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 an entity are always scheduled in the order in which they were pushed.
  45 *
  46 * Note that once a job was taken from the entities queue and pushed to the
  47 * hardware, i.e. the pending queue, the entity must not be referenced anymore
  48 * through the jobs entity pointer.
  49 */
  50
  51/**
  52 * DOC: Flow Control
  53 *
  54 * The DRM GPU scheduler provides a flow control mechanism to regulate the rate
  55 * in which the jobs fetched from scheduler entities are executed.
  56 *
  57 * In this context the &drm_gpu_scheduler keeps track of a driver specified
  58 * credit limit representing the capacity of this scheduler and a credit count;
  59 * every &drm_sched_job carries a driver specified number of credits.
  60 *
  61 * Once a job is executed (but not yet finished), the job's credits contribute
  62 * to the scheduler's credit count until the job is finished. If by executing
  63 * one more job the scheduler's credit count would exceed the scheduler's
  64 * credit limit, the job won't be executed. Instead, the scheduler will wait
  65 * until the credit count has decreased enough to not overflow its credit limit.
  66 * This implies waiting for previously executed jobs.
  67 */
  68
  69#include <linux/export.h>
  70#include <linux/wait.h>
  71#include <linux/sched.h>
  72#include <linux/completion.h>
  73#include <linux/dma-resv.h>
  74#include <uapi/linux/sched/types.h>
  75
  76#include <drm/drm_print.h>
  77#include <drm/drm_gem.h>
  78#include <drm/drm_syncobj.h>
  79#include <drm/gpu_scheduler.h>
  80#include <drm/spsc_queue.h>
  81
  82#include "sched_internal.h"
  83
  84#define CREATE_TRACE_POINTS
  85#include "gpu_scheduler_trace.h"
  86
  87int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
  88
  89/**
  90 * DOC: sched_policy (int)
  91 * Used to override default entities scheduling policy in a run queue.
  92 */
  93MODULE_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).");
  94module_param_named(sched_policy, drm_sched_policy, int, 0444);
  95
  96static u32 drm_sched_available_credits(struct drm_gpu_scheduler *sched)
  97{
  98	u32 credits;
  99
 100	WARN_ON(check_sub_overflow(sched->credit_limit,
 101				   atomic_read(&sched->credit_count),
 102				   &credits));
 103
 104	return credits;
 105}
 106
 107/**
 108 * drm_sched_can_queue -- Can we queue more to the hardware?
 109 * @sched: scheduler instance
 110 * @entity: the scheduler entity
 111 *
 112 * Return true if we can push at least one more job from @entity, false
 113 * otherwise.
 114 */
 115static bool drm_sched_can_queue(struct drm_gpu_scheduler *sched,
 116				struct drm_sched_entity *entity)
 117{
 118	struct drm_sched_job *s_job;
 119
 120	s_job = drm_sched_entity_queue_peek(entity);
 121	if (!s_job)
 122		return false;
 123
 124	/* If a job exceeds the credit limit, truncate it to the credit limit
 125	 * itself to guarantee forward progress.
 126	 */
 127	if (s_job->credits > sched->credit_limit) {
 128		dev_WARN(sched->dev,
 129			 "Jobs may not exceed the credit limit, truncate.\n");
 130		s_job->credits = sched->credit_limit;
 131	}
 132
 133	return drm_sched_available_credits(sched) >= s_job->credits;
 134}
 135
 136static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
 137							    const struct rb_node *b)
 138{
 139	struct drm_sched_entity *ent_a =  rb_entry((a), struct drm_sched_entity, rb_tree_node);
 140	struct drm_sched_entity *ent_b =  rb_entry((b), struct drm_sched_entity, rb_tree_node);
 141
 142	return ktime_before(ent_a->oldest_job_waiting, ent_b->oldest_job_waiting);
 143}
 144
 145static void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity,
 146					    struct drm_sched_rq *rq)
 147{
 148	if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
 149		rb_erase_cached(&entity->rb_tree_node, &rq->rb_tree_root);
 150		RB_CLEAR_NODE(&entity->rb_tree_node);
 151	}
 152}
 153
 154void drm_sched_rq_update_fifo_locked(struct drm_sched_entity *entity,
 155				     struct drm_sched_rq *rq,
 156				     ktime_t ts)
 157{
 158	/*
 159	 * Both locks need to be grabbed, one to protect from entity->rq change
 160	 * for entity from within concurrent drm_sched_entity_select_rq and the
 161	 * other to update the rb tree structure.
 162	 */
 163	lockdep_assert_held(&entity->lock);
 164	lockdep_assert_held(&rq->lock);
 165
 166	drm_sched_rq_remove_fifo_locked(entity, rq);
 167
 168	entity->oldest_job_waiting = ts;
 169
 170	rb_add_cached(&entity->rb_tree_node, &rq->rb_tree_root,
 171		      drm_sched_entity_compare_before);
 172}
 173
 174/**
 175 * drm_sched_rq_init - initialize a given run queue struct
 176 *
 177 * @sched: scheduler instance to associate with this run queue
 178 * @rq: scheduler run queue
 179 *
 180 * Initializes a scheduler runqueue.
 181 */
 182static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
 183			      struct drm_sched_rq *rq)
 184{
 185	spin_lock_init(&rq->lock);
 186	INIT_LIST_HEAD(&rq->entities);
 187	rq->rb_tree_root = RB_ROOT_CACHED;
 188	rq->current_entity = NULL;
 189	rq->sched = sched;
 190}
 191
 192/**
 193 * drm_sched_rq_add_entity - add an entity
 194 *
 195 * @rq: scheduler run queue
 196 * @entity: scheduler entity
 197 *
 198 * Adds a scheduler entity to the run queue.
 199 */
 200void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
 201			     struct drm_sched_entity *entity)
 202{
 203	lockdep_assert_held(&entity->lock);
 204	lockdep_assert_held(&rq->lock);
 205
 206	if (!list_empty(&entity->list))
 207		return;
 208
 209	atomic_inc(rq->sched->score);
 210	list_add_tail(&entity->list, &rq->entities);
 211}
 212
 213/**
 214 * drm_sched_rq_remove_entity - remove an entity
 215 *
 216 * @rq: scheduler run queue
 217 * @entity: scheduler entity
 218 *
 219 * Removes a scheduler entity from the run queue.
 220 */
 221void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 222				struct drm_sched_entity *entity)
 223{
 224	lockdep_assert_held(&entity->lock);
 225
 226	if (list_empty(&entity->list))
 227		return;
 228
 229	spin_lock(&rq->lock);
 230
 231	atomic_dec(rq->sched->score);
 232	list_del_init(&entity->list);
 233
 234	if (rq->current_entity == entity)
 235		rq->current_entity = NULL;
 236
 237	if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 238		drm_sched_rq_remove_fifo_locked(entity, rq);
 239
 240	spin_unlock(&rq->lock);
 241}
 242
 243/**
 244 * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
 245 *
 246 * @sched: the gpu scheduler
 247 * @rq: scheduler run queue to check.
 248 *
 249 * Try to find the next ready entity.
 250 *
 251 * Return an entity if one is found; return an error-pointer (!NULL) if an
 252 * entity was ready, but the scheduler had insufficient credits to accommodate
 253 * its job; return NULL, if no ready entity was found.
 254 */
 255static struct drm_sched_entity *
 256drm_sched_rq_select_entity_rr(struct drm_gpu_scheduler *sched,
 257			      struct drm_sched_rq *rq)
 258{
 259	struct drm_sched_entity *entity;
 260
 261	spin_lock(&rq->lock);
 262
 263	entity = rq->current_entity;
 264	if (entity) {
 265		list_for_each_entry_continue(entity, &rq->entities, list) {
 266			if (drm_sched_entity_is_ready(entity))
 267				goto found;
 268		}
 269	}
 270
 271	list_for_each_entry(entity, &rq->entities, list) {
 272		if (drm_sched_entity_is_ready(entity))
 273			goto found;
 274
 275		if (entity == rq->current_entity)
 276			break;
 277	}
 278
 279	spin_unlock(&rq->lock);
 280
 281	return NULL;
 282
 283found:
 284	if (!drm_sched_can_queue(sched, entity)) {
 285		/*
 286		 * If scheduler cannot take more jobs signal the caller to not
 287		 * consider lower priority queues.
 288		 */
 289		entity = ERR_PTR(-ENOSPC);
 290	} else {
 291		rq->current_entity = entity;
 292		reinit_completion(&entity->entity_idle);
 293	}
 294
 295	spin_unlock(&rq->lock);
 296
 297	return entity;
 298}
 299
 300/**
 301 * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
 302 *
 303 * @sched: the gpu scheduler
 304 * @rq: scheduler run queue to check.
 305 *
 306 * Find oldest waiting ready entity.
 307 *
 308 * Return an entity if one is found; return an error-pointer (!NULL) if an
 309 * entity was ready, but the scheduler had insufficient credits to accommodate
 310 * its job; return NULL, if no ready entity was found.
 311 */
 312static struct drm_sched_entity *
 313drm_sched_rq_select_entity_fifo(struct drm_gpu_scheduler *sched,
 314				struct drm_sched_rq *rq)
 315{
 316	struct rb_node *rb;
 317
 318	spin_lock(&rq->lock);
 319	for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
 320		struct drm_sched_entity *entity;
 321
 322		entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
 323		if (drm_sched_entity_is_ready(entity)) {
 324			/* If we can't queue yet, preserve the current entity in
 325			 * terms of fairness.
 326			 */
 327			if (!drm_sched_can_queue(sched, entity)) {
 328				spin_unlock(&rq->lock);
 329				return ERR_PTR(-ENOSPC);
 330			}
 331
 332			reinit_completion(&entity->entity_idle);
 333			break;
 334		}
 335	}
 336	spin_unlock(&rq->lock);
 337
 338	return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
 339}
 340
 341/**
 342 * drm_sched_run_job_queue - enqueue run-job work
 343 * @sched: scheduler instance
 344 */
 345static void drm_sched_run_job_queue(struct drm_gpu_scheduler *sched)
 346{
 347	if (!READ_ONCE(sched->pause_submit))
 348		queue_work(sched->submit_wq, &sched->work_run_job);
 349}
 350
 351/**
 352 * drm_sched_run_free_queue - enqueue free-job work
 353 * @sched: scheduler instance
 354 */
 355static void drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
 356{
 357	if (!READ_ONCE(sched->pause_submit))
 358		queue_work(sched->submit_wq, &sched->work_free_job);
 359}
 360
 361/**
 362 * drm_sched_job_done - complete a job
 363 * @s_job: pointer to the job which is done
 364 * @result: 0 on success, -ERRNO on error
 365 *
 366 * Finish the job's fence and resubmit the work items.
 367 */
 368static void drm_sched_job_done(struct drm_sched_job *s_job, int result)
 369{
 370	struct drm_sched_fence *s_fence = s_job->s_fence;
 371	struct drm_gpu_scheduler *sched = s_fence->sched;
 372
 373	atomic_sub(s_job->credits, &sched->credit_count);
 374	atomic_dec(sched->score);
 375
 376	trace_drm_sched_job_done(s_fence);
 377
 378	dma_fence_get(&s_fence->finished);
 379	drm_sched_fence_finished(s_fence, result);
 380	dma_fence_put(&s_fence->finished);
 381	drm_sched_run_free_queue(sched);
 382}
 383
 384/**
 385 * drm_sched_job_done_cb - the callback for a done job
 386 * @f: fence
 387 * @cb: fence callbacks
 388 */
 389static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
 390{
 391	struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
 392
 393	drm_sched_job_done(s_job, f->error);
 394}
 395
 396/**
 397 * drm_sched_start_timeout - start timeout for reset worker
 398 *
 399 * @sched: scheduler instance to start the worker for
 400 *
 401 * Start the timeout for the given scheduler.
 402 */
 403static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
 404{
 405	lockdep_assert_held(&sched->job_list_lock);
 406
 407	if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
 408	    !list_empty(&sched->pending_list))
 409		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, sched->timeout);
 410}
 411
 412static void drm_sched_start_timeout_unlocked(struct drm_gpu_scheduler *sched)
 413{
 414	spin_lock(&sched->job_list_lock);
 415	drm_sched_start_timeout(sched);
 416	spin_unlock(&sched->job_list_lock);
 417}
 418
 419/**
 420 * drm_sched_tdr_queue_imm: - immediately start job timeout handler
 421 *
 422 * @sched: scheduler for which the timeout handling should be started.
 423 *
 424 * Start timeout handling immediately for the named scheduler.
 425 */
 426void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched)
 427{
 428	spin_lock(&sched->job_list_lock);
 429	sched->timeout = 0;
 430	drm_sched_start_timeout(sched);
 431	spin_unlock(&sched->job_list_lock);
 432}
 433EXPORT_SYMBOL(drm_sched_tdr_queue_imm);
 434
 435/**
 436 * drm_sched_fault - immediately start timeout handler
 437 *
 438 * @sched: scheduler where the timeout handling should be started.
 439 *
 440 * Start timeout handling immediately when the driver detects a hardware fault.
 441 */
 442void drm_sched_fault(struct drm_gpu_scheduler *sched)
 443{
 444	if (sched->timeout_wq)
 445		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, 0);
 446}
 447EXPORT_SYMBOL(drm_sched_fault);
 448
 449/**
 450 * drm_sched_suspend_timeout - Suspend scheduler job timeout
 451 *
 452 * @sched: scheduler instance for which to suspend the timeout
 453 *
 454 * Suspend the delayed work timeout for the scheduler. This is done by
 455 * modifying the delayed work timeout to an arbitrary large value,
 456 * MAX_SCHEDULE_TIMEOUT in this case.
 457 *
 458 * Returns the timeout remaining
 459 *
 460 */
 461unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
 462{
 463	unsigned long sched_timeout, now = jiffies;
 464
 465	sched_timeout = sched->work_tdr.timer.expires;
 466
 467	/*
 468	 * Modify the timeout to an arbitrarily large value. This also prevents
 469	 * the timeout to be restarted when new submissions arrive
 470	 */
 471	if (mod_delayed_work(sched->timeout_wq, &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
 472			&& time_after(sched_timeout, now))
 473		return sched_timeout - now;
 474	else
 475		return sched->timeout;
 476}
 477EXPORT_SYMBOL(drm_sched_suspend_timeout);
 478
 479/**
 480 * drm_sched_resume_timeout - Resume scheduler job timeout
 481 *
 482 * @sched: scheduler instance for which to resume the timeout
 483 * @remaining: remaining timeout
 484 *
 485 * Resume the delayed work timeout for the scheduler.
 486 */
 487void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
 488		unsigned long remaining)
 489{
 490	spin_lock(&sched->job_list_lock);
 491
 492	if (list_empty(&sched->pending_list))
 493		cancel_delayed_work(&sched->work_tdr);
 494	else
 495		mod_delayed_work(sched->timeout_wq, &sched->work_tdr, remaining);
 496
 497	spin_unlock(&sched->job_list_lock);
 498}
 499EXPORT_SYMBOL(drm_sched_resume_timeout);
 500
 501static void drm_sched_job_begin(struct drm_sched_job *s_job)
 502{
 503	struct drm_gpu_scheduler *sched = s_job->sched;
 504
 505	spin_lock(&sched->job_list_lock);
 506	list_add_tail(&s_job->list, &sched->pending_list);
 507	drm_sched_start_timeout(sched);
 508	spin_unlock(&sched->job_list_lock);
 509}
 510
 511/**
 512 * drm_sched_job_reinsert_on_false_timeout - reinsert the job on a false timeout
 513 * @sched: scheduler instance
 514 * @job: job to be reinserted on the pending list
 515 *
 516 * In the case of a "false timeout" - when a timeout occurs but the GPU isn't
 517 * hung and is making progress, the scheduler must reinsert the job back into
 518 * @sched->pending_list. Otherwise, the job and its resources won't be freed
 519 * through the &struct drm_sched_backend_ops.free_job callback.
 520 *
 521 * This function must be used in "false timeout" cases only.
 522 */
 523static void drm_sched_job_reinsert_on_false_timeout(struct drm_gpu_scheduler *sched,
 524						    struct drm_sched_job *job)
 525{
 526	spin_lock(&sched->job_list_lock);
 527	list_add(&job->list, &sched->pending_list);
 528
 529	/* After reinserting the job, the scheduler enqueues the free-job work
 530	 * again if ready. Otherwise, a signaled job could be added to the
 531	 * pending list, but never freed.
 532	 */
 533	drm_sched_run_free_queue(sched);
 534	spin_unlock(&sched->job_list_lock);
 535}
 536
 537static void drm_sched_job_timedout(struct work_struct *work)
 538{
 539	struct drm_gpu_scheduler *sched;
 540	struct drm_sched_job *job;
 541	enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_RESET;
 542
 543	sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 544
 545	/* Protects against concurrent deletion in drm_sched_get_finished_job */
 546	spin_lock(&sched->job_list_lock);
 547	job = list_first_entry_or_null(&sched->pending_list,
 548				       struct drm_sched_job, list);
 549
 550	if (job) {
 551		/*
 552		 * Remove the bad job so it cannot be freed by a concurrent
 553		 * &struct drm_sched_backend_ops.free_job. It will be
 554		 * reinserted after the scheduler's work items have been
 555		 * cancelled, at which point it's safe.
 556		 */
 557		list_del_init(&job->list);
 558		spin_unlock(&sched->job_list_lock);
 559
 560		status = job->sched->ops->timedout_job(job);
 561
 562		/*
 563		 * Guilty job did complete and hence needs to be manually removed
 564		 * See drm_sched_stop doc.
 565		 */
 566		if (sched->free_guilty) {
 567			job->sched->ops->free_job(job);
 568			sched->free_guilty = false;
 569		}
 570
 571		if (status == DRM_GPU_SCHED_STAT_NO_HANG)
 572			drm_sched_job_reinsert_on_false_timeout(sched, job);
 573	} else {
 574		spin_unlock(&sched->job_list_lock);
 575	}
 576
 577	if (status != DRM_GPU_SCHED_STAT_ENODEV)
 578		drm_sched_start_timeout_unlocked(sched);
 579}
 580
 581/**
 582 * drm_sched_stop - stop the scheduler
 583 *
 584 * @sched: scheduler instance
 585 * @bad: job which caused the time out
 586 *
 587 * Stop the scheduler and also removes and frees all completed jobs.
 588 * Note: bad job will not be freed as it might be used later and so it's
 589 * callers responsibility to release it manually if it's not part of the
 590 * pending list any more.
 591 *
 592 * This function is typically used for reset recovery (see the docu of
 593 * drm_sched_backend_ops.timedout_job() for details). Do not call it for
 594 * scheduler teardown, i.e., before calling drm_sched_fini().
 595 *
 596 * As it's only used for reset recovery, drivers must not call this function
 597 * in their &struct drm_sched_backend_ops.timedout_job callback when they
 598 * skip a reset using &enum drm_gpu_sched_stat.DRM_GPU_SCHED_STAT_NO_HANG.
 599 */
 600void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
 601{
 602	struct drm_sched_job *s_job, *tmp;
 603
 604	drm_sched_wqueue_stop(sched);
 605
 606	/*
 607	 * Reinsert back the bad job here - now it's safe as
 608	 * drm_sched_get_finished_job() cannot race against us and release the
 609	 * bad job at this point - we parked (waited for) any in progress
 610	 * (earlier) cleanups and drm_sched_get_finished_job() will not be
 611	 * called now until the scheduler's work items are submitted again.
 612	 */
 613	if (bad && bad->sched == sched)
 614		/*
 615		 * Add at the head of the queue to reflect it was the earliest
 616		 * job extracted.
 617		 */
 618		list_add(&bad->list, &sched->pending_list);
 619
 620	/*
 621	 * Iterate the job list from later to  earlier one and either deactive
 622	 * their HW callbacks or remove them from pending list if they already
 623	 * signaled.
 624	 * This iteration is thread safe as the scheduler's work items have been
 625	 * cancelled.
 626	 */
 627	list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
 628					 list) {
 629		if (s_job->s_fence->parent &&
 630		    dma_fence_remove_callback(s_job->s_fence->parent,
 631					      &s_job->cb)) {
 632			dma_fence_put(s_job->s_fence->parent);
 633			s_job->s_fence->parent = NULL;
 634			atomic_sub(s_job->credits, &sched->credit_count);
 635		} else {
 636			/*
 637			 * remove job from pending_list.
 638			 * Locking here is for concurrent resume timeout
 639			 */
 640			spin_lock(&sched->job_list_lock);
 641			list_del_init(&s_job->list);
 642			spin_unlock(&sched->job_list_lock);
 643
 644			/*
 645			 * Wait for job's HW fence callback to finish using s_job
 646			 * before releasing it.
 647			 *
 648			 * Job is still alive so fence refcount at least 1
 649			 */
 650			dma_fence_wait(&s_job->s_fence->finished, false);
 651
 652			/*
 653			 * We must keep bad job alive for later use during
 654			 * recovery by some of the drivers but leave a hint
 655			 * that the guilty job must be released.
 656			 */
 657			if (bad != s_job)
 658				sched->ops->free_job(s_job);
 659			else
 660				sched->free_guilty = true;
 661		}
 662	}
 663
 664	/*
 665	 * Stop pending timer in flight as we rearm it in  drm_sched_start. This
 666	 * avoids the pending timeout work in progress to fire right away after
 667	 * this TDR finished and before the newly restarted jobs had a
 668	 * chance to complete.
 669	 */
 670	cancel_delayed_work(&sched->work_tdr);
 671}
 672EXPORT_SYMBOL(drm_sched_stop);
 673
 674/**
 675 * drm_sched_start - recover jobs after a reset
 676 *
 677 * @sched: scheduler instance
 678 * @errno: error to set on the pending fences
 679 *
 680 * This function is typically used for reset recovery (see the docu of
 681 * drm_sched_backend_ops.timedout_job() for details). Do not call it for
 682 * scheduler startup. The scheduler itself is fully operational after
 683 * drm_sched_init() succeeded.
 684 *
 685 * As it's only used for reset recovery, drivers must not call this function
 686 * in their &struct drm_sched_backend_ops.timedout_job callback when they
 687 * skip a reset using &enum drm_gpu_sched_stat.DRM_GPU_SCHED_STAT_NO_HANG.
 688 */
 689void drm_sched_start(struct drm_gpu_scheduler *sched, int errno)
 690{
 691	struct drm_sched_job *s_job, *tmp;
 692
 693	/*
 694	 * Locking the list is not required here as the scheduler's work items
 695	 * are currently not running, so no new jobs are being inserted or
 696	 * removed. Also concurrent GPU recovers can't run in parallel.
 697	 */
 698	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 699		struct dma_fence *fence = s_job->s_fence->parent;
 700
 701		atomic_add(s_job->credits, &sched->credit_count);
 702
 703		if (!fence) {
 704			drm_sched_job_done(s_job, errno ?: -ECANCELED);
 705			continue;
 706		}
 707
 708		if (dma_fence_add_callback(fence, &s_job->cb,
 709					   drm_sched_job_done_cb))
 710			drm_sched_job_done(s_job, fence->error ?: errno);
 711	}
 712
 713	drm_sched_start_timeout_unlocked(sched);
 714	drm_sched_wqueue_start(sched);
 715}
 716EXPORT_SYMBOL(drm_sched_start);
 717
 718/**
 719 * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
 720 *
 721 * @sched: scheduler instance
 722 *
 723 * Re-submitting jobs was a concept AMD came up as cheap way to implement
 724 * recovery after a job timeout.
 725 *
 726 * This turned out to be not working very well. First of all there are many
 727 * problem with the dma_fence implementation and requirements. Either the
 728 * implementation is risking deadlocks with core memory management or violating
 729 * documented implementation details of the dma_fence object.
 730 *
 731 * Drivers can still save and restore their state for recovery operations, but
 732 * we shouldn't make this a general scheduler feature around the dma_fence
 733 * interface.
 734 */
 735void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
 736{
 737	struct drm_sched_job *s_job, *tmp;
 738	uint64_t guilty_context;
 739	bool found_guilty = false;
 740	struct dma_fence *fence;
 741
 742	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
 743		struct drm_sched_fence *s_fence = s_job->s_fence;
 744
 745		if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
 746			found_guilty = true;
 747			guilty_context = s_job->s_fence->scheduled.context;
 748		}
 749
 750		if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 751			dma_fence_set_error(&s_fence->finished, -ECANCELED);
 752
 753		fence = sched->ops->run_job(s_job);
 754
 755		if (IS_ERR_OR_NULL(fence)) {
 756			if (IS_ERR(fence))
 757				dma_fence_set_error(&s_fence->finished, PTR_ERR(fence));
 758
 759			s_job->s_fence->parent = NULL;
 760		} else {
 761
 762			s_job->s_fence->parent = dma_fence_get(fence);
 763
 764			/* Drop for orignal kref_init */
 765			dma_fence_put(fence);
 766		}
 767	}
 768}
 769EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 770
 771/**
 772 * drm_sched_job_init - init a scheduler job
 773 * @job: scheduler job to init
 774 * @entity: scheduler entity to use
 775 * @credits: the number of credits this job contributes to the schedulers
 776 * credit limit
 777 * @owner: job owner for debugging
 778 * @drm_client_id: &struct drm_file.client_id of the owner (used by trace
 779 * events)
 780 *
 781 * Refer to drm_sched_entity_push_job() documentation
 782 * for locking considerations.
 783 *
 784 * Drivers must make sure drm_sched_job_cleanup() if this function returns
 785 * successfully, even when @job is aborted before drm_sched_job_arm() is called.
 786 *
 787 * Note that this function does not assign a valid value to each struct member
 788 * of struct drm_sched_job. Take a look at that struct's documentation to see
 789 * who sets which struct member with what lifetime.
 790 *
 791 * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
 792 * has died, which can mean that there's no valid runqueue for a @entity.
 793 * This function returns -ENOENT in this case (which probably should be -EIO as
 794 * a more meanigful return value).
 795 *
 796 * Returns 0 for success, negative error code otherwise.
 797 */
 798int drm_sched_job_init(struct drm_sched_job *job,
 799		       struct drm_sched_entity *entity,
 800		       u32 credits, void *owner,
 801		       uint64_t drm_client_id)
 802{
 803	if (!entity->rq) {
 804		/* This will most likely be followed by missing frames
 805		 * or worse--a blank screen--leave a trail in the
 806		 * logs, so this can be debugged easier.
 807		 */
 808		dev_err(job->sched->dev, "%s: entity has no rq!\n", __func__);
 809		return -ENOENT;
 810	}
 811
 812	if (unlikely(!credits)) {
 813		pr_err("*ERROR* %s: credits cannot be 0!\n", __func__);
 814		return -EINVAL;
 815	}
 816
 817	/*
 818	 * We don't know for sure how the user has allocated. Thus, zero the
 819	 * struct so that unallowed (i.e., too early) usage of pointers that
 820	 * this function does not set is guaranteed to lead to a NULL pointer
 821	 * exception instead of UB.
 822	 */
 823	memset(job, 0, sizeof(*job));
 824
 825	job->entity = entity;
 826	job->credits = credits;
 827	job->s_fence = drm_sched_fence_alloc(entity, owner, drm_client_id);
 828	if (!job->s_fence)
 829		return -ENOMEM;
 830
 831	INIT_LIST_HEAD(&job->list);
 832
 833	xa_init_flags(&job->dependencies, XA_FLAGS_ALLOC);
 834
 835	return 0;
 836}
 837EXPORT_SYMBOL(drm_sched_job_init);
 838
 839/**
 840 * drm_sched_job_arm - arm a scheduler job for execution
 841 * @job: scheduler job to arm
 842 *
 843 * This arms a scheduler job for execution. Specifically it initializes the
 844 * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
 845 * or other places that need to track the completion of this job. It also
 846 * initializes sequence numbers, which are fundamental for fence ordering.
 847 *
 848 * Refer to drm_sched_entity_push_job() documentation for locking
 849 * considerations.
 850 *
 851 * Once this function was called, you *must* submit @job with
 852 * drm_sched_entity_push_job().
 853 *
 854 * This can only be called if drm_sched_job_init() succeeded.
 855 */
 856void drm_sched_job_arm(struct drm_sched_job *job)
 857{
 858	struct drm_gpu_scheduler *sched;
 859	struct drm_sched_entity *entity = job->entity;
 860
 861	BUG_ON(!entity);
 862	drm_sched_entity_select_rq(entity);
 863	sched = entity->rq->sched;
 864
 865	job->sched = sched;
 866	job->s_priority = entity->priority;
 867
 868	drm_sched_fence_init(job->s_fence, job->entity);
 869}
 870EXPORT_SYMBOL(drm_sched_job_arm);
 871
 872/**
 873 * drm_sched_job_add_dependency - adds the fence as a job dependency
 874 * @job: scheduler job to add the dependencies to
 875 * @fence: the dma_fence to add to the list of dependencies.
 876 *
 877 * Note that @fence is consumed in both the success and error cases.
 878 *
 879 * Returns:
 880 * 0 on success, or an error on failing to expand the array.
 881 */
 882int drm_sched_job_add_dependency(struct drm_sched_job *job,
 883				 struct dma_fence *fence)
 884{
 885	struct dma_fence *entry;
 886	unsigned long index;
 887	u32 id = 0;
 888	int ret;
 889
 890	if (!fence)
 891		return 0;
 892
 893	/* Deduplicate if we already depend on a fence from the same context.
 894	 * This lets the size of the array of deps scale with the number of
 895	 * engines involved, rather than the number of BOs.
 896	 */
 897	xa_for_each(&job->dependencies, index, entry) {
 898		if (entry->context != fence->context)
 899			continue;
 900
 901		if (dma_fence_is_later(fence, entry)) {
 902			dma_fence_put(entry);
 903			xa_store(&job->dependencies, index, fence, GFP_KERNEL);
 904		} else {
 905			dma_fence_put(fence);
 906		}
 907		return 0;
 908	}
 909
 910	ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
 911	if (ret != 0)
 912		dma_fence_put(fence);
 913
 914	return ret;
 915}
 916EXPORT_SYMBOL(drm_sched_job_add_dependency);
 917
 918/**
 919 * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
 920 * @job: scheduler job to add the dependencies to
 921 * @file: drm file private pointer
 922 * @handle: syncobj handle to lookup
 923 * @point: timeline point
 924 *
 925 * This adds the fence matching the given syncobj to @job.
 926 *
 927 * Returns:
 928 * 0 on success, or an error on failing to expand the array.
 929 */
 930int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
 931					 struct drm_file *file,
 932					 u32 handle,
 933					 u32 point)
 934{
 935	struct dma_fence *fence;
 936	int ret;
 937
 938	ret = drm_syncobj_find_fence(file, handle, point, 0, &fence);
 939	if (ret)
 940		return ret;
 941
 942	return drm_sched_job_add_dependency(job, fence);
 943}
 944EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
 945
 946/**
 947 * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
 948 * @job: scheduler job to add the dependencies to
 949 * @resv: the dma_resv object to get the fences from
 950 * @usage: the dma_resv_usage to use to filter the fences
 951 *
 952 * This adds all fences matching the given usage from @resv to @job.
 953 * Must be called with the @resv lock held.
 954 *
 955 * Returns:
 956 * 0 on success, or an error on failing to expand the array.
 957 */
 958int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
 959					struct dma_resv *resv,
 960					enum dma_resv_usage usage)
 961{
 962	struct dma_resv_iter cursor;
 963	struct dma_fence *fence;
 964	int ret;
 965
 966	dma_resv_assert_held(resv);
 967
 968	dma_resv_for_each_fence(&cursor, resv, usage, fence) {
 969		/*
 970		 * As drm_sched_job_add_dependency always consumes the fence
 971		 * reference (even when it fails), and dma_resv_for_each_fence
 972		 * is not obtaining one, we need to grab one before calling.
 973		 */
 974		ret = drm_sched_job_add_dependency(job, dma_fence_get(fence));
 975		if (ret)
 976			return ret;
 977	}
 978	return 0;
 979}
 980EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
 981
 982/**
 983 * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
 984 *   dependencies
 985 * @job: scheduler job to add the dependencies to
 986 * @obj: the gem object to add new dependencies from.
 987 * @write: whether the job might write the object (so we need to depend on
 988 * shared fences in the reservation object).
 989 *
 990 * This should be called after drm_gem_lock_reservations() on your array of
 991 * GEM objects used in the job but before updating the reservations with your
 992 * own fences.
 993 *
 994 * Returns:
 995 * 0 on success, or an error on failing to expand the array.
 996 */
 997int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
 998					    struct drm_gem_object *obj,
 999					    bool write)
1000{
1001	return drm_sched_job_add_resv_dependencies(job, obj->resv,
1002						   dma_resv_usage_rw(write));
1003}
1004EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
1005
1006/**
1007 * drm_sched_job_has_dependency - check whether fence is the job's dependency
1008 * @job: scheduler job to check
1009 * @fence: fence to look for
1010 *
1011 * Returns:
1012 * True if @fence is found within the job's dependencies, or otherwise false.
1013 */
1014bool drm_sched_job_has_dependency(struct drm_sched_job *job,
1015				  struct dma_fence *fence)
1016{
1017	struct dma_fence *f;
1018	unsigned long index;
1019
1020	xa_for_each(&job->dependencies, index, f) {
1021		if (f == fence)
1022			return true;
1023	}
1024
1025	return false;
1026}
1027EXPORT_SYMBOL(drm_sched_job_has_dependency);
1028
1029/**
1030 * drm_sched_job_cleanup - clean up scheduler job resources
1031 * @job: scheduler job to clean up
1032 *
1033 * Cleans up the resources allocated with drm_sched_job_init().
1034 *
1035 * Drivers should call this from their error unwind code if @job is aborted
1036 * before drm_sched_job_arm() is called.
1037 *
1038 * drm_sched_job_arm() is a point of no return since it initializes the fences
1039 * and their sequence number etc. Once that function has been called, you *must*
1040 * submit it with drm_sched_entity_push_job() and cannot simply abort it by
1041 * calling drm_sched_job_cleanup().
1042 *
1043 * This function should be called in the &drm_sched_backend_ops.free_job callback.
1044 */
1045void drm_sched_job_cleanup(struct drm_sched_job *job)
1046{
1047	struct dma_fence *fence;
1048	unsigned long index;
1049
1050	if (kref_read(&job->s_fence->finished.refcount)) {
1051		/* The job has been processed by the scheduler, i.e.,
1052		 * drm_sched_job_arm() and drm_sched_entity_push_job() have
1053		 * been called.
1054		 */
1055		dma_fence_put(&job->s_fence->finished);
1056	} else {
1057		/* The job was aborted before it has been committed to be run;
1058		 * notably, drm_sched_job_arm() has not been called.
1059		 */
1060		drm_sched_fence_free(job->s_fence);
1061	}
1062
1063	job->s_fence = NULL;
1064
1065	xa_for_each(&job->dependencies, index, fence) {
1066		dma_fence_put(fence);
1067	}
1068	xa_destroy(&job->dependencies);
1069
1070}
1071EXPORT_SYMBOL(drm_sched_job_cleanup);
1072
1073/**
1074 * drm_sched_wakeup - Wake up the scheduler if it is ready to queue
1075 * @sched: scheduler instance
1076 *
1077 * Wake up the scheduler if we can queue jobs.
1078 */
1079void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
1080{
1081	drm_sched_run_job_queue(sched);
1082}
1083
1084/**
1085 * drm_sched_select_entity - Select next entity to process
1086 *
1087 * @sched: scheduler instance
1088 *
1089 * Return an entity to process or NULL if none are found.
1090 *
1091 * Note, that we break out of the for-loop when "entity" is non-null, which can
1092 * also be an error-pointer--this assures we don't process lower priority
1093 * run-queues. See comments in the respectively called functions.
1094 */
1095static struct drm_sched_entity *
1096drm_sched_select_entity(struct drm_gpu_scheduler *sched)
1097{
1098	struct drm_sched_entity *entity;
1099	int i;
1100
1101	/* Start with the highest priority.
1102	 */
1103	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1104		entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
1105			drm_sched_rq_select_entity_fifo(sched, sched->sched_rq[i]) :
1106			drm_sched_rq_select_entity_rr(sched, sched->sched_rq[i]);
1107		if (entity)
1108			break;
1109	}
1110
1111	return IS_ERR(entity) ? NULL : entity;
1112}
1113
1114/**
1115 * drm_sched_get_finished_job - fetch the next finished job to be destroyed
1116 *
1117 * @sched: scheduler instance
1118 * @have_more: are there more finished jobs on the list
1119 *
1120 * Informs the caller through @have_more whether there are more finished jobs
1121 * besides the returned one.
1122 *
1123 * Returns the next finished job from the pending list (if there is one)
1124 * ready for it to be destroyed.
1125 */
1126static struct drm_sched_job *
1127drm_sched_get_finished_job(struct drm_gpu_scheduler *sched, bool *have_more)
1128{
1129	struct drm_sched_job *job, *next;
1130
1131	spin_lock(&sched->job_list_lock);
1132
1133	job = list_first_entry_or_null(&sched->pending_list,
1134				       struct drm_sched_job, list);
1135	if (job && dma_fence_is_signaled(&job->s_fence->finished)) {
1136		/* remove job from pending_list */
1137		list_del_init(&job->list);
1138
1139		/* cancel this job's TO timer */
1140		cancel_delayed_work(&sched->work_tdr);
1141
1142		*have_more = false;
1143		next = list_first_entry_or_null(&sched->pending_list,
1144						typeof(*next), list);
1145		if (next) {
1146			/* make the scheduled timestamp more accurate */
1147			if (test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT,
1148				     &next->s_fence->scheduled.flags))
1149				next->s_fence->scheduled.timestamp =
1150					dma_fence_timestamp(&job->s_fence->finished);
1151
1152			*have_more = dma_fence_is_signaled(&next->s_fence->finished);
1153
1154			/* start TO timer for next job */
1155			drm_sched_start_timeout(sched);
1156		}
1157	} else {
1158		job = NULL;
1159	}
1160
1161	spin_unlock(&sched->job_list_lock);
1162
1163	return job;
1164}
1165
1166/**
1167 * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
1168 * @sched_list: list of drm_gpu_schedulers
1169 * @num_sched_list: number of drm_gpu_schedulers in the sched_list
1170 *
1171 * Returns pointer of the sched with the least load or NULL if none of the
1172 * drm_gpu_schedulers are ready
1173 */
1174struct drm_gpu_scheduler *
1175drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
1176		     unsigned int num_sched_list)
1177{
1178	struct drm_gpu_scheduler *sched, *picked_sched = NULL;
1179	int i;
1180	unsigned int min_score = UINT_MAX, num_score;
1181
1182	for (i = 0; i < num_sched_list; ++i) {
1183		sched = sched_list[i];
1184
1185		if (!sched->ready) {
1186			DRM_WARN("scheduler %s is not ready, skipping",
1187				 sched->name);
1188			continue;
1189		}
1190
1191		num_score = atomic_read(sched->score);
1192		if (num_score < min_score) {
1193			min_score = num_score;
1194			picked_sched = sched;
1195		}
1196	}
1197
1198	return picked_sched;
1199}
1200EXPORT_SYMBOL(drm_sched_pick_best);
1201
1202/**
1203 * drm_sched_free_job_work - worker to call free_job
1204 *
1205 * @w: free job work
1206 */
1207static void drm_sched_free_job_work(struct work_struct *w)
1208{
1209	struct drm_gpu_scheduler *sched =
1210		container_of(w, struct drm_gpu_scheduler, work_free_job);
1211	struct drm_sched_job *job;
1212	bool have_more;
1213
1214	job = drm_sched_get_finished_job(sched, &have_more);
1215	if (job) {
1216		sched->ops->free_job(job);
1217		if (have_more)
1218			drm_sched_run_free_queue(sched);
1219	}
1220
1221	drm_sched_run_job_queue(sched);
1222}
1223
1224/**
1225 * drm_sched_run_job_work - worker to call run_job
1226 *
1227 * @w: run job work
1228 */
1229static void drm_sched_run_job_work(struct work_struct *w)
1230{
1231	struct drm_gpu_scheduler *sched =
1232		container_of(w, struct drm_gpu_scheduler, work_run_job);
1233	struct drm_sched_entity *entity;
1234	struct dma_fence *fence;
1235	struct drm_sched_fence *s_fence;
1236	struct drm_sched_job *sched_job;
1237	int r;
1238
1239	/* Find entity with a ready job */
1240	entity = drm_sched_select_entity(sched);
1241	if (!entity) {
1242		/*
1243		 * Either no more work to do, or the next ready job needs more
1244		 * credits than the scheduler has currently available.
1245		 */
1246		return;
1247	}
1248
1249	sched_job = drm_sched_entity_pop_job(entity);
1250	if (!sched_job) {
1251		complete_all(&entity->entity_idle);
1252		drm_sched_run_job_queue(sched);
1253		return;
1254	}
1255
1256	s_fence = sched_job->s_fence;
1257
1258	atomic_add(sched_job->credits, &sched->credit_count);
1259	drm_sched_job_begin(sched_job);
1260
1261	trace_drm_sched_job_run(sched_job, entity);
1262	/*
1263	 * The run_job() callback must by definition return a fence whose
1264	 * refcount has been incremented for the scheduler already.
1265	 */
1266	fence = sched->ops->run_job(sched_job);
1267	complete_all(&entity->entity_idle);
1268	drm_sched_fence_scheduled(s_fence, fence);
1269
1270	if (!IS_ERR_OR_NULL(fence)) {
1271		r = dma_fence_add_callback(fence, &sched_job->cb,
1272					   drm_sched_job_done_cb);
1273		if (r == -ENOENT)
1274			drm_sched_job_done(sched_job, fence->error);
1275		else if (r)
1276			DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
1277
1278		dma_fence_put(fence);
1279	} else {
1280		drm_sched_job_done(sched_job, IS_ERR(fence) ?
1281				   PTR_ERR(fence) : 0);
1282	}
1283
1284	wake_up(&sched->job_scheduled);
1285	drm_sched_run_job_queue(sched);
1286}
1287
1288static struct workqueue_struct *drm_sched_alloc_wq(const char *name)
1289{
1290#if (IS_ENABLED(CONFIG_LOCKDEP))
1291	static struct lockdep_map map = {
1292		.name = "drm_sched_lockdep_map"
1293	};
1294
1295	/*
1296	 * Avoid leaking a lockdep map on each drm sched creation and
1297	 * destruction by using a single lockdep map for all drm sched
1298	 * allocated submit_wq.
1299	 */
1300
1301	return alloc_ordered_workqueue_lockdep_map(name, WQ_MEM_RECLAIM, &map);
1302#else
1303	return alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
1304#endif
1305}
1306
1307/**
1308 * drm_sched_init - Init a gpu scheduler instance
1309 *
1310 * @sched: scheduler instance
1311 * @args: scheduler initialization arguments
1312 *
1313 * Return 0 on success, otherwise error code.
1314 */
1315int drm_sched_init(struct drm_gpu_scheduler *sched, const struct drm_sched_init_args *args)
1316{
1317	int i;
1318
1319	sched->ops = args->ops;
1320	sched->credit_limit = args->credit_limit;
1321	sched->name = args->name;
1322	sched->timeout = args->timeout;
1323	sched->hang_limit = args->hang_limit;
1324	sched->timeout_wq = args->timeout_wq ? args->timeout_wq : system_percpu_wq;
1325	sched->score = args->score ? args->score : &sched->_score;
1326	sched->dev = args->dev;
1327
1328	if (args->num_rqs > DRM_SCHED_PRIORITY_COUNT) {
1329		/* This is a gross violation--tell drivers what the  problem is.
1330		 */
1331		dev_err(sched->dev, "%s: num_rqs cannot be greater than DRM_SCHED_PRIORITY_COUNT\n",
1332			__func__);
1333		return -EINVAL;
1334	} else if (sched->sched_rq) {
1335		/* Not an error, but warn anyway so drivers can
1336		 * fine-tune their DRM calling order, and return all
1337		 * is good.
1338		 */
1339		dev_warn(sched->dev, "%s: scheduler already initialized!\n", __func__);
1340		return 0;
1341	}
1342
1343	if (args->submit_wq) {
1344		sched->submit_wq = args->submit_wq;
1345		sched->own_submit_wq = false;
1346	} else {
1347		sched->submit_wq = drm_sched_alloc_wq(args->name);
1348		if (!sched->submit_wq)
1349			return -ENOMEM;
1350
1351		sched->own_submit_wq = true;
1352	}
1353
1354	sched->sched_rq = kmalloc_array(args->num_rqs, sizeof(*sched->sched_rq),
1355					GFP_KERNEL | __GFP_ZERO);
1356	if (!sched->sched_rq)
1357		goto Out_check_own;
1358	sched->num_rqs = args->num_rqs;
1359	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1360		sched->sched_rq[i] = kzalloc(sizeof(*sched->sched_rq[i]), GFP_KERNEL);
1361		if (!sched->sched_rq[i])
1362			goto Out_unroll;
1363		drm_sched_rq_init(sched, sched->sched_rq[i]);
1364	}
1365
1366	init_waitqueue_head(&sched->job_scheduled);
1367	INIT_LIST_HEAD(&sched->pending_list);
1368	spin_lock_init(&sched->job_list_lock);
1369	atomic_set(&sched->credit_count, 0);
1370	INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1371	INIT_WORK(&sched->work_run_job, drm_sched_run_job_work);
1372	INIT_WORK(&sched->work_free_job, drm_sched_free_job_work);
1373	atomic_set(&sched->_score, 0);
1374	atomic64_set(&sched->job_id_count, 0);
1375	sched->pause_submit = false;
1376
1377	sched->ready = true;
1378	return 0;
1379Out_unroll:
1380	for (--i ; i >= DRM_SCHED_PRIORITY_KERNEL; i--)
1381		kfree(sched->sched_rq[i]);
1382
1383	kfree(sched->sched_rq);
1384	sched->sched_rq = NULL;
1385Out_check_own:
1386	if (sched->own_submit_wq)
1387		destroy_workqueue(sched->submit_wq);
1388	dev_err(sched->dev, "%s: Failed to setup GPU scheduler--out of memory\n", __func__);
1389	return -ENOMEM;
1390}
1391EXPORT_SYMBOL(drm_sched_init);
1392
1393static void drm_sched_cancel_remaining_jobs(struct drm_gpu_scheduler *sched)
1394{
1395	struct drm_sched_job *job, *tmp;
1396
1397	/* All other accessors are stopped. No locking necessary. */
1398	list_for_each_entry_safe_reverse(job, tmp, &sched->pending_list, list) {
1399		sched->ops->cancel_job(job);
1400		list_del(&job->list);
1401		sched->ops->free_job(job);
1402	}
1403}
1404
1405/**
1406 * drm_sched_fini - Destroy a gpu scheduler
1407 *
1408 * @sched: scheduler instance
1409 *
1410 * Tears down and cleans up the scheduler.
1411 *
1412 * This stops submission of new jobs to the hardware through &struct
1413 * drm_sched_backend_ops.run_job. If &struct drm_sched_backend_ops.cancel_job
1414 * is implemented, all jobs will be canceled through it and afterwards cleaned
1415 * up through &struct drm_sched_backend_ops.free_job. If cancel_job is not
1416 * implemented, memory could leak.
1417 */
1418void drm_sched_fini(struct drm_gpu_scheduler *sched)
1419{
1420	struct drm_sched_entity *s_entity;
1421	int i;
1422
1423	drm_sched_wqueue_stop(sched);
1424
1425	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1426		struct drm_sched_rq *rq = sched->sched_rq[i];
1427
1428		spin_lock(&rq->lock);
1429		list_for_each_entry(s_entity, &rq->entities, list) {
1430			/*
1431			 * Prevents reinsertion and marks job_queue as idle,
1432			 * it will be removed from the rq in drm_sched_entity_fini()
1433			 * eventually
1434			 *
1435			 * FIXME:
1436			 * This lacks the proper spin_lock(&s_entity->lock) and
1437			 * is, therefore, a race condition. Most notably, it
1438			 * can race with drm_sched_entity_push_job(). The lock
1439			 * cannot be taken here, however, because this would
1440			 * lead to lock inversion -> deadlock.
1441			 *
1442			 * The best solution probably is to enforce the life
1443			 * time rule of all entities having to be torn down
1444			 * before their scheduler. Then, however, locking could
1445			 * be dropped alltogether from this function.
1446			 *
1447			 * For now, this remains a potential race in all
1448			 * drivers that keep entities alive for longer than
1449			 * the scheduler.
1450			 *
1451			 * The READ_ONCE() is there to make the lockless read
1452			 * (warning about the lockless write below) slightly
1453			 * less broken...
1454			 */
1455			if (!READ_ONCE(s_entity->stopped))
1456				dev_warn(sched->dev, "Tearing down scheduler with active entities!\n");
1457			s_entity->stopped = true;
1458		}
1459		spin_unlock(&rq->lock);
1460		kfree(sched->sched_rq[i]);
1461	}
1462
1463	/* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1464	wake_up_all(&sched->job_scheduled);
1465
1466	/* Confirm no work left behind accessing device structures */
1467	cancel_delayed_work_sync(&sched->work_tdr);
1468
1469	/* Avoid memory leaks if supported by the driver. */
1470	if (sched->ops->cancel_job)
1471		drm_sched_cancel_remaining_jobs(sched);
1472
1473	if (sched->own_submit_wq)
1474		destroy_workqueue(sched->submit_wq);
1475	sched->ready = false;
1476	kfree(sched->sched_rq);
1477	sched->sched_rq = NULL;
1478
1479	if (!list_empty(&sched->pending_list))
1480		dev_warn(sched->dev, "Tearing down scheduler while jobs are pending!\n");
1481}
1482EXPORT_SYMBOL(drm_sched_fini);
1483
1484/**
1485 * drm_sched_increase_karma - Update sched_entity guilty flag
1486 *
1487 * @bad: The job guilty of time out
1488 *
1489 * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1490 * limit of the scheduler then the respective sched entity is marked guilty and
1491 * jobs from it will not be scheduled further
1492 */
1493void drm_sched_increase_karma(struct drm_sched_job *bad)
1494{
1495	int i;
1496	struct drm_sched_entity *tmp;
1497	struct drm_sched_entity *entity;
1498	struct drm_gpu_scheduler *sched = bad->sched;
1499
1500	/* don't change @bad's karma if it's from KERNEL RQ,
1501	 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1502	 * corrupt but keep in mind that kernel jobs always considered good.
1503	 */
1504	if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1505		atomic_inc(&bad->karma);
1506
1507		for (i = DRM_SCHED_PRIORITY_HIGH; i < sched->num_rqs; i++) {
1508			struct drm_sched_rq *rq = sched->sched_rq[i];
1509
1510			spin_lock(&rq->lock);
1511			list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1512				if (bad->s_fence->scheduled.context ==
1513				    entity->fence_context) {
1514					if (entity->guilty)
1515						atomic_set(entity->guilty, 1);
1516					break;
1517				}
1518			}
1519			spin_unlock(&rq->lock);
1520			if (&entity->list != &rq->entities)
1521				break;
1522		}
1523	}
1524}
1525EXPORT_SYMBOL(drm_sched_increase_karma);
1526
1527/**
1528 * drm_sched_wqueue_ready - Is the scheduler ready for submission
1529 *
1530 * @sched: scheduler instance
1531 *
1532 * Returns true if submission is ready
1533 */
1534bool drm_sched_wqueue_ready(struct drm_gpu_scheduler *sched)
1535{
1536	return sched->ready;
1537}
1538EXPORT_SYMBOL(drm_sched_wqueue_ready);
1539
1540/**
1541 * drm_sched_wqueue_stop - stop scheduler submission
1542 * @sched: scheduler instance
1543 *
1544 * Stops the scheduler from pulling new jobs from entities. It also stops
1545 * freeing jobs automatically through drm_sched_backend_ops.free_job().
1546 */
1547void drm_sched_wqueue_stop(struct drm_gpu_scheduler *sched)
1548{
1549	WRITE_ONCE(sched->pause_submit, true);
1550	cancel_work_sync(&sched->work_run_job);
1551	cancel_work_sync(&sched->work_free_job);
1552}
1553EXPORT_SYMBOL(drm_sched_wqueue_stop);
1554
1555/**
1556 * drm_sched_wqueue_start - start scheduler submission
1557 * @sched: scheduler instance
1558 *
1559 * Restarts the scheduler after drm_sched_wqueue_stop() has stopped it.
1560 *
1561 * This function is not necessary for 'conventional' startup. The scheduler is
1562 * fully operational after drm_sched_init() succeeded.
1563 */
1564void drm_sched_wqueue_start(struct drm_gpu_scheduler *sched)
1565{
1566	WRITE_ONCE(sched->pause_submit, false);
1567	queue_work(sched->submit_wq, &sched->work_run_job);
1568	queue_work(sched->submit_wq, &sched->work_free_job);
1569}
1570EXPORT_SYMBOL(drm_sched_wqueue_start);