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/drm_syncobj.h>
  57 #include <drm/gpu_scheduler.h>
  58 #include <drm/spsc_queue.h>
  59
  60 #define CREATE_TRACE_POINTS
  61 #include "gpu_scheduler_trace.h"
  62
  63 #define to_drm_sched_job(sched_job)             \
  64                 container_of((sched_job), struct drm_sched_job, queue_node)
  65
  66 int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
  67
  68 /**
  69  * DOC: sched_policy (int)
  70  * Used to override default entities scheduling policy in a run queue.
  71  */
  72 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).");
  73 module_param_named(sched_policy, drm_sched_policy, int, 0444);
  74
  75 static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
  76                                                             const struct rb_node *b)
  77 {
  78         struct drm_sched_entity *ent_a =  rb_entry((a), struct drm_sched_entity, rb_tree_node);
  79         struct drm_sched_entity *ent_b =  rb_entry((b), struct drm_sched_entity, rb_tree_node);
  80
  81         return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
  82 }
  83
  84 static inline void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity)
  85 {
  86         struct drm_sched_rq *rq = entity->rq;
  87
  88         if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
  89                 rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
  90                 RB_CLEAR_NODE(&entity->rb_tree_node);
  91         }
  92 }
  93
  94 void drm_sched_rq_update_fifo(struct drm_sched_entity *entity, ktime_t ts)
  95 {
  96         /*
  97          * Both locks need to be grabbed, one to protect from entity->rq change
  98          * for entity from within concurrent drm_sched_entity_select_rq and the
  99          * other to update the rb tree structure.
 100          */
 101         spin_lock(&entity->rq_lock);
 102         spin_lock(&entity->rq->lock);
 103
 104         drm_sched_rq_remove_fifo_locked(entity);
 105
 106         entity->oldest_job_waiting = ts;
 107
 108         rb_add_cached(&entity->rb_tree_node, &entity->rq->rb_tree_root,
 109                       drm_sched_entity_compare_before);
 110
 111         spin_unlock(&entity->rq->lock);
 112         spin_unlock(&entity->rq_lock);
 113 }
 114
 115 /**
 116  * drm_sched_rq_init - initialize a given run queue struct
 117  *
 118  * @sched: scheduler instance to associate with this run queue
 119  * @rq: scheduler run queue
 120  *
 121  * Initializes a scheduler runqueue.
 122  */
 123 static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
 124                               struct drm_sched_rq *rq)
 125 {
 126         spin_lock_init(&rq->lock);
 127         INIT_LIST_HEAD(&rq->entities);
 128         rq->rb_tree_root = RB_ROOT_CACHED;
 129         rq->current_entity = NULL;
 130         rq->sched = sched;
 131 }
 132
 133 /**
 134  * drm_sched_rq_add_entity - add an entity
 135  *
 136  * @rq: scheduler run queue
 137  * @entity: scheduler entity
 138  *
 139  * Adds a scheduler entity to the run queue.
 140  */
 141 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
 142                              struct drm_sched_entity *entity)
 143 {
 144         if (!list_empty(&entity->list))
 145                 return;
 146
 147         spin_lock(&rq->lock);
 148
 149         atomic_inc(rq->sched->score);
 150         list_add_tail(&entity->list, &rq->entities);
 151
 152         spin_unlock(&rq->lock);
 153 }
 154
 155 /**
 156  * drm_sched_rq_remove_entity - remove an entity
 157  *
 158  * @rq: scheduler run queue
 159  * @entity: scheduler entity
 160  *
 161  * Removes a scheduler entity from the run queue.
 162  */
 163 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 164                                 struct drm_sched_entity *entity)
 165 {
 166         if (list_empty(&entity->list))
 167                 return;
 168
 169         spin_lock(&rq->lock);
 170
 171         atomic_dec(rq->sched->score);
 172         list_del_init(&entity->list);
 173
 174         if (rq->current_entity == entity)
 175                 rq->current_entity = NULL;
 176
 177         if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 178                 drm_sched_rq_remove_fifo_locked(entity);
 179
 180         spin_unlock(&rq->lock);
 181 }
 182
 183 /**
 184  * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
 185  *
 186  * @rq: scheduler run queue to check.
 187  *
 188  * Try to find a ready entity, returns NULL if none found.
 189  */
 190 static struct drm_sched_entity *
 191 drm_sched_rq_select_entity_rr(struct drm_sched_rq *rq)
 192 {
 193         struct drm_sched_entity *entity;
 194
 195         spin_lock(&rq->lock);
 196
 197         entity = rq->current_entity;
 198         if (entity) {
 199                 list_for_each_entry_continue(entity, &rq->entities, list) {
 200                         if (drm_sched_entity_is_ready(entity)) {
 201                                 rq->current_entity = entity;
 202                                 reinit_completion(&entity->entity_idle);
 203                                 spin_unlock(&rq->lock);
 204                                 return entity;
 205                         }
 206                 }
 207         }
 208
 209         list_for_each_entry(entity, &rq->entities, list) {
 210
 211                 if (drm_sched_entity_is_ready(entity)) {
 212                         rq->current_entity = entity;
 213                         reinit_completion(&entity->entity_idle);
 214                         spin_unlock(&rq->lock);
 215                         return entity;
 216                 }
 217
 218                 if (entity == rq->current_entity)
 219                         break;
 220         }
 221
 222         spin_unlock(&rq->lock);
 223
 224         return NULL;
 225 }
 226
 227 /**
 228  * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
 229  *
 230  * @rq: scheduler run queue to check.
 231  *
 232  * Find oldest waiting ready entity, returns NULL if none found.
 233  */
 234 static struct drm_sched_entity *
 235 drm_sched_rq_select_entity_fifo(struct drm_sched_rq *rq)
 236 {
 237         struct rb_node *rb;
 238
 239         spin_lock(&rq->lock);
 240         for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
 241                 struct drm_sched_entity *entity;
 242
 243                 entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
 244                 if (drm_sched_entity_is_ready(entity)) {
 245                         rq->current_entity = entity;
 246                         reinit_completion(&entity->entity_idle);
 247                         break;
 248                 }
 249         }
 250         spin_unlock(&rq->lock);
 251
 252         return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
 253 }
 254
 255 /**
 256  * drm_sched_job_done - complete a job
 257  * @s_job: pointer to the job which is done
 258  *
 259  * Finish the job's fence and wake up the worker thread.
 260  */
 261 static void drm_sched_job_done(struct drm_sched_job *s_job)
 262 {
 263         struct drm_sched_fence *s_fence = s_job->s_fence;
 264         struct drm_gpu_scheduler *sched = s_fence->sched;
 265
 266         atomic_dec(&sched->hw_rq_count);
 267         atomic_dec(sched->score);
 268
 269         trace_drm_sched_process_job(s_fence);
 270
 271         dma_fence_get(&s_fence->finished);
 272         drm_sched_fence_finished(s_fence);
 273         dma_fence_put(&s_fence->finished);
 274         wake_up_interruptible(&sched->wake_up_worker);
 275 }
 276
 277 /**
 278  * drm_sched_job_done_cb - the callback for a done job
 279  * @f: fence
 280  * @cb: fence callbacks
 281  */
 282 static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
 283 {
 284         struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
 285
 286         drm_sched_job_done(s_job);
 287 }
 288
 289 /**
 290  * drm_sched_start_timeout - start timeout for reset worker
 291  *
 292  * @sched: scheduler instance to start the worker for
 293  *
 294  * Start the timeout for the given scheduler.
 295  */
 296 static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
 297 {
 298         if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 299             !list_empty(&sched->pending_list))
 300                 queue_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
 301 }
 302
 303 /**
 304  * drm_sched_fault - immediately start timeout handler
 305  *
 306  * @sched: scheduler where the timeout handling should be started.
 307  *
 308  * Start timeout handling immediately when the driver detects a hardware fault.
 309  */
 310 void drm_sched_fault(struct drm_gpu_scheduler *sched)
 311 {
 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_syncobj_dependency - adds a syncobj's fence as a job dependency
 724  * @job: scheduler job to add the dependencies to
 725  * @file_private: drm file private pointer
 726  * @handle: syncobj handle to lookup
 727  * @point: timeline point
 728  *
 729  * This adds the fence matching the given syncobj to @job.
 730  *
 731  * Returns:
 732  * 0 on success, or an error on failing to expand the array.
 733  */
 734 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
 735                                          struct drm_file *file,
 736                                          u32 handle,
 737                                          u32 point)
 738 {
 739         struct dma_fence *fence;
 740         int ret;
 741
 742         ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
 743         if (ret)
 744                 return ret;
 745
 746         return drm_sched_job_add_dependency(job, fence);
 747 }
 748 EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
 749
 750 /**
 751  * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
 752  * @job: scheduler job to add the dependencies to
 753  * @resv: the dma_resv object to get the fences from
 754  * @usage: the dma_resv_usage to use to filter the fences
 755  *
 756  * This adds all fences matching the given usage from @resv to @job.
 757  * Must be called with the @resv lock held.
 758  *
 759  * Returns:
 760  * 0 on success, or an error on failing to expand the array.
 761  */
 762 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
 763                                         struct dma_resv *resv,
 764                                         enum dma_resv_usage usage)
 765 {
 766         struct dma_resv_iter cursor;
 767         struct dma_fence *fence;
 768         int ret;
 769
 770         dma_resv_assert_held(resv);
 771
 772         dma_resv_for_each_fence(&cursor, resv, usage, fence) {
 773                 /* Make sure to grab an additional ref on the added fence */
 774                 dma_fence_get(fence);
 775                 ret = drm_sched_job_add_dependency(job, fence);
 776                 if (ret) {
 777                         dma_fence_put(fence);
 778                         return ret;
 779                 }
 780         }
 781         return 0;
 782 }
 783 EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
 784
 785 /**
 786  * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
 787  *   dependencies
 788  * @job: scheduler job to add the dependencies to
 789  * @obj: the gem object to add new dependencies from.
 790  * @write: whether the job might write the object (so we need to depend on
 791  * shared fences in the reservation object).
 792  *
 793  * This should be called after drm_gem_lock_reservations() on your array of
 794  * GEM objects used in the job but before updating the reservations with your
 795  * own fences.
 796  *
 797  * Returns:
 798  * 0 on success, or an error on failing to expand the array.
 799  */
 800 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
 801                                             struct drm_gem_object *obj,
 802                                             bool write)
 803 {
 804         return drm_sched_job_add_resv_dependencies(job, obj->resv,
 805                                                    dma_resv_usage_rw(write));
 806 }
 807 EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
 808
 809 /**
 810  * drm_sched_job_cleanup - clean up scheduler job resources
 811  * @job: scheduler job to clean up
 812  *
 813  * Cleans up the resources allocated with drm_sched_job_init().
 814  *
 815  * Drivers should call this from their error unwind code if @job is aborted
 816  * before drm_sched_job_arm() is called.
 817  *
 818  * After that point of no return @job is committed to be executed by the
 819  * scheduler, and this function should be called from the
 820  * &drm_sched_backend_ops.free_job callback.
 821  */
 822 void drm_sched_job_cleanup(struct drm_sched_job *job)
 823 {
 824         struct dma_fence *fence;
 825         unsigned long index;
 826
 827         if (kref_read(&job->s_fence->finished.refcount)) {
 828                 /* drm_sched_job_arm() has been called */
 829                 dma_fence_put(&job->s_fence->finished);
 830         } else {
 831                 /* aborted job before committing to run it */
 832                 drm_sched_fence_free(job->s_fence);
 833         }
 834
 835         job->s_fence = NULL;
 836
 837         xa_for_each(&job->dependencies, index, fence) {
 838                 dma_fence_put(fence);
 839         }
 840         xa_destroy(&job->dependencies);
 841
 842 }
 843 EXPORT_SYMBOL(drm_sched_job_cleanup);
 844
 845 /**
 846  * drm_sched_ready - is the scheduler ready
 847  *
 848  * @sched: scheduler instance
 849  *
 850  * Return true if we can push more jobs to the hw, otherwise false.
 851  */
 852 static bool drm_sched_ready(struct drm_gpu_scheduler *sched)
 853 {
 854         return atomic_read(&sched->hw_rq_count) <
 855                 sched->hw_submission_limit;
 856 }
 857
 858 /**
 859  * drm_sched_wakeup - Wake up the scheduler when it is ready
 860  *
 861  * @sched: scheduler instance
 862  *
 863  */
 864 void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
 865 {
 866         if (drm_sched_ready(sched))
 867                 wake_up_interruptible(&sched->wake_up_worker);
 868 }
 869
 870 /**
 871  * drm_sched_select_entity - Select next entity to process
 872  *
 873  * @sched: scheduler instance
 874  *
 875  * Returns the entity to process or NULL if none are found.
 876  */
 877 static struct drm_sched_entity *
 878 drm_sched_select_entity(struct drm_gpu_scheduler *sched)
 879 {
 880         struct drm_sched_entity *entity;
 881         int i;
 882
 883         if (!drm_sched_ready(sched))
 884                 return NULL;
 885
 886         /* Kernel run queue has higher priority than normal run queue*/
 887         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
 888                 entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
 889                         drm_sched_rq_select_entity_fifo(&sched->sched_rq[i]) :
 890                         drm_sched_rq_select_entity_rr(&sched->sched_rq[i]);
 891                 if (entity)
 892                         break;
 893         }
 894
 895         return entity;
 896 }
 897
 898 /**
 899  * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
 900  *
 901  * @sched: scheduler instance
 902  *
 903  * Returns the next finished job from the pending list (if there is one)
 904  * ready for it to be destroyed.
 905  */
 906 static struct drm_sched_job *
 907 drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched)
 908 {
 909         struct drm_sched_job *job, *next;
 910
 911         spin_lock(&sched->job_list_lock);
 912
 913         job = list_first_entry_or_null(&sched->pending_list,
 914                                        struct drm_sched_job, list);
 915
 916         if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
 917                 /* remove job from pending_list */
 918                 list_del_init(&job->list);
 919
 920                 /* cancel this job's TO timer */
 921                 cancel_delayed_work(&sched->work_tdr);
 922                 /* make the scheduled timestamp more accurate */
 923                 next = list_first_entry_or_null(&sched->pending_list,
 924                                                 typeof(*next), list);
 925
 926                 if (next) {
 927                         next->s_fence->scheduled.timestamp =
 928                                 job->s_fence->finished.timestamp;
 929                         /* start TO timer for next job */
 930                         drm_sched_start_timeout(sched);
 931                 }
 932         } else {
 933                 job = NULL;
 934         }
 935
 936         spin_unlock(&sched->job_list_lock);
 937
 938         if (job) {
 939                 job->entity->elapsed_ns += ktime_to_ns(
 940                         ktime_sub(job->s_fence->finished.timestamp,
 941                                   job->s_fence->scheduled.timestamp));
 942         }
 943
 944         return job;
 945 }
 946
 947 /**
 948  * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
 949  * @sched_list: list of drm_gpu_schedulers
 950  * @num_sched_list: number of drm_gpu_schedulers in the sched_list
 951  *
 952  * Returns pointer of the sched with the least load or NULL if none of the
 953  * drm_gpu_schedulers are ready
 954  */
 955 struct drm_gpu_scheduler *
 956 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
 957                      unsigned int num_sched_list)
 958 {
 959         struct drm_gpu_scheduler *sched, *picked_sched = NULL;
 960         int i;
 961         unsigned int min_score = UINT_MAX, num_score;
 962
 963         for (i = 0; i < num_sched_list; ++i) {
 964                 sched = sched_list[i];
 965
 966                 if (!sched->ready) {
 967                         DRM_WARN("scheduler %s is not ready, skipping",
 968                                  sched->name);
 969                         continue;
 970                 }
 971
 972                 num_score = atomic_read(sched->score);
 973                 if (num_score < min_score) {
 974                         min_score = num_score;
 975                         picked_sched = sched;
 976                 }
 977         }
 978
 979         return picked_sched;
 980 }
 981 EXPORT_SYMBOL(drm_sched_pick_best);
 982
 983 /**
 984  * drm_sched_blocked - check if the scheduler is blocked
 985  *
 986  * @sched: scheduler instance
 987  *
 988  * Returns true if blocked, otherwise false.
 989  */
 990 static bool drm_sched_blocked(struct drm_gpu_scheduler *sched)
 991 {
 992         if (kthread_should_park()) {
 993                 kthread_parkme();
 994                 return true;
 995         }
 996
 997         return false;
 998 }
 999
1000 /**
1001  * drm_sched_main - main scheduler thread
1002  *
1003  * @param: scheduler instance
1004  *
1005  * Returns 0.
1006  */
1007 static int drm_sched_main(void *param)
1008 {
1009         struct drm_gpu_scheduler *sched = (struct drm_gpu_scheduler *)param;
1010         int r;
1011
1012         sched_set_fifo_low(current);
1013
1014         while (!kthread_should_stop()) {
1015                 struct drm_sched_entity *entity = NULL;
1016                 struct drm_sched_fence *s_fence;
1017                 struct drm_sched_job *sched_job;
1018                 struct dma_fence *fence;
1019                 struct drm_sched_job *cleanup_job = NULL;
1020
1021                 wait_event_interruptible(sched->wake_up_worker,
1022                                          (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
1023                                          (!drm_sched_blocked(sched) &&
1024                                           (entity = drm_sched_select_entity(sched))) ||
1025                                          kthread_should_stop());
1026
1027                 if (cleanup_job)
1028                         sched->ops->free_job(cleanup_job);
1029
1030                 if (!entity)
1031                         continue;
1032
1033                 sched_job = drm_sched_entity_pop_job(entity);
1034
1035                 if (!sched_job) {
1036                         complete_all(&entity->entity_idle);
1037                         continue;
1038                 }
1039
1040                 s_fence = sched_job->s_fence;
1041
1042                 atomic_inc(&sched->hw_rq_count);
1043                 drm_sched_job_begin(sched_job);
1044
1045                 trace_drm_run_job(sched_job, entity);
1046                 fence = sched->ops->run_job(sched_job);
1047                 complete_all(&entity->entity_idle);
1048                 drm_sched_fence_scheduled(s_fence);
1049
1050                 if (!IS_ERR_OR_NULL(fence)) {
1051                         s_fence->parent = dma_fence_get(fence);
1052                         /* Drop for original kref_init of the fence */
1053                         dma_fence_put(fence);
1054
1055                         r = dma_fence_add_callback(fence, &sched_job->cb,
1056                                                    drm_sched_job_done_cb);
1057                         if (r == -ENOENT)
1058                                 drm_sched_job_done(sched_job);
1059                         else if (r)
1060                                 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n",
1061                                           r);
1062                 } else {
1063                         if (IS_ERR(fence))
1064                                 dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
1065
1066                         drm_sched_job_done(sched_job);
1067                 }
1068
1069                 wake_up(&sched->job_scheduled);
1070         }
1071         return 0;
1072 }
1073
1074 /**
1075  * drm_sched_init - Init a gpu scheduler instance
1076  *
1077  * @sched: scheduler instance
1078  * @ops: backend operations for this scheduler
1079  * @hw_submission: number of hw submissions that can be in flight
1080  * @hang_limit: number of times to allow a job to hang before dropping it
1081  * @timeout: timeout value in jiffies for the scheduler
1082  * @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
1083  *              used
1084  * @score: optional score atomic shared with other schedulers
1085  * @name: name used for debugging
1086  * @dev: target &struct device
1087  *
1088  * Return 0 on success, otherwise error code.
1089  */
1090 int drm_sched_init(struct drm_gpu_scheduler *sched,
1091                    const struct drm_sched_backend_ops *ops,
1092                    unsigned hw_submission, unsigned hang_limit,
1093                    long timeout, struct workqueue_struct *timeout_wq,
1094                    atomic_t *score, const char *name, struct device *dev)
1095 {
1096         int i, ret;
1097         sched->ops = ops;
1098         sched->hw_submission_limit = hw_submission;
1099         sched->name = name;
1100         sched->timeout = timeout;
1101         sched->timeout_wq = timeout_wq ? : system_wq;
1102         sched->hang_limit = hang_limit;
1103         sched->score = score ? score : &sched->_score;
1104         sched->dev = dev;
1105         for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; i++)
1106                 drm_sched_rq_init(sched, &sched->sched_rq[i]);
1107
1108         init_waitqueue_head(&sched->wake_up_worker);
1109         init_waitqueue_head(&sched->job_scheduled);
1110         INIT_LIST_HEAD(&sched->pending_list);
1111         spin_lock_init(&sched->job_list_lock);
1112         atomic_set(&sched->hw_rq_count, 0);
1113         INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1114         atomic_set(&sched->_score, 0);
1115         atomic64_set(&sched->job_id_count, 0);
1116
1117         /* Each scheduler will run on a seperate kernel thread */
1118         sched->thread = kthread_run(drm_sched_main, sched, sched->name);
1119         if (IS_ERR(sched->thread)) {
1120                 ret = PTR_ERR(sched->thread);
1121                 sched->thread = NULL;
1122                 DRM_DEV_ERROR(sched->dev, "Failed to create scheduler for %s.\n", name);
1123                 return ret;
1124         }
1125
1126         sched->ready = true;
1127         return 0;
1128 }
1129 EXPORT_SYMBOL(drm_sched_init);
1130
1131 /**
1132  * drm_sched_fini - Destroy a gpu scheduler
1133  *
1134  * @sched: scheduler instance
1135  *
1136  * Tears down and cleans up the scheduler.
1137  */
1138 void drm_sched_fini(struct drm_gpu_scheduler *sched)
1139 {
1140         struct drm_sched_entity *s_entity;
1141         int i;
1142
1143         if (sched->thread)
1144                 kthread_stop(sched->thread);
1145
1146         for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
1147                 struct drm_sched_rq *rq = &sched->sched_rq[i];
1148
1149                 if (!rq)
1150                         continue;
1151
1152                 spin_lock(&rq->lock);
1153                 list_for_each_entry(s_entity, &rq->entities, list)
1154                         /*
1155                          * Prevents reinsertion and marks job_queue as idle,
1156                          * it will removed from rq in drm_sched_entity_fini
1157                          * eventually
1158                          */
1159                         s_entity->stopped = true;
1160                 spin_unlock(&rq->lock);
1161
1162         }
1163
1164         /* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1165         wake_up_all(&sched->job_scheduled);
1166
1167         /* Confirm no work left behind accessing device structures */
1168         cancel_delayed_work_sync(&sched->work_tdr);
1169
1170         sched->ready = false;
1171 }
1172 EXPORT_SYMBOL(drm_sched_fini);
1173
1174 /**
1175  * drm_sched_increase_karma - Update sched_entity guilty flag
1176  *
1177  * @bad: The job guilty of time out
1178  *
1179  * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1180  * limit of the scheduler then the respective sched entity is marked guilty and
1181  * jobs from it will not be scheduled further
1182  */
1183 void drm_sched_increase_karma(struct drm_sched_job *bad)
1184 {
1185         int i;
1186         struct drm_sched_entity *tmp;
1187         struct drm_sched_entity *entity;
1188         struct drm_gpu_scheduler *sched = bad->sched;
1189
1190         /* don't change @bad's karma if it's from KERNEL RQ,
1191          * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1192          * corrupt but keep in mind that kernel jobs always considered good.
1193          */
1194         if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1195                 atomic_inc(&bad->karma);
1196
1197                 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
1198                      i++) {
1199                         struct drm_sched_rq *rq = &sched->sched_rq[i];
1200
1201                         spin_lock(&rq->lock);
1202                         list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1203                                 if (bad->s_fence->scheduled.context ==
1204                                     entity->fence_context) {
1205                                         if (entity->guilty)
1206                                                 atomic_set(entity->guilty, 1);
1207                                         break;
1208                                 }
1209                         }
1210                         spin_unlock(&rq->lock);
1211                         if (&entity->list != &rq->entities)
1212                                 break;
1213                 }
1214         }
1215 }
1216 EXPORT_SYMBOL(drm_sched_increase_karma);