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 <linux/completion.h>
  51 #include <linux/dma-resv.h>
  52 #include <uapi/linux/sched/types.h>
  53
  54 #include <drm/drm_print.h>
  55 #include <drm/drm_gem.h>
  56 #include <drm/gpu_scheduler.h>
  57 #include <drm/spsc_queue.h>
  58
  59 #define CREATE_TRACE_POINTS
  60 #include "gpu_scheduler_trace.h"
  61
  62 #define to_drm_sched_job(sched_job)             \
  63                 container_of((sched_job), struct drm_sched_job, queue_node)
  64
  65 int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
  66
  67 /**
  68  * DOC: sched_policy (int)
  69  * Used to override default entities scheduling policy in a run queue.
  70  */
  71 MODULE_PARM_DESC(sched_policy, "Specify the scheduling policy for entities on a run-queue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin, " __stringify(DRM_SCHED_POLICY_FIFO) " = FIFO (default).");
  72 module_param_named(sched_policy, drm_sched_policy, int, 0444);
  73
  74 static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
  75                                                             const struct rb_node *b)
  76 {
  77         struct drm_sched_entity *ent_a =  rb_entry((a), struct drm_sched_entity, rb_tree_node);
  78         struct drm_sched_entity *ent_b =  rb_entry((b), struct drm_sched_entity, rb_tree_node);
  79
  80         return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
  81 }
  82
  83 static inline void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity)
  84 {
  85         struct drm_sched_rq *rq = entity->rq;
  86
  87         if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
  88                 rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
  89                 RB_CLEAR_NODE(&entity->rb_tree_node);
  90         }
  91 }
  92
  93 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts)
  94 {
  95         /*
  96          * Both locks need to be grabbed, one to protect from entity->rq change
  97          * for entity from within concurrent drm_sched_entity_select_rq and the
  98          * other to update the rb tree structure.
  99          */
 100         spin_lock(&entity->rq_lock);
 101         spin_lock(&entity->rq->lock);
 102
 103         drm_sched_rq_remove_fifo_locked(entity);
 104
 105         entity->oldest_job_waiting = ts;
 106
 107         rb_add_cached(&entity->rb_tree_node, &entity->rq->rb_tree_root,
 108                       drm_sched_entity_compare_before);
 109
 110         spin_unlock(&entity->rq->lock);
 111         spin_unlock(&entity->rq_lock);
 112 }
 113
 114 /**
 115  * drm_sched_rq_init - initialize a given run queue struct
 116  *
 117  * @sched: scheduler instance to associate with this run queue
 118  * @rq: scheduler run queue
 119  *
 120  * Initializes a scheduler runqueue.
 121  */
 122 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
 123                               struct drm_sched_rq *rq)
 124 {
 125         spin_lock_init(&rq->lock);
 126         INIT_LIST_HEAD(&rq->entities);
 127         rq->rb_tree_root = RB_ROOT_CACHED;
 128         rq->current_entity = NULL;
 129         rq->sched = sched;
 130 }
 131
 132 /**
 133  * drm_sched_rq_add_entity - add an entity
 134  *
 135  * @rq: scheduler run queue
 136  * @entity: scheduler entity
 137  *
 138  * Adds a scheduler entity to the run queue.
 139  */
 140 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
 141                              struct drm_sched_entity *entity)
 142 {
 143         if (!list_empty(&entity->list))
 144                 return;
 145
 146         spin_lock(&rq->lock);
 147
 148         atomic_inc(rq->sched->score);
 149         list_add_tail(&entity->list, &rq->entities);
 150
 151         spin_unlock(&rq->lock);
 152 }
 153
 154 /**
 155  * drm_sched_rq_remove_entity - remove an entity
 156  *
 157  * @rq: scheduler run queue
 158  * @entity: scheduler entity
 159  *
 160  * Removes a scheduler entity from the run queue.
 161  */
 162 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 163                                 struct drm_sched_entity *entity)
 164 {
 165         if (list_empty(&entity->list))
 166                 return;
 167
 168         spin_lock(&rq->lock);
 169
 170         atomic_dec(rq->sched->score);
 171         list_del_init(&entity->list);
 172
 173         if (rq->current_entity == entity)
 174                 rq->current_entity = NULL;
 175
 176         if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 177                 drm_sched_rq_remove_fifo_locked(entity);
 178
 179         spin_unlock(&rq->lock);
 180 }
 181
 182 /**
 183  * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
 184  *
 185  * @rq: scheduler run queue to check.
 186  *
 187  * Try to find a ready entity, returns NULL if none found.
 188  */
 189 static struct drm_sched_entity *
 190 drm_sched_rq_select_entity_rr(struct drm_sched_rq *rq)
 191 {
 192         struct drm_sched_entity *entity;
 193
 194         spin_lock(&rq->lock);
 195
 196         entity = rq->current_entity;
 197         if (entity) {
 198                 list_for_each_entry_continue(entity, &rq->entities, list) {
 199                         if (drm_sched_entity_is_ready(entity)) {
 200                                 rq->current_entity = entity;
 201                                 reinit_completion(&entity->entity_idle);
 202                                 spin_unlock(&rq->lock);
 203                                 return entity;
 204                         }
 205                 }
 206         }
 207
 208         list_for_each_entry(entity, &rq->entities, list) {
 209
 210                 if (drm_sched_entity_is_ready(entity)) {
 211                         rq->current_entity = entity;
 212                         reinit_completion(&entity->entity_idle);
 213                         spin_unlock(&rq->lock);
 214                         return entity;
 215                 }
 216
 217                 if (entity == rq->current_entity)
 218                         break;
 219         }
 220
 221         spin_unlock(&rq->lock);
 222
 223         return NULL;
 224 }
 225
 226 /**
 227  * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
 228  *
 229  * @rq: scheduler run queue to check.
 230  *
 231  * Find oldest waiting ready entity, returns NULL if none found.
 232  */
 233 static struct drm_sched_entity *
 234 drm_sched_rq_select_entity_fifo(struct drm_sched_rq *rq)
 235 {
 236         struct rb_node *rb;
 237
 238         spin_lock(&rq->lock);
 239         for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
 240                 struct drm_sched_entity *entity;
 241
 242                 entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
 243                 if (drm_sched_entity_is_ready(entity)) {
 244                         rq->current_entity = entity;
 245                         reinit_completion(&entity->entity_idle);
 246                         break;
 247                 }
 248         }
 249         spin_unlock(&rq->lock);
 250
 251         return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
 252 }
 253
 254 /**
 255  * drm_sched_job_done - complete a job
 256  * @s_job: pointer to the job which is done
 257  *
 258  * Finish the job's fence and wake up the worker thread.
 259  */
 260 static void drm_sched_job_done(struct drm_sched_job *s_job)
 261 {
 262         struct drm_sched_fence *s_fence = s_job->s_fence;
 263         struct drm_gpu_scheduler *sched = s_fence->sched;
 264
 265         atomic_dec(&sched->hw_rq_count);
 266         atomic_dec(sched->score);
 267
 268         trace_drm_sched_process_job(s_fence);
 269
 270         dma_fence_get(&s_fence->finished);
 271         drm_sched_fence_finished(s_fence);
 272         dma_fence_put(&s_fence->finished);
 273         wake_up_interruptible(&sched->wake_up_worker);
 274 }
 275
 276 /**
 277  * drm_sched_job_done_cb - the callback for a done job
 278  * @f: fence
 279  * @cb: fence callbacks
 280  */
 281 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
 282 {
 283         struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
 284
 285         drm_sched_job_done(s_job);
 286 }
 287
 288 /**
 289  * drm_sched_start_timeout - start timeout for reset worker
 290  *
 291  * @sched: scheduler instance to start the worker for
 292  *
 293  * Start the timeout for the given scheduler.
 294  */
 295 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
 296 {
 297         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 298             !list_empty(&sched->pending_list))
 299                 queue_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
 300 }
 301
 302 /**
 303  * drm_sched_fault - immediately start timeout handler
 304  *
 305  * @sched: scheduler where the timeout handling should be started.
 306  *
 307  * Start timeout handling immediately when the driver detects a hardware fault.
 308  */
 309 void drm_sched_fault(struct drm_gpu_scheduler *sched)
 310 {
 311         if (sched->ready)
 312                 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
 313 }
 314 EXPORT_SYMBOL(drm_sched_fault);
 315
 316 /**
 317  * drm_sched_suspend_timeout - Suspend scheduler job timeout
 318  *
 319  * @sched: scheduler instance for which to suspend the timeout
 320  *
 321  * Suspend the delayed work timeout for the scheduler. This is done by
 322  * modifying the delayed work timeout to an arbitrary large value,
 323  * MAX_SCHEDULE_TIMEOUT in this case.
 324  *
 325  * Returns the timeout remaining
 326  *
 327  */
 328 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
 329 {
 330         unsigned long sched_timeout, now = jiffies;
 331
 332         sched_timeout = sched->work_tdr.timer.expires;
 333
 334         /*
 335          * Modify the timeout to an arbitrarily large value. This also prevents
 336          * the timeout to be restarted when new submissions arrive
 337          */
 338         if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
 339                         && time_after(sched_timeout, now))
 340                 return sched_timeout - now;
 341         else
 342                 return sched->timeout;
 343 }
 344 EXPORT_SYMBOL(drm_sched_suspend_timeout);
 345
 346 /**
 347  * drm_sched_resume_timeout - Resume scheduler job timeout
 348  *
 349  * @sched: scheduler instance for which to resume the timeout
 350  * @remaining: remaining timeout
 351  *
 352  * Resume the delayed work timeout for the scheduler.
 353  */
 354 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
 355                 unsigned long remaining)
 356 {
 357         spin_lock(&sched->job_list_lock);
 358
 359         if (list_empty(&sched->pending_list))
 360                 cancel_delayed_work(&sched->work_tdr);
 361         else
 362                 mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
 363
 364         spin_unlock(&sched->job_list_lock);
 365 }
 366 EXPORT_SYMBOL(drm_sched_resume_timeout);
 367
 368 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 369 {
 370         struct drm_gpu_scheduler *sched = s_job->sched;
 371
 372         spin_lock(&sched->job_list_lock);
 373         list_add_tail(&s_job->list, &sched->pending_list);
 374         drm_sched_start_timeout(sched);
 375         spin_unlock(&sched->job_list_lock);
 376 }
 377
 378 static void drm_sched_job_timedout(struct work_struct *work)
 379 {
 380         struct drm_gpu_scheduler *sched;
 381         struct drm_sched_job *job;
 382         enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_NOMINAL;
 383
 384         sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 385
 386         /* Protects against concurrent deletion in drm_sched_get_cleanup_job */
 387         spin_lock(&sched->job_list_lock);
 388         job = list_first_entry_or_null(&sched->pending_list,
 389                                        struct drm_sched_job, list);
 390
 391         if (job) {
 392                 /*
 393                  * Remove the bad job so it cannot be freed by concurrent
 394                  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
 395                  * is parked at which point it's safe.
 396                  */
 397                 list_del_init(&job->list);
 398                 spin_unlock(&sched->job_list_lock);
 399
 400                 status = job->sched->ops->timedout_job(job);
 401
 402                 /*
 403                  * Guilty job did complete and hence needs to be manually removed
 404                  * See drm_sched_stop doc.
 405                  */
 406                 if (sched->free_guilty) {
 407                         job->sched->ops->free_job(job);
 408                         sched->free_guilty = false;
 409                 }
 410         } else {
 411                 spin_unlock(&sched->job_list_lock);
 412         }
 413
 414         if (status != DRM_GPU_SCHED_STAT_ENODEV) {
 415                 spin_lock(&sched->job_list_lock);
 416                 drm_sched_start_timeout(sched);
 417                 spin_unlock(&sched->job_list_lock);
 418         }
 419 }
 420
 421 /**
 422  * drm_sched_stop - stop the scheduler
 423  *
 424  * @sched: scheduler instance
 425  * @bad: job which caused the time out
 426  *
 427  * Stop the scheduler and also removes and frees all completed jobs.
 428  * Note: bad job will not be freed as it might be used later and so it's
 429  * callers responsibility to release it manually if it's not part of the
 430  * pending list any more.
 431  *
 432  */
 433 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 434 {
 435         struct drm_sched_job *s_job, *tmp;
 436
 437         kthread_park(sched->thread);
 438
 439         /*
 440          * Reinsert back the bad job here - now it's safe as
 441          * drm_sched_get_cleanup_job cannot race against us and release the
 442          * bad job at this point - we parked (waited for) any in progress
 443          * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
 444          * now until the scheduler thread is unparked.
 445          */
 446         if (bad && bad->sched == sched)
 447                 /*
 448                  * Add at the head of the queue to reflect it was the earliest
 449                  * job extracted.
 450                  */
 451                 list_add(&bad->list, &sched->pending_list);
 452
 453         /*
 454          * Iterate the job list from later to  earlier one and either deactive
 455          * their HW callbacks or remove them from pending list if they already
 456          * signaled.
 457          * This iteration is thread safe as sched thread is stopped.
 458          */
 459         list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
 460                                          list) {
 461                 if (s_job->s_fence->parent &&
 462                     dma_fence_remove_callback(s_job->s_fence->parent,
 463                                               &s_job->cb)) {
 464                         dma_fence_put(s_job->s_fence->parent);
 465                         s_job->s_fence->parent = NULL;
 466                         atomic_dec(&sched->hw_rq_count);
 467                 } else {
 468                         /*
 469                          * remove job from pending_list.
 470                          * Locking here is for concurrent resume timeout
 471                          */
 472                         spin_lock(&sched->job_list_lock);
 473                         list_del_init(&s_job->list);
 474                         spin_unlock(&sched->job_list_lock);
 475
 476                         /*
 477                          * Wait for job's HW fence callback to finish using s_job
 478                          * before releasing it.
 479                          *
 480                          * Job is still alive so fence refcount at least 1
 481                          */
 482                         dma_fence_wait(&s_job->s_fence->finished, false);
 483
 484                         /*
 485                          * We must keep bad job alive for later use during
 486                          * recovery by some of the drivers but leave a hint
 487                          * that the guilty job must be released.
 488                          */
 489                         if (bad != s_job)
 490                                 sched->ops->free_job(s_job);
 491                         else
 492                                 sched->free_guilty = true;
 493                 }
 494         }
 495
 496         /*
 497          * Stop pending timer in flight as we rearm it in  drm_sched_start. This
 498          * avoids the pending timeout work in progress to fire right away after
 499          * this TDR finished and before the newly restarted jobs had a
 500          * chance to complete.
 501          */
 502         cancel_delayed_work(&sched->work_tdr);
 503 }
 504
 505 EXPORT_SYMBOL(drm_sched_stop);
 506
 507 /**
 508  * drm_sched_start - recover jobs after a reset
 509  *
 510  * @sched: scheduler instance
 511  * @full_recovery: proceed with complete sched restart
 512  *
 513  */
 514 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery)
 515 {
 516         struct drm_sched_job *s_job, *tmp;
 517         int r;
 518
 519         /*
 520          * Locking the list is not required here as the sched thread is parked
 521          * so no new jobs are being inserted or removed. Also concurrent
 522          * GPU recovers can't run in parallel.
 523          */
 524         list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 525                 struct dma_fence *fence = s_job->s_fence->parent;
 526
 527                 atomic_inc(&sched->hw_rq_count);
 528
 529                 if (!full_recovery)
 530                         continue;
 531
 532                 if (fence) {
 533                         r = dma_fence_add_callback(fence, &s_job->cb,
 534                                                    drm_sched_job_done_cb);
 535                         if (r == -ENOENT)
 536                                 drm_sched_job_done(s_job);
 537                         else if (r)
 538                                 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
 539                                           r);
 540                 } else
 541                         drm_sched_job_done(s_job);
 542         }
 543
 544         if (full_recovery) {
 545                 spin_lock(&sched->job_list_lock);
 546                 drm_sched_start_timeout(sched);
 547                 spin_unlock(&sched->job_list_lock);
 548         }
 549
 550         kthread_unpark(sched->thread);
 551 }
 552 EXPORT_SYMBOL(drm_sched_start);
 553
 554 /**
 555  * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
 556  *
 557  * @sched: scheduler instance
 558  *
 559  * Re-submitting jobs was a concept AMD came up as cheap way to implement
 560  * recovery after a job timeout.
 561  *
 562  * This turned out to be not working very well. First of all there are many
 563  * problem with the dma_fence implementation and requirements. Either the
 564  * implementation is risking deadlocks with core memory management or violating
 565  * documented implementation details of the dma_fence object.
 566  *
 567  * Drivers can still save and restore their state for recovery operations, but
 568  * we shouldn't make this a general scheduler feature around the dma_fence
 569  * interface.
 570  */
 571 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
 572 {
 573         struct drm_sched_job *s_job, *tmp;
 574         uint64_t guilty_context;
 575         bool found_guilty = false;
 576         struct dma_fence *fence;
 577
 578         list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 579                 struct drm_sched_fence *s_fence = s_job->s_fence;
 580
 581                 if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 582                         found_guilty = true;
 583                         guilty_context = s_job->s_fence->scheduled.context;
 584                 }
 585
 586                 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 587                         dma_fence_set_error(&s_fence->finished, -ECANCELED);
 588
 589                 fence = sched->ops->run_job(s_job);
 590
 591                 if (IS_ERR_OR_NULL(fence)) {
 592                         if (IS_ERR(fence))
 593                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
 594
 595                         s_job->s_fence->parent = NULL;
 596                 } else {
 597
 598                         s_job->s_fence->parent = dma_fence_get(fence);
 599
 600                         /* Drop for orignal kref_init */
 601                         dma_fence_put(fence);
 602                 }
 603         }
 604 }
 605 EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 606
 607 /**
 608  * drm_sched_job_init - init a scheduler job
 609  * @job: scheduler job to init
 610  * @entity: scheduler entity to use
 611  * @owner: job owner for debugging
 612  *
 613  * Refer to drm_sched_entity_push_job() documentation
 614  * for locking considerations.
 615  *
 616  * Drivers must make sure drm_sched_job_cleanup() if this function returns
 617  * successfully, even when @job is aborted before drm_sched_job_arm() is called.
 618  *
 619  * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
 620  * has died, which can mean that there's no valid runqueue for a @entity.
 621  * This function returns -ENOENT in this case (which probably should be -EIO as
 622  * a more meanigful return value).
 623  *
 624  * Returns 0 for success, negative error code otherwise.
 625  */
 626 int drm_sched_job_init(struct drm_sched_job *job,
 627                        struct drm_sched_entity *entity,
 628                        void *owner)
 629 {
 630         if (!entity->rq)
 631                 return -ENOENT;
 632
 633         job->entity = entity;
 634         job->s_fence = drm_sched_fence_alloc(entity, owner);
 635         if (!job->s_fence)
 636                 return -ENOMEM;
 637
 638         INIT_LIST_HEAD(&job->list);
 639
 640         xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
 641
 642         return 0;
 643 }
 644 EXPORT_SYMBOL(drm_sched_job_init);
 645
 646 /**
 647  * drm_sched_job_arm - arm a scheduler job for execution
 648  * @job: scheduler job to arm
 649  *
 650  * This arms a scheduler job for execution. Specifically it initializes the
 651  * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
 652  * or other places that need to track the completion of this job.
 653  *
 654  * Refer to drm_sched_entity_push_job() documentation for locking
 655  * considerations.
 656  *
 657  * This can only be called if drm_sched_job_init() succeeded.
 658  */
 659 void drm_sched_job_arm(struct drm_sched_job *job)
 660 {
 661         struct drm_gpu_scheduler *sched;
 662         struct drm_sched_entity *entity = job->entity;
 663
 664         BUG_ON(!entity);
 665         drm_sched_entity_select_rq(entity);
 666         sched = entity->rq->sched;
 667
 668         job->sched = sched;
 669         job->s_priority = entity->rq - sched->sched_rq;
 670         job->id = atomic64_inc_return(&sched->job_id_count);
 671
 672         drm_sched_fence_init(job->s_fence, job->entity);
 673 }
 674 EXPORT_SYMBOL(drm_sched_job_arm);
 675
 676 /**
 677  * drm_sched_job_add_dependency - adds the fence as a job dependency
 678  * @job: scheduler job to add the dependencies to
 679  * @fence: the dma_fence to add to the list of dependencies.
 680  *
 681  * Note that @fence is consumed in both the success and error cases.
 682  *
 683  * Returns:
 684  * 0 on success, or an error on failing to expand the array.
 685  */
 686 int drm_sched_job_add_dependency(struct drm_sched_job *job,
 687                                  struct dma_fence *fence)
 688 {
 689         struct dma_fence *entry;
 690         unsigned long index;
 691         u32 id = 0;
 692         int ret;
 693
 694         if (!fence)
 695                 return 0;
 696
 697         /* Deduplicate if we already depend on a fence from the same context.
 698          * This lets the size of the array of deps scale with the number of
 699          * engines involved, rather than the number of BOs.
 700          */
 701         xa_for_each(&job->dependencies, index, entry) {
 702                 if (entry->context != fence->context)
 703                         continue;
 704
 705                 if (dma_fence_is_later(fence, entry)) {
 706                         dma_fence_put(entry);
 707                         xa_store(&job->dependencies, index, fence, GFP_KERNEL);
 708                 } else {
 709                         dma_fence_put(fence);
 710                 }
 711                 return 0;
 712         }
 713
 714         ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
 715         if (ret != 0)
 716                 dma_fence_put(fence);
 717
 718         return ret;
 719 }
 720 EXPORT_SYMBOL(drm_sched_job_add_dependency);
 721
 722 /**
 723  * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
 724  * @job: scheduler job to add the dependencies to
 725  * @resv: the dma_resv object to get the fences from
 726  * @usage: the dma_resv_usage to use to filter the fences
 727  *
 728  * This adds all fences matching the given usage from @resv to @job.
 729  * Must be called with the @resv lock held.
 730  *
 731  * Returns:
 732  * 0 on success, or an error on failing to expand the array.
 733  */
 734 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
 735                                         struct dma_resv *resv,
 736                                         enum dma_resv_usage usage)
 737 {
 738         struct dma_resv_iter cursor;
 739         struct dma_fence *fence;
 740         int ret;
 741
 742         dma_resv_assert_held(resv);
 743
 744         dma_resv_for_each_fence(&cursor, resv, usage, fence) {
 745                 /* Make sure to grab an additional ref on the added fence */
 746                 dma_fence_get(fence);
 747                 ret = drm_sched_job_add_dependency(job, fence);
 748                 if (ret) {
 749                         dma_fence_put(fence);
 750                         return ret;
 751                 }
 752         }
 753         return 0;
 754 }
 755 EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
 756
 757 /**
 758  * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
 759  *   dependencies
 760  * @job: scheduler job to add the dependencies to
 761  * @obj: the gem object to add new dependencies from.
 762  * @write: whether the job might write the object (so we need to depend on
 763  * shared fences in the reservation object).
 764  *
 765  * This should be called after drm_gem_lock_reservations() on your array of
 766  * GEM objects used in the job but before updating the reservations with your
 767  * own fences.
 768  *
 769  * Returns:
 770  * 0 on success, or an error on failing to expand the array.
 771  */
 772 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
 773                                             struct drm_gem_object *obj,
 774                                             bool write)
 775 {
 776         return drm_sched_job_add_resv_dependencies(job, obj->resv,
 777                                                    dma_resv_usage_rw(write));
 778 }
 779 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
 780
 781 /**
 782  * drm_sched_job_cleanup - clean up scheduler job resources
 783  * @job: scheduler job to clean up
 784  *
 785  * Cleans up the resources allocated with drm_sched_job_init().
 786  *
 787  * Drivers should call this from their error unwind code if @job is aborted
 788  * before drm_sched_job_arm() is called.
 789  *
 790  * After that point of no return @job is committed to be executed by the
 791  * scheduler, and this function should be called from the
 792  * &drm_sched_backend_ops.free_job callback.
 793  */
 794 void drm_sched_job_cleanup(struct drm_sched_job *job)
 795 {
 796         struct dma_fence *fence;
 797         unsigned long index;
 798
 799         if (kref_read(&job->s_fence->finished.refcount)) {
 800                 /* drm_sched_job_arm() has been called */
 801                 dma_fence_put(&job->s_fence->finished);
 802         } else {
 803                 /* aborted job before committing to run it */
 804                 drm_sched_fence_free(job->s_fence);
 805         }
 806
 807         job->s_fence = NULL;
 808
 809         xa_for_each(&job->dependencies, index, fence) {
 810                 dma_fence_put(fence);
 811         }
 812         xa_destroy(&job->dependencies);
 813
 814 }
 815 EXPORT_SYMBOL(drm_sched_job_cleanup);
 816
 817 /**
 818  * drm_sched_ready - is the scheduler ready
 819  *
 820  * @sched: scheduler instance
 821  *
 822  * Return true if we can push more jobs to the hw, otherwise false.
 823  */
 824 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 825 {
 826         return atomic_read(&sched->hw_rq_count) <
 827                 sched->hw_submission_limit;
 828 }
 829
 830 /**
 831  * drm_sched_wakeup - Wake up the scheduler when it is ready
 832  *
 833  * @sched: scheduler instance
 834  *
 835  */
 836 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 837 {
 838         if (drm_sched_ready(sched))
 839                 wake_up_interruptible(&sched->wake_up_worker);
 840 }
 841
 842 /**
 843  * drm_sched_select_entity - Select next entity to process
 844  *
 845  * @sched: scheduler instance
 846  *
 847  * Returns the entity to process or NULL if none are found.
 848  */
 849 static struct drm_sched_entity *
 850 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 851 {
 852         struct drm_sched_entity *entity;
 853         int i;
 854
 855         if (!drm_sched_ready(sched))
 856                 return NULL;
 857
 858         /* Kernel run queue has higher priority than normal run queue*/
 859         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 860                 entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
 861                         drm_sched_rq_select_entity_fifo(&sched->sched_rq[i]) :
 862                         drm_sched_rq_select_entity_rr(&sched->sched_rq[i]);
 863                 if (entity)
 864                         break;
 865         }
 866
 867         return entity;
 868 }
 869
 870 /**
 871  * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
 872  *
 873  * @sched: scheduler instance
 874  *
 875  * Returns the next finished job from the pending list (if there is one)
 876  * ready for it to be destroyed.
 877  */
 878 static struct drm_sched_job *
 879 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
 880 {
 881         struct drm_sched_job *job, *next;
 882
 883         spin_lock(&sched->job_list_lock);
 884
 885         job = list_first_entry_or_null(&sched->pending_list,
 886                                        struct drm_sched_job, list);
 887
 888         if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
 889                 /* remove job from pending_list */
 890                 list_del_init(&job->list);
 891
 892                 /* cancel this job's TO timer */
 893                 cancel_delayed_work(&sched->work_tdr);
 894                 /* make the scheduled timestamp more accurate */
 895                 next = list_first_entry_or_null(&sched->pending_list,
 896                                                 typeof(*next), list);
 897
 898                 if (next) {
 899                         next->s_fence->scheduled.timestamp =
 900                                 job->s_fence->finished.timestamp;
 901                         /* start TO timer for next job */
 902                         drm_sched_start_timeout(sched);
 903                 }
 904         } else {
 905                 job = NULL;
 906         }
 907
 908         spin_unlock(&sched->job_list_lock);
 909
 910         return job;
 911 }
 912
 913 /**
 914  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
 915  * @sched_list: list of drm_gpu_schedulers
 916  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
 917  *
 918  * Returns pointer of the sched with the least load or NULL if none of the
 919  * drm_gpu_schedulers are ready
 920  */
 921 struct drm_gpu_scheduler *
 922 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
 923                      unsigned int num_sched_list)
 924 {
 925         struct drm_gpu_scheduler *sched, *picked_sched = NULL;
 926         int i;
 927         unsigned int min_score = UINT_MAX, num_score;
 928
 929         for (i = 0; i < num_sched_list; ++i) {
 930                 sched = sched_list[i];
 931
 932                 if (!sched->ready) {
 933                         DRM_WARN("scheduler %s is not ready, skipping",
 934                                  sched->name);
 935                         continue;
 936                 }
 937
 938                 num_score = atomic_read(sched->score);
 939                 if (num_score < min_score) {
 940                         min_score = num_score;
 941                         picked_sched = sched;
 942                 }
 943         }
 944
 945         return picked_sched;
 946 }
 947 EXPORT_SYMBOL(drm_sched_pick_best);
 948
 949 /**
 950  * drm_sched_blocked - check if the scheduler is blocked
 951  *
 952  * @sched: scheduler instance
 953  *
 954  * Returns true if blocked, otherwise false.
 955  */
 956 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 957 {
 958         if (kthread_should_park()) {
 959                 kthread_parkme();
 960                 return true;
 961         }
 962
 963         return false;
 964 }
 965
 966 /**
 967  * drm_sched_main - main scheduler thread
 968  *
 969  * @param: scheduler instance
 970  *
 971  * Returns 0.
 972  */
 973 static int drm_sched_main(void *param)
 974 {
 975         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
 976         int r;
 977
 978         sched_set_fifo_low(current);
 979
 980         while (!kthread_should_stop()) {
 981                 struct drm_sched_entity *entity = NULL;
 982                 struct drm_sched_fence *s_fence;
 983                 struct drm_sched_job *sched_job;
 984                 struct dma_fence *fence;
 985                 struct drm_sched_job *cleanup_job = NULL;
 986
 987                 wait_event_interruptible(sched->wake_up_worker,
 988                                          (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
 989                                          (!drm_sched_blocked(sched) &&
 990                                           (entity = drm_sched_select_entity(sched))) ||
 991                                          kthread_should_stop());
 992
 993                 if (cleanup_job)
 994                         sched->ops->free_job(cleanup_job);
 995
 996                 if (!entity)
 997                         continue;
 998
 999                 sched_job = drm_sched_entity_pop_job(entity);
1000
1001                 if (!sched_job) {
1002                         complete_all(&entity->entity_idle);
1003                         continue;
1004                 }
1005
1006                 s_fence = sched_job->s_fence;
1007
1008                 atomic_inc(&sched->hw_rq_count);
1009                 drm_sched_job_begin(sched_job);
1010
1011                 trace_drm_run_job(sched_job, entity);
1012                 fence = sched->ops->run_job(sched_job);
1013                 complete_all(&entity->entity_idle);
1014                 drm_sched_fence_scheduled(s_fence);
1015
1016                 if (!IS_ERR_OR_NULL(fence)) {
1017                         s_fence->parent = dma_fence_get(fence);
1018                         /* Drop for original kref_init of the fence */
1019                         dma_fence_put(fence);
1020
1021                         r = dma_fence_add_callback(fence, &sched_job->cb,
1022                                                    drm_sched_job_done_cb);
1023                         if (r == -ENOENT)
1024                                 drm_sched_job_done(sched_job);
1025                         else if (r)
1026                                 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
1027                                           r);
1028                 } else {
1029                         if (IS_ERR(fence))
1030                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
1031
1032                         drm_sched_job_done(sched_job);
1033                 }
1034
1035                 wake_up(&sched->job_scheduled);
1036         }
1037         return 0;
1038 }
1039
1040 /**
1041  * drm_sched_init - Init a gpu scheduler instance
1042  *
1043  * @sched: scheduler instance
1044  * @ops: backend operations for this scheduler
1045  * @hw_submission: number of hw submissions that can be in flight
1046  * @hang_limit: number of times to allow a job to hang before dropping it
1047  * @timeout: timeout value in jiffies for the scheduler
1048  * @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
1049  *              used
1050  * @score: optional score atomic shared with other schedulers
1051  * @name: name used for debugging
1052  * @dev: target &struct device
1053  *
1054  * Return 0 on success, otherwise error code.
1055  */
1056 int drm_sched_init(struct drm_gpu_scheduler *sched,
1057                    const struct drm_sched_backend_ops *ops,
1058                    unsigned hw_submission, unsigned hang_limit,
1059                    long timeout, struct workqueue_struct *timeout_wq,
1060                    atomic_t *score, const char *name, struct device *dev)
1061 {
1062         int i, ret;
1063         sched->ops = ops;
1064         sched->hw_submission_limit = hw_submission;
1065         sched->name = name;
1066         sched->timeout = timeout;
1067         sched->timeout_wq = timeout_wq ? : system_wq;
1068         sched->hang_limit = hang_limit;
1069         sched->score = score ? score : &sched->_score;
1070         sched->dev = dev;
1071         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
1072                 drm_sched_rq_init(sched, &sched->sched_rq[i]);
1073
1074         init_waitqueue_head(&sched->wake_up_worker);
1075         init_waitqueue_head(&sched->job_scheduled);
1076         INIT_LIST_HEAD(&sched->pending_list);
1077         spin_lock_init(&sched->job_list_lock);
1078         atomic_set(&sched->hw_rq_count, 0);
1079         INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1080         atomic_set(&sched->_score, 0);
1081         atomic64_set(&sched->job_id_count, 0);
1082
1083         /* Each scheduler will run on a seperate kernel thread */
1084         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
1085         if (IS_ERR(sched->thread)) {
1086                 ret = PTR_ERR(sched->thread);
1087                 sched->thread = NULL;
1088                 DRM_DEV_ERROR(sched->dev, "Failed to create scheduler for %s.\n", name);
1089                 return ret;
1090         }
1091
1092         sched->ready = true;
1093         return 0;
1094 }
1095 EXPORT_SYMBOL(drm_sched_init);
1096
1097 /**
1098  * drm_sched_fini - Destroy a gpu scheduler
1099  *
1100  * @sched: scheduler instance
1101  *
1102  * Tears down and cleans up the scheduler.
1103  */
1104 void drm_sched_fini(struct drm_gpu_scheduler *sched)
1105 {
1106         struct drm_sched_entity *s_entity;
1107         int i;
1108
1109         if (sched->thread)
1110                 kthread_stop(sched->thread);
1111
1112         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
1113                 struct drm_sched_rq *rq = &sched->sched_rq[i];
1114
1115                 if (!rq)
1116                         continue;
1117
1118                 spin_lock(&rq->lock);
1119                 list_for_each_entry(s_entity, &rq->entities, list)
1120                         /*
1121                          * Prevents reinsertion and marks job_queue as idle,
1122                          * it will removed from rq in drm_sched_entity_fini
1123                          * eventually
1124                          */
1125                         s_entity->stopped = true;
1126                 spin_unlock(&rq->lock);
1127
1128         }
1129
1130         /* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1131         wake_up_all(&sched->job_scheduled);
1132
1133         /* Confirm no work left behind accessing device structures */
1134         cancel_delayed_work_sync(&sched->work_tdr);
1135
1136         sched->ready = false;
1137 }
1138 EXPORT_SYMBOL(drm_sched_fini);
1139
1140 /**
1141  * drm_sched_increase_karma - Update sched_entity guilty flag
1142  *
1143  * @bad: The job guilty of time out
1144  *
1145  * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1146  * limit of the scheduler then the respective sched entity is marked guilty and
1147  * jobs from it will not be scheduled further
1148  */
1149 void drm_sched_increase_karma(struct drm_sched_job *bad)
1150 {
1151         int i;
1152         struct drm_sched_entity *tmp;
1153         struct drm_sched_entity *entity;
1154         struct drm_gpu_scheduler *sched = bad->sched;
1155
1156         /* don't change @bad's karma if it's from KERNEL RQ,
1157          * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1158          * corrupt but keep in mind that kernel jobs always considered good.
1159          */
1160         if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1161                 atomic_inc(&bad->karma);
1162
1163                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
1164                      i++) {
1165                         struct drm_sched_rq *rq = &sched->sched_rq[i];
1166
1167                         spin_lock(&rq->lock);
1168                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1169                                 if (bad->s_fence->scheduled.context ==
1170                                     entity->fence_context) {
1171                                         if (entity->guilty)
1172                                                 atomic_set(entity->guilty, 1);
1173                                         break;
1174                                 }
1175                         }
1176                         spin_unlock(&rq->lock);
1177                         if (&entity->list != &rq->entities)
1178                                 break;
1179                 }
1180         }
1181 }
1182 EXPORT_SYMBOL(drm_sched_increase_karma);