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 #define CREATE_TRACE_POINTS
  31 #include "gpu_sched_trace.h"
  32
  33 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity);
  34 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched);
  35 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb);
  36
  37 struct kmem_cache *sched_fence_slab;
  38 atomic_t sched_fence_slab_ref = ATOMIC_INIT(0);
  39
  40 /* Initialize a given run queue struct */
  41 static void amd_sched_rq_init(struct amd_sched_rq *rq)
  42 {
  43         spin_lock_init(&rq->lock);
  44         INIT_LIST_HEAD(&rq->entities);
  45         rq->current_entity = NULL;
  46 }
  47
  48 static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
  49                                     struct amd_sched_entity *entity)
  50 {
  51         if (!list_empty(&entity->list))
  52                 return;
  53         spin_lock(&rq->lock);
  54         list_add_tail(&entity->list, &rq->entities);
  55         spin_unlock(&rq->lock);
  56 }
  57
  58 static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
  59                                        struct amd_sched_entity *entity)
  60 {
  61         if (list_empty(&entity->list))
  62                 return;
  63         spin_lock(&rq->lock);
  64         list_del_init(&entity->list);
  65         if (rq->current_entity == entity)
  66                 rq->current_entity = NULL;
  67         spin_unlock(&rq->lock);
  68 }
  69
  70 /**
  71  * Select an entity which could provide a job to run
  72  *
  73  * @rq          The run queue to check.
  74  *
  75  * Try to find a ready entity, returns NULL if none found.
  76  */
  77 static struct amd_sched_entity *
  78 amd_sched_rq_select_entity(struct amd_sched_rq *rq)
  79 {
  80         struct amd_sched_entity *entity;
  81
  82         spin_lock(&rq->lock);
  83
  84         entity = rq->current_entity;
  85         if (entity) {
  86                 list_for_each_entry_continue(entity, &rq->entities, list) {
  87                         if (amd_sched_entity_is_ready(entity)) {
  88                                 rq->current_entity = entity;
  89                                 spin_unlock(&rq->lock);
  90                                 return entity;
  91                         }
  92                 }
  93         }
  94
  95         list_for_each_entry(entity, &rq->entities, list) {
  96
  97                 if (amd_sched_entity_is_ready(entity)) {
  98                         rq->current_entity = entity;
  99                         spin_unlock(&rq->lock);
 100                         return entity;
 101                 }
 102
 103                 if (entity == rq->current_entity)
 104                         break;
 105         }
 106
 107         spin_unlock(&rq->lock);
 108
 109         return NULL;
 110 }
 111
 112 /**
 113  * Init a context entity used by scheduler when submit to HW ring.
 114  *
 115  * @sched       The pointer to the scheduler
 116  * @entity      The pointer to a valid amd_sched_entity
 117  * @rq          The run queue this entity belongs
 118  * @kernel      If this is an entity for the kernel
 119  * @jobs        The max number of jobs in the job queue
 120  *
 121  * return 0 if succeed. negative error code on failure
 122 */
 123 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 124                           struct amd_sched_entity *entity,
 125                           struct amd_sched_rq *rq,
 126                           uint32_t jobs)
 127 {
 128         int r;
 129
 130         if (!(sched && entity && rq))
 131                 return -EINVAL;
 132
 133         memset(entity, 0, sizeof(struct amd_sched_entity));
 134         INIT_LIST_HEAD(&entity->list);
 135         entity->rq = rq;
 136         entity->sched = sched;
 137
 138         spin_lock_init(&entity->queue_lock);
 139         r = kfifo_alloc(&entity->job_queue, jobs * sizeof(void *), GFP_KERNEL);
 140         if (r)
 141                 return r;
 142
 143         atomic_set(&entity->fence_seq, 0);
 144         entity->fence_context = fence_context_alloc(2);
 145
 146         return 0;
 147 }
 148
 149 /**
 150  * Query if entity is initialized
 151  *
 152  * @sched       Pointer to scheduler instance
 153  * @entity      The pointer to a valid scheduler entity
 154  *
 155  * return true if entity is initialized, false otherwise
 156 */
 157 static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
 158                                             struct amd_sched_entity *entity)
 159 {
 160         return entity->sched == sched &&
 161                 entity->rq != NULL;
 162 }
 163
 164 /**
 165  * Check if entity is idle
 166  *
 167  * @entity      The pointer to a valid scheduler entity
 168  *
 169  * Return true if entity don't has any unscheduled jobs.
 170  */
 171 static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
 172 {
 173         rmb();
 174         if (kfifo_is_empty(&entity->job_queue))
 175                 return true;
 176
 177         return false;
 178 }
 179
 180 /**
 181  * Check if entity is ready
 182  *
 183  * @entity      The pointer to a valid scheduler entity
 184  *
 185  * Return true if entity could provide a job.
 186  */
 187 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity)
 188 {
 189         if (kfifo_is_empty(&entity->job_queue))
 190                 return false;
 191
 192         if (ACCESS_ONCE(entity->dependency))
 193                 return false;
 194
 195         return true;
 196 }
 197
 198 /**
 199  * Destroy a context entity
 200  *
 201  * @sched       Pointer to scheduler instance
 202  * @entity      The pointer to a valid scheduler entity
 203  *
 204  * Cleanup and free the allocated resources.
 205  */
 206 void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
 207                            struct amd_sched_entity *entity)
 208 {
 209         struct amd_sched_rq *rq = entity->rq;
 210
 211         if (!amd_sched_entity_is_initialized(sched, entity))
 212                 return;
 213
 214         /**
 215          * The client will not queue more IBs during this fini, consume existing
 216          * queued IBs
 217         */
 218         wait_event(sched->job_scheduled, amd_sched_entity_is_idle(entity));
 219
 220         amd_sched_rq_remove_entity(rq, entity);
 221         kfifo_free(&entity->job_queue);
 222 }
 223
 224 static void amd_sched_entity_wakeup(struct fence *f, struct fence_cb *cb)
 225 {
 226         struct amd_sched_entity *entity =
 227                 container_of(cb, struct amd_sched_entity, cb);
 228         entity->dependency = NULL;
 229         fence_put(f);
 230         amd_sched_wakeup(entity->sched);
 231 }
 232
 233 static void amd_sched_entity_clear_dep(struct fence *f, struct fence_cb *cb)
 234 {
 235         struct amd_sched_entity *entity =
 236                 container_of(cb, struct amd_sched_entity, cb);
 237         entity->dependency = NULL;
 238         fence_put(f);
 239 }
 240
 241 static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
 242 {
 243         struct amd_gpu_scheduler *sched = entity->sched;
 244         struct fence * fence = entity->dependency;
 245         struct amd_sched_fence *s_fence;
 246
 247         if (fence->context == entity->fence_context) {
 248                 /* We can ignore fences from ourself */
 249                 fence_put(entity->dependency);
 250                 return false;
 251         }
 252
 253         s_fence = to_amd_sched_fence(fence);
 254         if (s_fence && s_fence->sched == sched) {
 255
 256                 /*
 257                  * Fence is from the same scheduler, only need to wait for
 258                  * it to be scheduled
 259                  */
 260                 fence = fence_get(&s_fence->scheduled);
 261                 fence_put(entity->dependency);
 262                 entity->dependency = fence;
 263                 if (!fence_add_callback(fence, &entity->cb,
 264                                         amd_sched_entity_clear_dep))
 265                         return true;
 266
 267                 /* Ignore it when it is already scheduled */
 268                 fence_put(fence);
 269                 return false;
 270         }
 271
 272         if (!fence_add_callback(entity->dependency, &entity->cb,
 273                                 amd_sched_entity_wakeup))
 274                 return true;
 275
 276         fence_put(entity->dependency);
 277         return false;
 278 }
 279
 280 static struct amd_sched_job *
 281 amd_sched_entity_pop_job(struct amd_sched_entity *entity)
 282 {
 283         struct amd_gpu_scheduler *sched = entity->sched;
 284         struct amd_sched_job *sched_job;
 285
 286         if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job)))
 287                 return NULL;
 288
 289         while ((entity->dependency = sched->ops->dependency(sched_job)))
 290                 if (amd_sched_entity_add_dependency_cb(entity))
 291                         return NULL;
 292
 293         return sched_job;
 294 }
 295
 296 /**
 297  * Helper to submit a job to the job queue
 298  *
 299  * @sched_job           The pointer to job required to submit
 300  *
 301  * Returns true if we could submit the job.
 302  */
 303 static bool amd_sched_entity_in(struct amd_sched_job *sched_job)
 304 {
 305         struct amd_gpu_scheduler *sched = sched_job->sched;
 306         struct amd_sched_entity *entity = sched_job->s_entity;
 307         bool added, first = false;
 308
 309         spin_lock(&entity->queue_lock);
 310         added = kfifo_in(&entity->job_queue, &sched_job,
 311                         sizeof(sched_job)) == sizeof(sched_job);
 312
 313         if (added && kfifo_len(&entity->job_queue) == sizeof(sched_job))
 314                 first = true;
 315
 316         spin_unlock(&entity->queue_lock);
 317
 318         /* first job wakes up scheduler */
 319         if (first) {
 320                 /* Add the entity to the run queue */
 321                 amd_sched_rq_add_entity(entity->rq, entity);
 322                 amd_sched_wakeup(sched);
 323         }
 324         return added;
 325 }
 326
 327 /* job_finish is called after hw fence signaled, and
 328  * the job had already been deleted from ring_mirror_list
 329  */
 330 static void amd_sched_job_finish(struct work_struct *work)
 331 {
 332         struct amd_sched_job *s_job = container_of(work, struct amd_sched_job,
 333                                                    finish_work);
 334         struct amd_gpu_scheduler *sched = s_job->sched;
 335
 336         /* remove job from ring_mirror_list */
 337         spin_lock(&sched->job_list_lock);
 338         list_del_init(&s_job->node);
 339         if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
 340                 struct amd_sched_job *next;
 341
 342                 spin_unlock(&sched->job_list_lock);
 343                 cancel_delayed_work_sync(&s_job->work_tdr);
 344                 spin_lock(&sched->job_list_lock);
 345
 346                 /* queue TDR for next job */
 347                 next = list_first_entry_or_null(&sched->ring_mirror_list,
 348                                                 struct amd_sched_job, node);
 349
 350                 if (next)
 351                         schedule_delayed_work(&next->work_tdr, sched->timeout);
 352         }
 353         spin_unlock(&sched->job_list_lock);
 354         sched->ops->free_job(s_job);
 355 }
 356
 357 static void amd_sched_job_finish_cb(struct fence *f, struct fence_cb *cb)
 358 {
 359         struct amd_sched_job *job = container_of(cb, struct amd_sched_job,
 360                                                  finish_cb);
 361         schedule_work(&job->finish_work);
 362 }
 363
 364 static void amd_sched_job_begin(struct amd_sched_job *s_job)
 365 {
 366         struct amd_gpu_scheduler *sched = s_job->sched;
 367
 368         spin_lock(&sched->job_list_lock);
 369         list_add_tail(&s_job->node, &sched->ring_mirror_list);
 370         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 371             list_first_entry_or_null(&sched->ring_mirror_list,
 372                                      struct amd_sched_job, node) == s_job)
 373                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
 374         spin_unlock(&sched->job_list_lock);
 375 }
 376
 377 static void amd_sched_job_timedout(struct work_struct *work)
 378 {
 379         struct amd_sched_job *job = container_of(work, struct amd_sched_job,
 380                                                  work_tdr.work);
 381
 382         job->sched->ops->timedout_job(job);
 383 }
 384
 385 void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched)
 386 {
 387         struct amd_sched_job *s_job;
 388
 389         spin_lock(&sched->job_list_lock);
 390         list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
 391                 if (fence_remove_callback(s_job->s_fence->parent, &s_job->s_fence->cb)) {
 392                         fence_put(s_job->s_fence->parent);
 393                         s_job->s_fence->parent = NULL;
 394                 }
 395         }
 396         atomic_set(&sched->hw_rq_count, 0);
 397         spin_unlock(&sched->job_list_lock);
 398 }
 399
 400 void amd_sched_job_recovery(struct amd_gpu_scheduler *sched)
 401 {
 402         struct amd_sched_job *s_job, *tmp;
 403         int r;
 404
 405         spin_lock(&sched->job_list_lock);
 406         s_job = list_first_entry_or_null(&sched->ring_mirror_list,
 407                                          struct amd_sched_job, node);
 408         if (s_job && sched->timeout != MAX_SCHEDULE_TIMEOUT)
 409                 schedule_delayed_work(&s_job->work_tdr, sched->timeout);
 410
 411         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 412                 struct amd_sched_fence *s_fence = s_job->s_fence;
 413                 struct fence *fence;
 414
 415                 spin_unlock(&sched->job_list_lock);
 416                 fence = sched->ops->run_job(s_job);
 417                 atomic_inc(&sched->hw_rq_count);
 418                 if (fence) {
 419                         s_fence->parent = fence_get(fence);
 420                         r = fence_add_callback(fence, &s_fence->cb,
 421                                                amd_sched_process_job);
 422                         if (r == -ENOENT)
 423                                 amd_sched_process_job(fence, &s_fence->cb);
 424                         else if (r)
 425                                 DRM_ERROR("fence add callback failed (%d)\n",
 426                                           r);
 427                         fence_put(fence);
 428                 } else {
 429                         DRM_ERROR("Failed to run job!\n");
 430                         amd_sched_process_job(NULL, &s_fence->cb);
 431                 }
 432                 spin_lock(&sched->job_list_lock);
 433         }
 434         spin_unlock(&sched->job_list_lock);
 435 }
 436
 437 /**
 438  * Submit a job to the job queue
 439  *
 440  * @sched_job           The pointer to job required to submit
 441  *
 442  * Returns 0 for success, negative error code otherwise.
 443  */
 444 void amd_sched_entity_push_job(struct amd_sched_job *sched_job)
 445 {
 446         struct amd_sched_entity *entity = sched_job->s_entity;
 447
 448         trace_amd_sched_job(sched_job);
 449         fence_add_callback(&sched_job->s_fence->finished, &sched_job->finish_cb,
 450                            amd_sched_job_finish_cb);
 451         wait_event(entity->sched->job_scheduled,
 452                    amd_sched_entity_in(sched_job));
 453 }
 454
 455 /* init a sched_job with basic field */
 456 int amd_sched_job_init(struct amd_sched_job *job,
 457                        struct amd_gpu_scheduler *sched,
 458                        struct amd_sched_entity *entity,
 459                        void *owner)
 460 {
 461         job->sched = sched;
 462         job->s_entity = entity;
 463         job->s_fence = amd_sched_fence_create(entity, owner);
 464         if (!job->s_fence)
 465                 return -ENOMEM;
 466
 467         INIT_WORK(&job->finish_work, amd_sched_job_finish);
 468         INIT_LIST_HEAD(&job->node);
 469         INIT_DELAYED_WORK(&job->work_tdr, amd_sched_job_timedout);
 470
 471         return 0;
 472 }
 473
 474 /**
 475  * Return ture if we can push more jobs to the hw.
 476  */
 477 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
 478 {
 479         return atomic_read(&sched->hw_rq_count) <
 480                 sched->hw_submission_limit;
 481 }
 482
 483 /**
 484  * Wake up the scheduler when it is ready
 485  */
 486 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched)
 487 {
 488         if (amd_sched_ready(sched))
 489                 wake_up_interruptible(&sched->wake_up_worker);
 490 }
 491
 492 /**
 493  * Select next entity to process
 494 */
 495 static struct amd_sched_entity *
 496 amd_sched_select_entity(struct amd_gpu_scheduler *sched)
 497 {
 498         struct amd_sched_entity *entity;
 499         int i;
 500
 501         if (!amd_sched_ready(sched))
 502                 return NULL;
 503
 504         /* Kernel run queue has higher priority than normal run queue*/
 505         for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++) {
 506                 entity = amd_sched_rq_select_entity(&sched->sched_rq[i]);
 507                 if (entity)
 508                         break;
 509         }
 510
 511         return entity;
 512 }
 513
 514 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 515 {
 516         struct amd_sched_fence *s_fence =
 517                 container_of(cb, struct amd_sched_fence, cb);
 518         struct amd_gpu_scheduler *sched = s_fence->sched;
 519
 520         atomic_dec(&sched->hw_rq_count);
 521         amd_sched_fence_finished(s_fence);
 522
 523         trace_amd_sched_process_job(s_fence);
 524         fence_put(&s_fence->finished);
 525         wake_up_interruptible(&sched->wake_up_worker);
 526 }
 527
 528 static bool amd_sched_blocked(struct amd_gpu_scheduler *sched)
 529 {
 530         if (kthread_should_park()) {
 531                 kthread_parkme();
 532                 return true;
 533         }
 534
 535         return false;
 536 }
 537
 538 static int amd_sched_main(void *param)
 539 {
 540         struct sched_param sparam = {.sched_priority = 1};
 541         struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
 542         int r, count;
 543
 544         sched_setscheduler(current, SCHED_FIFO, &sparam);
 545
 546         while (!kthread_should_stop()) {
 547                 struct amd_sched_entity *entity = NULL;
 548                 struct amd_sched_fence *s_fence;
 549                 struct amd_sched_job *sched_job;
 550                 struct fence *fence;
 551
 552                 wait_event_interruptible(sched->wake_up_worker,
 553                                          (!amd_sched_blocked(sched) &&
 554                                           (entity = amd_sched_select_entity(sched))) ||
 555                                          kthread_should_stop());
 556
 557                 if (!entity)
 558                         continue;
 559
 560                 sched_job = amd_sched_entity_pop_job(entity);
 561                 if (!sched_job)
 562                         continue;
 563
 564                 s_fence = sched_job->s_fence;
 565
 566                 atomic_inc(&sched->hw_rq_count);
 567                 amd_sched_job_begin(sched_job);
 568
 569                 fence = sched->ops->run_job(sched_job);
 570                 amd_sched_fence_scheduled(s_fence);
 571                 if (fence) {
 572                         s_fence->parent = fence_get(fence);
 573                         r = fence_add_callback(fence, &s_fence->cb,
 574                                                amd_sched_process_job);
 575                         if (r == -ENOENT)
 576                                 amd_sched_process_job(fence, &s_fence->cb);
 577                         else if (r)
 578                                 DRM_ERROR("fence add callback failed (%d)\n",
 579                                           r);
 580                         fence_put(fence);
 581                 } else {
 582                         DRM_ERROR("Failed to run job!\n");
 583                         amd_sched_process_job(NULL, &s_fence->cb);
 584                 }
 585
 586                 count = kfifo_out(&entity->job_queue, &sched_job,
 587                                 sizeof(sched_job));
 588                 WARN_ON(count != sizeof(sched_job));
 589                 wake_up(&sched->job_scheduled);
 590         }
 591         return 0;
 592 }
 593
 594 /**
 595  * Init a gpu scheduler instance
 596  *
 597  * @sched               The pointer to the scheduler
 598  * @ops                 The backend operations for this scheduler.
 599  * @hw_submissions      Number of hw submissions to do.
 600  * @name                Name used for debugging
 601  *
 602  * Return 0 on success, otherwise error code.
 603 */
 604 int amd_sched_init(struct amd_gpu_scheduler *sched,
 605                    const struct amd_sched_backend_ops *ops,
 606                    unsigned hw_submission, long timeout, const char *name)
 607 {
 608         int i;
 609         sched->ops = ops;
 610         sched->hw_submission_limit = hw_submission;
 611         sched->name = name;
 612         sched->timeout = timeout;
 613         for (i = 0; i < AMD_SCHED_MAX_PRIORITY; i++)
 614                 amd_sched_rq_init(&sched->sched_rq[i]);
 615
 616         init_waitqueue_head(&sched->wake_up_worker);
 617         init_waitqueue_head(&sched->job_scheduled);
 618         INIT_LIST_HEAD(&sched->ring_mirror_list);
 619         spin_lock_init(&sched->job_list_lock);
 620         atomic_set(&sched->hw_rq_count, 0);
 621         if (atomic_inc_return(&sched_fence_slab_ref) == 1) {
 622                 sched_fence_slab = kmem_cache_create(
 623                         "amd_sched_fence", sizeof(struct amd_sched_fence), 0,
 624                         SLAB_HWCACHE_ALIGN, NULL);
 625                 if (!sched_fence_slab)
 626                         return -ENOMEM;
 627         }
 628
 629         /* Each scheduler will run on a seperate kernel thread */
 630         sched->thread = kthread_run(amd_sched_main, sched, sched->name);
 631         if (IS_ERR(sched->thread)) {
 632                 DRM_ERROR("Failed to create scheduler for %s.\n", name);
 633                 return PTR_ERR(sched->thread);
 634         }
 635
 636         return 0;
 637 }
 638
 639 /**
 640  * Destroy a gpu scheduler
 641  *
 642  * @sched       The pointer to the scheduler
 643  */
 644 void amd_sched_fini(struct amd_gpu_scheduler *sched)
 645 {
 646         if (sched->thread)
 647                 kthread_stop(sched->thread);
 648         if (atomic_dec_and_test(&sched_fence_slab_ref))
 649                 kmem_cache_destroy(sched_fence_slab);
 650 }