drivers/gpu/drm/scheduler/sched_main.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 /**
  25  * DOC: Overview
  26  *
  27  * The GPU scheduler provides entities which allow userspace to push jobs
  28  * into software queues which are then scheduled on a hardware run queue.
  29  * The software queues have a priority among them. The scheduler selects the entities
  30  * from the run queue using a FIFO. The scheduler provides dependency handling
  31  * features among jobs. The driver is supposed to provide callback functions for
  32  * backend operations to the scheduler like submitting a job to hardware run queue,
  33  * returning the dependencies of a job etc.
  34  *
  35  * The organisation of the scheduler is the following:
  36  *
  37  * 1. Each hw run queue has one scheduler
  38  * 2. Each scheduler has multiple run queues with different priorities
  39  *    (e.g., HIGH_HW,HIGH_SW, KERNEL, NORMAL)
  40  * 3. Each scheduler run queue has a queue of entities to schedule
  41  * 4. Entities themselves maintain a queue of jobs that will be scheduled on
  42  *    the hardware.
  43  *
  44  * The jobs in a entity are always scheduled in the order that they were pushed.
  45  */
  46
  47 #include <linux/kthread.h>
  48 #include <linux/wait.h>
  49 #include <linux/sched.h>
  50 #include <uapi/linux/sched/types.h>
  51 #include <drm/drmP.h>
  52 #include <drm/gpu_scheduler.h>
  53 #include <drm/spsc_queue.h>
  54
  55 #define CREATE_TRACE_POINTS
  56 #include "gpu_scheduler_trace.h"
  57
  58 #define to_drm_sched_job(sched_job)             \
  59                 container_of((sched_job), struct drm_sched_job, queue_node)
  60
  61 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
  62
  63 /**
  64  * drm_sched_rq_init - initialize a given run queue struct
  65  *
  66  * @rq: scheduler run queue
  67  *
  68  * Initializes a scheduler runqueue.
  69  */
  70 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
  71                               struct drm_sched_rq *rq)
  72 {
  73         spin_lock_init(&rq->lock);
  74         INIT_LIST_HEAD(&rq->entities);
  75         rq->current_entity = NULL;
  76         rq->sched = sched;
  77 }
  78
  79 /**
  80  * drm_sched_rq_add_entity - add an entity
  81  *
  82  * @rq: scheduler run queue
  83  * @entity: scheduler entity
  84  *
  85  * Adds a scheduler entity to the run queue.
  86  */
  87 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
  88                              struct drm_sched_entity *entity)
  89 {
  90         if (!list_empty(&entity->list))
  91                 return;
  92         spin_lock(&rq->lock);
  93         list_add_tail(&entity->list, &rq->entities);
  94         spin_unlock(&rq->lock);
  95 }
  96
  97 /**
  98  * drm_sched_rq_remove_entity - remove an entity
  99  *
 100  * @rq: scheduler run queue
 101  * @entity: scheduler entity
 102  *
 103  * Removes a scheduler entity from the run queue.
 104  */
 105 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 106                                 struct drm_sched_entity *entity)
 107 {
 108         if (list_empty(&entity->list))
 109                 return;
 110         spin_lock(&rq->lock);
 111         list_del_init(&entity->list);
 112         if (rq->current_entity == entity)
 113                 rq->current_entity = NULL;
 114         spin_unlock(&rq->lock);
 115 }
 116
 117 /**
 118  * drm_sched_rq_select_entity - Select an entity which could provide a job to run
 119  *
 120  * @rq: scheduler run queue to check.
 121  *
 122  * Try to find a ready entity, returns NULL if none found.
 123  */
 124 static struct drm_sched_entity *
 125 drm_sched_rq_select_entity(struct drm_sched_rq *rq)
 126 {
 127         struct drm_sched_entity *entity;
 128
 129         spin_lock(&rq->lock);
 130
 131         entity = rq->current_entity;
 132         if (entity) {
 133                 list_for_each_entry_continue(entity, &rq->entities, list) {
 134                         if (drm_sched_entity_is_ready(entity)) {
 135                                 rq->current_entity = entity;
 136                                 spin_unlock(&rq->lock);
 137                                 return entity;
 138                         }
 139                 }
 140         }
 141
 142         list_for_each_entry(entity, &rq->entities, list) {
 143
 144                 if (drm_sched_entity_is_ready(entity)) {
 145                         rq->current_entity = entity;
 146                         spin_unlock(&rq->lock);
 147                         return entity;
 148                 }
 149
 150                 if (entity == rq->current_entity)
 151                         break;
 152         }
 153
 154         spin_unlock(&rq->lock);
 155
 156         return NULL;
 157 }
 158
 159 /**
 160  * drm_sched_dependency_optimized
 161  *
 162  * @fence: the dependency fence
 163  * @entity: the entity which depends on the above fence
 164  *
 165  * Returns true if the dependency can be optimized and false otherwise
 166  */
 167 bool drm_sched_dependency_optimized(struct dma_fence* fence,
 168                                     struct drm_sched_entity *entity)
 169 {
 170         struct drm_gpu_scheduler *sched = entity->rq->sched;
 171         struct drm_sched_fence *s_fence;
 172
 173         if (!fence || dma_fence_is_signaled(fence))
 174                 return false;
 175         if (fence->context == entity->fence_context)
 176                 return true;
 177         s_fence = to_drm_sched_fence(fence);
 178         if (s_fence && s_fence->sched == sched)
 179                 return true;
 180
 181         return false;
 182 }
 183 EXPORT_SYMBOL(drm_sched_dependency_optimized);
 184
 185 /**
 186  * drm_sched_start_timeout - start timeout for reset worker
 187  *
 188  * @sched: scheduler instance to start the worker for
 189  *
 190  * Start the timeout for the given scheduler.
 191  */
 192 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
 193 {
 194         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 195             !list_empty(&sched->ring_mirror_list))
 196                 schedule_delayed_work(&sched->work_tdr, sched->timeout);
 197 }
 198
 199 /* job_finish is called after hw fence signaled
 200  */
 201 static void drm_sched_job_finish(struct work_struct *work)
 202 {
 203         struct drm_sched_job *s_job = container_of(work, struct drm_sched_job,
 204                                                    finish_work);
 205         struct drm_gpu_scheduler *sched = s_job->sched;
 206
 207         /*
 208          * Canceling the timeout without removing our job from the ring mirror
 209          * list is safe, as we will only end up in this worker if our jobs
 210          * finished fence has been signaled. So even if some another worker
 211          * manages to find this job as the next job in the list, the fence
 212          * signaled check below will prevent the timeout to be restarted.
 213          */
 214         cancel_delayed_work_sync(&sched->work_tdr);
 215
 216         spin_lock(&sched->job_list_lock);
 217         /* remove job from ring_mirror_list */
 218         list_del(&s_job->node);
 219         /* queue TDR for next job */
 220         drm_sched_start_timeout(sched);
 221         spin_unlock(&sched->job_list_lock);
 222
 223         dma_fence_put(&s_job->s_fence->finished);
 224         sched->ops->free_job(s_job);
 225 }
 226
 227 static void drm_sched_job_finish_cb(struct dma_fence *f,
 228                                     struct dma_fence_cb *cb)
 229 {
 230         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 231                                                  finish_cb);
 232         schedule_work(&job->finish_work);
 233 }
 234
 235 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 236 {
 237         struct drm_gpu_scheduler *sched = s_job->sched;
 238
 239         dma_fence_add_callback(&s_job->s_fence->finished, &s_job->finish_cb,
 240                                drm_sched_job_finish_cb);
 241
 242         spin_lock(&sched->job_list_lock);
 243         list_add_tail(&s_job->node, &sched->ring_mirror_list);
 244         drm_sched_start_timeout(sched);
 245         spin_unlock(&sched->job_list_lock);
 246 }
 247
 248 static void drm_sched_job_timedout(struct work_struct *work)
 249 {
 250         struct drm_gpu_scheduler *sched;
 251         struct drm_sched_job *job;
 252         int r;
 253
 254         sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 255
 256         spin_lock(&sched->job_list_lock);
 257         list_for_each_entry_reverse(job, &sched->ring_mirror_list, node) {
 258                 struct drm_sched_fence *fence = job->s_fence;
 259
 260                 if (!dma_fence_remove_callback(fence->parent, &fence->cb))
 261                         goto already_signaled;
 262         }
 263
 264         job = list_first_entry_or_null(&sched->ring_mirror_list,
 265                                        struct drm_sched_job, node);
 266         spin_unlock(&sched->job_list_lock);
 267
 268         if (job)
 269                 sched->ops->timedout_job(job);
 270
 271         spin_lock(&sched->job_list_lock);
 272         list_for_each_entry(job, &sched->ring_mirror_list, node) {
 273                 struct drm_sched_fence *fence = job->s_fence;
 274
 275                 if (!fence->parent || !list_empty(&fence->cb.node))
 276                         continue;
 277
 278                 r = dma_fence_add_callback(fence->parent, &fence->cb,
 279                                            drm_sched_process_job);
 280                 if (r)
 281                         drm_sched_process_job(fence->parent, &fence->cb);
 282
 283 already_signaled:
 284                 ;
 285         }
 286         spin_unlock(&sched->job_list_lock);
 287 }
 288
 289 /**
 290  * drm_sched_hw_job_reset - stop the scheduler if it contains the bad job
 291  *
 292  * @sched: scheduler instance
 293  * @bad: bad scheduler job
 294  *
 295  */
 296 void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 297 {
 298         struct drm_sched_job *s_job;
 299         struct drm_sched_entity *entity, *tmp;
 300         int i;
 301
 302         spin_lock(&sched->job_list_lock);
 303         list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
 304                 if (s_job->s_fence->parent &&
 305                     dma_fence_remove_callback(s_job->s_fence->parent,
 306                                               &s_job->s_fence->cb)) {
 307                         dma_fence_put(s_job->s_fence->parent);
 308                         s_job->s_fence->parent = NULL;
 309                         atomic_dec(&sched->hw_rq_count);
 310                 }
 311         }
 312         spin_unlock(&sched->job_list_lock);
 313
 314         if (bad && bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 315                 atomic_inc(&bad->karma);
 316                 /* don't increase @bad's karma if it's from KERNEL RQ,
 317                  * becuase sometimes GPU hang would cause kernel jobs (like VM updating jobs)
 318                  * corrupt but keep in mind that kernel jobs always considered good.
 319                  */
 320                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL; i++ ) {
 321                         struct drm_sched_rq *rq = &sched->sched_rq[i];
 322
 323                         spin_lock(&rq->lock);
 324                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
 325                                 if (bad->s_fence->scheduled.context == entity->fence_context) {
 326                                     if (atomic_read(&bad->karma) > bad->sched->hang_limit)
 327                                                 if (entity->guilty)
 328                                                         atomic_set(entity->guilty, 1);
 329                                         break;
 330                                 }
 331                         }
 332                         spin_unlock(&rq->lock);
 333                         if (&entity->list != &rq->entities)
 334                                 break;
 335                 }
 336         }
 337 }
 338 EXPORT_SYMBOL(drm_sched_hw_job_reset);
 339
 340 /**
 341  * drm_sched_job_recovery - recover jobs after a reset
 342  *
 343  * @sched: scheduler instance
 344  *
 345  */
 346 void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
 347 {
 348         struct drm_sched_job *s_job, *tmp;
 349         bool found_guilty = false;
 350         int r;
 351
 352         spin_lock(&sched->job_list_lock);
 353         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 354                 struct drm_sched_fence *s_fence = s_job->s_fence;
 355                 struct dma_fence *fence;
 356                 uint64_t guilty_context;
 357
 358                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 359                         found_guilty = true;
 360                         guilty_context = s_job->s_fence->scheduled.context;
 361                 }
 362
 363                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 364                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 365
 366                 spin_unlock(&sched->job_list_lock);
 367                 fence = sched->ops->run_job(s_job);
 368                 atomic_inc(&sched->hw_rq_count);
 369
 370                 if (fence) {
 371                         s_fence->parent = dma_fence_get(fence);
 372                         r = dma_fence_add_callback(fence, &s_fence->cb,
 373                                                    drm_sched_process_job);
 374                         if (r == -ENOENT)
 375                                 drm_sched_process_job(fence, &s_fence->cb);
 376                         else if (r)
 377                                 DRM_ERROR("fence add callback failed (%d)\n",
 378                                           r);
 379                         dma_fence_put(fence);
 380                 } else {
 381                         drm_sched_process_job(NULL, &s_fence->cb);
 382                 }
 383                 spin_lock(&sched->job_list_lock);
 384         }
 385         drm_sched_start_timeout(sched);
 386         spin_unlock(&sched->job_list_lock);
 387 }
 388 EXPORT_SYMBOL(drm_sched_job_recovery);
 389
 390 /**
 391  * drm_sched_job_init - init a scheduler job
 392  *
 393  * @job: scheduler job to init
 394  * @entity: scheduler entity to use
 395  * @owner: job owner for debugging
 396  *
 397  * Refer to drm_sched_entity_push_job() documentation
 398  * for locking considerations.
 399  *
 400  * Returns 0 for success, negative error code otherwise.
 401  */
 402 int drm_sched_job_init(struct drm_sched_job *job,
 403                        struct drm_sched_entity *entity,
 404                        void *owner)
 405 {
 406         struct drm_gpu_scheduler *sched;
 407
 408         drm_sched_entity_select_rq(entity);
 409         sched = entity->rq->sched;
 410
 411         job->sched = sched;
 412         job->entity = entity;
 413         job->s_priority = entity->rq - sched->sched_rq;
 414         job->s_fence = drm_sched_fence_create(entity, owner);
 415         if (!job->s_fence)
 416                 return -ENOMEM;
 417         job->id = atomic64_inc_return(&sched->job_id_count);
 418
 419         INIT_WORK(&job->finish_work, drm_sched_job_finish);
 420         INIT_LIST_HEAD(&job->node);
 421
 422         return 0;
 423 }
 424 EXPORT_SYMBOL(drm_sched_job_init);
 425
 426 /**
 427  * drm_sched_ready - is the scheduler ready
 428  *
 429  * @sched: scheduler instance
 430  *
 431  * Return true if we can push more jobs to the hw, otherwise false.
 432  */
 433 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 434 {
 435         return atomic_read(&sched->hw_rq_count) <
 436                 sched->hw_submission_limit;
 437 }
 438
 439 /**
 440  * drm_sched_wakeup - Wake up the scheduler when it is ready
 441  *
 442  * @sched: scheduler instance
 443  *
 444  */
 445 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 446 {
 447         if (drm_sched_ready(sched))
 448                 wake_up_interruptible(&sched->wake_up_worker);
 449 }
 450
 451 /**
 452  * drm_sched_select_entity - Select next entity to process
 453  *
 454  * @sched: scheduler instance
 455  *
 456  * Returns the entity to process or NULL if none are found.
 457  */
 458 static struct drm_sched_entity *
 459 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 460 {
 461         struct drm_sched_entity *entity;
 462         int i;
 463
 464         if (!drm_sched_ready(sched))
 465                 return NULL;
 466
 467         /* Kernel run queue has higher priority than normal run queue*/
 468         for (i = DRM_SCHED_PRIORITY_MAX - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 469                 entity = drm_sched_rq_select_entity(&sched->sched_rq[i]);
 470                 if (entity)
 471                         break;
 472         }
 473
 474         return entity;
 475 }
 476
 477 /**
 478  * drm_sched_process_job - process a job
 479  *
 480  * @f: fence
 481  * @cb: fence callbacks
 482  *
 483  * Called after job has finished execution.
 484  */
 485 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
 486 {
 487         struct drm_sched_fence *s_fence =
 488                 container_of(cb, struct drm_sched_fence, cb);
 489         struct drm_gpu_scheduler *sched = s_fence->sched;
 490
 491         dma_fence_get(&s_fence->finished);
 492         atomic_dec(&sched->hw_rq_count);
 493         atomic_dec(&sched->num_jobs);
 494         drm_sched_fence_finished(s_fence);
 495
 496         trace_drm_sched_process_job(s_fence);
 497         dma_fence_put(&s_fence->finished);
 498         wake_up_interruptible(&sched->wake_up_worker);
 499 }
 500
 501 /**
 502  * drm_sched_blocked - check if the scheduler is blocked
 503  *
 504  * @sched: scheduler instance
 505  *
 506  * Returns true if blocked, otherwise false.
 507  */
 508 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 509 {
 510         if (kthread_should_park()) {
 511                 kthread_parkme();
 512                 return true;
 513         }
 514
 515         return false;
 516 }
 517
 518 /**
 519  * drm_sched_main - main scheduler thread
 520  *
 521  * @param: scheduler instance
 522  *
 523  * Returns 0.
 524  */
 525 static int drm_sched_main(void *param)
 526 {
 527         struct sched_param sparam = {.sched_priority = 1};
 528         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
 529         int r;
 530
 531         sched_setscheduler(current, SCHED_FIFO, &sparam);
 532
 533         while (!kthread_should_stop()) {
 534                 struct drm_sched_entity *entity = NULL;
 535                 struct drm_sched_fence *s_fence;
 536                 struct drm_sched_job *sched_job;
 537                 struct dma_fence *fence;
 538
 539                 wait_event_interruptible(sched->wake_up_worker,
 540                                          (!drm_sched_blocked(sched) &&
 541                                           (entity = drm_sched_select_entity(sched))) ||
 542                                          kthread_should_stop());
 543
 544                 if (!entity)
 545                         continue;
 546
 547                 sched_job = drm_sched_entity_pop_job(entity);
 548                 if (!sched_job)
 549                         continue;
 550
 551                 s_fence = sched_job->s_fence;
 552
 553                 atomic_inc(&sched->hw_rq_count);
 554                 drm_sched_job_begin(sched_job);
 555
 556                 fence = sched->ops->run_job(sched_job);
 557                 drm_sched_fence_scheduled(s_fence);
 558
 559                 if (fence) {
 560                         s_fence->parent = dma_fence_get(fence);
 561                         r = dma_fence_add_callback(fence, &s_fence->cb,
 562                                                    drm_sched_process_job);
 563                         if (r == -ENOENT)
 564                                 drm_sched_process_job(fence, &s_fence->cb);
 565                         else if (r)
 566                                 DRM_ERROR("fence add callback failed (%d)\n",
 567                                           r);
 568                         dma_fence_put(fence);
 569                 } else {
 570                         drm_sched_process_job(NULL, &s_fence->cb);
 571                 }
 572
 573                 wake_up(&sched->job_scheduled);
 574         }
 575         return 0;
 576 }
 577
 578 /**
 579  * drm_sched_init - Init a gpu scheduler instance
 580  *
 581  * @sched: scheduler instance
 582  * @ops: backend operations for this scheduler
 583  * @hw_submission: number of hw submissions that can be in flight
 584  * @hang_limit: number of times to allow a job to hang before dropping it
 585  * @timeout: timeout value in jiffies for the scheduler
 586  * @name: name used for debugging
 587  *
 588  * Return 0 on success, otherwise error code.
 589  */
 590 int drm_sched_init(struct drm_gpu_scheduler *sched,
 591                    const struct drm_sched_backend_ops *ops,
 592                    unsigned hw_submission,
 593                    unsigned hang_limit,
 594                    long timeout,
 595                    const char *name)
 596 {
 597         int i;
 598         sched->ops = ops;
 599         sched->hw_submission_limit = hw_submission;
 600         sched->name = name;
 601         sched->timeout = timeout;
 602         sched->hang_limit = hang_limit;
 603         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_MAX; i++)
 604                 drm_sched_rq_init(sched, &sched->sched_rq[i]);
 605
 606         init_waitqueue_head(&sched->wake_up_worker);
 607         init_waitqueue_head(&sched->job_scheduled);
 608         INIT_LIST_HEAD(&sched->ring_mirror_list);
 609         spin_lock_init(&sched->job_list_lock);
 610         atomic_set(&sched->hw_rq_count, 0);
 611         INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
 612         atomic_set(&sched->num_jobs, 0);
 613         atomic64_set(&sched->job_id_count, 0);
 614
 615         /* Each scheduler will run on a seperate kernel thread */
 616         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
 617         if (IS_ERR(sched->thread)) {
 618                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 619                 return PTR_ERR(sched->thread);
 620         }
 621
 622         return 0;
 623 }
 624 EXPORT_SYMBOL(drm_sched_init);
 625
 626 /**
 627  * drm_sched_fini - Destroy a gpu scheduler
 628  *
 629  * @sched: scheduler instance
 630  *
 631  * Tears down and cleans up the scheduler.
 632  */
 633 void drm_sched_fini(struct drm_gpu_scheduler *sched)
 634 {
 635         if (sched->thread)
 636                 kthread_stop(sched->thread);
 637 }
 638 EXPORT_SYMBOL(drm_sched_fini);