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