread_should_park()) {
 681                 kthread_parkme();
 682                 return true;
 683         }
 684
 685         return false;
 686 }
 687
 688 static int drm_sched_main(void *param)
 689 {
 690         struct sched_param sparam = {.sched_priority = 1};
 691         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
 692         int r;
 693
 694         sched_setscheduler(current, SCHED_FIFO, &sparam);
 695
 696         while (!kthread_should_stop()) {
 697                 struct drm_sched_entity *entity = NULL;
 698                 struct drm_sched_fence *s_fence;
 699                 struct drm_sched_job *sched_job;
 700                 struct dma_fence *fence;
 701
 702                 wait_event_interruptible(sched->wake_up_worker,
 703                                          (!drm_sched_blocked(sched) &&
 704                                           (entity = drm_sched_select_entity(sched))) ||
 705                                          kthread_should_stop());
 706
 707                 if (!entity)
 708                         continue;
 709
 710                 sched_job = drm_sched_entity_pop_job(entity);
 711                 if (!sched_job)
 712                         continue;
 713
 714                 s_fence = sched_job->s_fence;
 715
 716                 atomic_inc(&sched->hw_rq_count);
 717                 drm_sched_job_begin(sched_job);
 718
 719                 fence = sched->ops->run_job(sched_job);
 720                 drm_sched_fence_scheduled(s_fence);
 721
 722                 if (fence) {
 723                         s_fence->parent = dma_fence_get(fence);
 724                         r = dma_fence_add_callback(fence, &s_fence->cb,
 725                                                    drm_sched_process_job);
 726                         if (r == -ENOENT)
 727                                 drm_sched_process_job(fence, &s_fence->cb);
 728                         else if (r)
 729                                 DRM_ERROR("fence add callback failed (%d)\n",
 730                                           r);
 731                         dma_fence_put(fence);
 732                 } else {
 733                         drm_sched_process_job(NULL, &s_fence->cb);
 734                 }
 735
 736                 wake_up(&sched->job_scheduled);
 737         }
 738         return 0;
 739 }
 740
 741 /**
 742  * Init a gpu scheduler instance
 743  *
 744  * @sched               The pointer to the scheduler
 745  * @ops                 The backend operations for this scheduler.
 746  * @hw_submissions      Number of hw submissions to do.
 747  * @name                Name used for debugging
 748  *
 749  * Return 0 on success, otherwise error code.
 750 */
 751 int drm_sched_init(struct drm_gpu_scheduler *sched,
 752                    const struct drm_sched_backend_ops *ops,
 753                    unsigned hw_submission,
 754                    unsigned hang_limit,
 755                    long timeout,
 756                    const char *name)
 757 {
 758         int i;
 759         sched->ops = ops;
 760         sched->hw_submission_limit = hw_submission;
 761         sched->name = name;
 762         sched->timeout = timeout;
 763         sched->hang_limit = hang_limit;
 764         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++)
 765                 drm_sched_rq_init(&sched->sched_rq[i]);
 766
 767         init_waitqueue_head(&sched->wake_up_worker);
 768         init_waitqueue_head(&sched->job_scheduled);
 769         INIT_LIST_HEAD(&sched->ring_mirror_list);
 770         spin_lock_init(&sched->job_list_lock);
 771         atomic_set(&sched->hw_rq_count, 0);
 772         atomic64_set(&sched->job_id_count, 0);
 773
 774         /* Each scheduler will run on a seperate kernel thread */
 775         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
 776         if (IS_ERR(sched->thread)) {
 777                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 778                 return PTR_ERR(sched->thread);
 779         }
 780
 781         return 0;
 782 }
 783 EXPORT_SYMBOL(drm_sched_init);
 784
 785 /**
 786  * Destroy a gpu scheduler
 787  *
 788  * @sched       The pointer to the scheduler
 789  */
 790 void drm_sched_fini(struct drm_gpu_scheduler *sched)
 791 {
 792         if (sched->thread)
 793                 kthread_stop(sched->thread);
 794 }
 795 EXPORT_SYMBOL(drm_sched_fini);