Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   1/*
   2 * Copyright © 2008-2015 Intel Corporation
   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 (including the next
  12 * paragraph) shall be included in all copies or substantial portions of the
  13 * Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 */
  24
  25#include <linux/prefetch.h>
  26#include <linux/dma-fence-array.h>
  27#include <linux/sched.h>
  28#include <linux/sched/clock.h>
  29#include <linux/sched/signal.h>
  30
  31#include "i915_drv.h"
  32
  33static const char *i915_fence_get_driver_name(struct dma_fence *fence)
  34{
  35	return "i915";
  36}
  37
  38static const char *i915_fence_get_timeline_name(struct dma_fence *fence)
  39{
  40	/*
  41	 * The timeline struct (as part of the ppgtt underneath a context)
  42	 * may be freed when the request is no longer in use by the GPU.
  43	 * We could extend the life of a context to beyond that of all
  44	 * fences, possibly keeping the hw resource around indefinitely,
  45	 * or we just give them a false name. Since
  46	 * dma_fence_ops.get_timeline_name is a debug feature, the occasional
  47	 * lie seems justifiable.
  48	 */
  49	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
  50		return "signaled";
  51
  52	return to_request(fence)->timeline->common->name;
  53}
  54
  55static bool i915_fence_signaled(struct dma_fence *fence)
  56{
  57	return i915_request_completed(to_request(fence));
  58}
  59
  60static bool i915_fence_enable_signaling(struct dma_fence *fence)
  61{
  62	if (i915_fence_signaled(fence))
  63		return false;
  64
  65	intel_engine_enable_signaling(to_request(fence), true);
  66	return !i915_fence_signaled(fence);
  67}
  68
  69static signed long i915_fence_wait(struct dma_fence *fence,
  70				   bool interruptible,
  71				   signed long timeout)
  72{
  73	return i915_request_wait(to_request(fence), interruptible, timeout);
  74}
  75
  76static void i915_fence_release(struct dma_fence *fence)
  77{
  78	struct i915_request *rq = to_request(fence);
  79
  80	/*
  81	 * The request is put onto a RCU freelist (i.e. the address
  82	 * is immediately reused), mark the fences as being freed now.
  83	 * Otherwise the debugobjects for the fences are only marked as
  84	 * freed when the slab cache itself is freed, and so we would get
  85	 * caught trying to reuse dead objects.
  86	 */
  87	i915_sw_fence_fini(&rq->submit);
  88
  89	kmem_cache_free(rq->i915->requests, rq);
  90}
  91
  92const struct dma_fence_ops i915_fence_ops = {
  93	.get_driver_name = i915_fence_get_driver_name,
  94	.get_timeline_name = i915_fence_get_timeline_name,
  95	.enable_signaling = i915_fence_enable_signaling,
  96	.signaled = i915_fence_signaled,
  97	.wait = i915_fence_wait,
  98	.release = i915_fence_release,
  99};
 100
 101static inline void
 102i915_request_remove_from_client(struct i915_request *request)
 103{
 104	struct drm_i915_file_private *file_priv;
 105
 106	file_priv = request->file_priv;
 107	if (!file_priv)
 108		return;
 109
 110	spin_lock(&file_priv->mm.lock);
 111	if (request->file_priv) {
 112		list_del(&request->client_link);
 113		request->file_priv = NULL;
 114	}
 115	spin_unlock(&file_priv->mm.lock);
 116}
 117
 118static struct i915_dependency *
 119i915_dependency_alloc(struct drm_i915_private *i915)
 120{
 121	return kmem_cache_alloc(i915->dependencies, GFP_KERNEL);
 122}
 123
 124static void
 125i915_dependency_free(struct drm_i915_private *i915,
 126		     struct i915_dependency *dep)
 127{
 128	kmem_cache_free(i915->dependencies, dep);
 129}
 130
 131static void
 132__i915_priotree_add_dependency(struct i915_priotree *pt,
 133			       struct i915_priotree *signal,
 134			       struct i915_dependency *dep,
 135			       unsigned long flags)
 136{
 137	INIT_LIST_HEAD(&dep->dfs_link);
 138	list_add(&dep->wait_link, &signal->waiters_list);
 139	list_add(&dep->signal_link, &pt->signalers_list);
 140	dep->signaler = signal;
 141	dep->flags = flags;
 142}
 143
 144static int
 145i915_priotree_add_dependency(struct drm_i915_private *i915,
 146			     struct i915_priotree *pt,
 147			     struct i915_priotree *signal)
 148{
 149	struct i915_dependency *dep;
 150
 151	dep = i915_dependency_alloc(i915);
 152	if (!dep)
 153		return -ENOMEM;
 154
 155	__i915_priotree_add_dependency(pt, signal, dep, I915_DEPENDENCY_ALLOC);
 156	return 0;
 157}
 158
 159static void
 160i915_priotree_fini(struct drm_i915_private *i915, struct i915_priotree *pt)
 161{
 162	struct i915_dependency *dep, *next;
 163
 164	GEM_BUG_ON(!list_empty(&pt->link));
 165
 166	/*
 167	 * Everyone we depended upon (the fences we wait to be signaled)
 168	 * should retire before us and remove themselves from our list.
 169	 * However, retirement is run independently on each timeline and
 170	 * so we may be called out-of-order.
 171	 */
 172	list_for_each_entry_safe(dep, next, &pt->signalers_list, signal_link) {
 173		GEM_BUG_ON(!i915_priotree_signaled(dep->signaler));
 174		GEM_BUG_ON(!list_empty(&dep->dfs_link));
 175
 176		list_del(&dep->wait_link);
 177		if (dep->flags & I915_DEPENDENCY_ALLOC)
 178			i915_dependency_free(i915, dep);
 179	}
 180
 181	/* Remove ourselves from everyone who depends upon us */
 182	list_for_each_entry_safe(dep, next, &pt->waiters_list, wait_link) {
 183		GEM_BUG_ON(dep->signaler != pt);
 184		GEM_BUG_ON(!list_empty(&dep->dfs_link));
 185
 186		list_del(&dep->signal_link);
 187		if (dep->flags & I915_DEPENDENCY_ALLOC)
 188			i915_dependency_free(i915, dep);
 189	}
 190}
 191
 192static void
 193i915_priotree_init(struct i915_priotree *pt)
 194{
 195	INIT_LIST_HEAD(&pt->signalers_list);
 196	INIT_LIST_HEAD(&pt->waiters_list);
 197	INIT_LIST_HEAD(&pt->link);
 198	pt->priority = I915_PRIORITY_INVALID;
 199}
 200
 201static int reset_all_global_seqno(struct drm_i915_private *i915, u32 seqno)
 202{
 203	struct intel_engine_cs *engine;
 204	enum intel_engine_id id;
 205	int ret;
 206
 207	/* Carefully retire all requests without writing to the rings */
 208	ret = i915_gem_wait_for_idle(i915,
 209				     I915_WAIT_INTERRUPTIBLE |
 210				     I915_WAIT_LOCKED);
 211	if (ret)
 212		return ret;
 213
 214	/* If the seqno wraps around, we need to clear the breadcrumb rbtree */
 215	for_each_engine(engine, i915, id) {
 216		struct i915_gem_timeline *timeline;
 217		struct intel_timeline *tl = engine->timeline;
 218
 219		if (!i915_seqno_passed(seqno, tl->seqno)) {
 220			/* Flush any waiters before we reuse the seqno */
 221			intel_engine_disarm_breadcrumbs(engine);
 222			GEM_BUG_ON(!list_empty(&engine->breadcrumbs.signals));
 223		}
 224
 225		/* Check we are idle before we fiddle with hw state! */
 226		GEM_BUG_ON(!intel_engine_is_idle(engine));
 227		GEM_BUG_ON(i915_gem_active_isset(&engine->timeline->last_request));
 228
 229		/* Finally reset hw state */
 230		intel_engine_init_global_seqno(engine, seqno);
 231		tl->seqno = seqno;
 232
 233		list_for_each_entry(timeline, &i915->gt.timelines, link)
 234			memset(timeline->engine[id].global_sync, 0,
 235			       sizeof(timeline->engine[id].global_sync));
 236	}
 237
 238	return 0;
 239}
 240
 241int i915_gem_set_global_seqno(struct drm_device *dev, u32 seqno)
 242{
 243	struct drm_i915_private *i915 = to_i915(dev);
 244
 245	lockdep_assert_held(&i915->drm.struct_mutex);
 246
 247	if (seqno == 0)
 248		return -EINVAL;
 249
 250	/* HWS page needs to be set less than what we will inject to ring */
 251	return reset_all_global_seqno(i915, seqno - 1);
 252}
 253
 254static void mark_busy(struct drm_i915_private *i915)
 255{
 256	if (i915->gt.awake)
 257		return;
 258
 259	GEM_BUG_ON(!i915->gt.active_requests);
 260
 261	intel_runtime_pm_get_noresume(i915);
 262
 263	/*
 264	 * It seems that the DMC likes to transition between the DC states a lot
 265	 * when there are no connected displays (no active power domains) during
 266	 * command submission.
 267	 *
 268	 * This activity has negative impact on the performance of the chip with
 269	 * huge latencies observed in the interrupt handler and elsewhere.
 270	 *
 271	 * Work around it by grabbing a GT IRQ power domain whilst there is any
 272	 * GT activity, preventing any DC state transitions.
 273	 */
 274	intel_display_power_get(i915, POWER_DOMAIN_GT_IRQ);
 275
 276	i915->gt.awake = true;
 277	if (unlikely(++i915->gt.epoch == 0)) /* keep 0 as invalid */
 278		i915->gt.epoch = 1;
 279
 280	intel_enable_gt_powersave(i915);
 281	i915_update_gfx_val(i915);
 282	if (INTEL_GEN(i915) >= 6)
 283		gen6_rps_busy(i915);
 284	i915_pmu_gt_unparked(i915);
 285
 286	intel_engines_unpark(i915);
 287
 288	i915_queue_hangcheck(i915);
 289
 290	queue_delayed_work(i915->wq,
 291			   &i915->gt.retire_work,
 292			   round_jiffies_up_relative(HZ));
 293}
 294
 295static int reserve_engine(struct intel_engine_cs *engine)
 296{
 297	struct drm_i915_private *i915 = engine->i915;
 298	u32 active = ++engine->timeline->inflight_seqnos;
 299	u32 seqno = engine->timeline->seqno;
 300	int ret;
 301
 302	/* Reservation is fine until we need to wrap around */
 303	if (unlikely(add_overflows(seqno, active))) {
 304		ret = reset_all_global_seqno(i915, 0);
 305		if (ret) {
 306			engine->timeline->inflight_seqnos--;
 307			return ret;
 308		}
 309	}
 310
 311	if (!i915->gt.active_requests++)
 312		mark_busy(i915);
 313
 314	return 0;
 315}
 316
 317static void unreserve_engine(struct intel_engine_cs *engine)
 318{
 319	struct drm_i915_private *i915 = engine->i915;
 320
 321	if (!--i915->gt.active_requests) {
 322		/* Cancel the mark_busy() from our reserve_engine() */
 323		GEM_BUG_ON(!i915->gt.awake);
 324		mod_delayed_work(i915->wq,
 325				 &i915->gt.idle_work,
 326				 msecs_to_jiffies(100));
 327	}
 328
 329	GEM_BUG_ON(!engine->timeline->inflight_seqnos);
 330	engine->timeline->inflight_seqnos--;
 331}
 332
 333void i915_gem_retire_noop(struct i915_gem_active *active,
 334			  struct i915_request *request)
 335{
 336	/* Space left intentionally blank */
 337}
 338
 339static void advance_ring(struct i915_request *request)
 340{
 341	unsigned int tail;
 342
 343	/*
 344	 * We know the GPU must have read the request to have
 345	 * sent us the seqno + interrupt, so use the position
 346	 * of tail of the request to update the last known position
 347	 * of the GPU head.
 348	 *
 349	 * Note this requires that we are always called in request
 350	 * completion order.
 351	 */
 352	if (list_is_last(&request->ring_link, &request->ring->request_list)) {
 353		/*
 354		 * We may race here with execlists resubmitting this request
 355		 * as we retire it. The resubmission will move the ring->tail
 356		 * forwards (to request->wa_tail). We either read the
 357		 * current value that was written to hw, or the value that
 358		 * is just about to be. Either works, if we miss the last two
 359		 * noops - they are safe to be replayed on a reset.
 360		 */
 361		tail = READ_ONCE(request->ring->tail);
 362	} else {
 363		tail = request->postfix;
 364	}
 365	list_del(&request->ring_link);
 366
 367	request->ring->head = tail;
 368}
 369
 370static void free_capture_list(struct i915_request *request)
 371{
 372	struct i915_capture_list *capture;
 373
 374	capture = request->capture_list;
 375	while (capture) {
 376		struct i915_capture_list *next = capture->next;
 377
 378		kfree(capture);
 379		capture = next;
 380	}
 381}
 382
 383static void i915_request_retire(struct i915_request *request)
 384{
 385	struct intel_engine_cs *engine = request->engine;
 386	struct i915_gem_active *active, *next;
 387
 388	lockdep_assert_held(&request->i915->drm.struct_mutex);
 389	GEM_BUG_ON(!i915_sw_fence_signaled(&request->submit));
 390	GEM_BUG_ON(!i915_request_completed(request));
 391	GEM_BUG_ON(!request->i915->gt.active_requests);
 392
 393	trace_i915_request_retire(request);
 394
 395	spin_lock_irq(&engine->timeline->lock);
 396	list_del_init(&request->link);
 397	spin_unlock_irq(&engine->timeline->lock);
 398
 399	unreserve_engine(request->engine);
 400	advance_ring(request);
 401
 402	free_capture_list(request);
 403
 404	/*
 405	 * Walk through the active list, calling retire on each. This allows
 406	 * objects to track their GPU activity and mark themselves as idle
 407	 * when their *last* active request is completed (updating state
 408	 * tracking lists for eviction, active references for GEM, etc).
 409	 *
 410	 * As the ->retire() may free the node, we decouple it first and
 411	 * pass along the auxiliary information (to avoid dereferencing
 412	 * the node after the callback).
 413	 */
 414	list_for_each_entry_safe(active, next, &request->active_list, link) {
 415		/*
 416		 * In microbenchmarks or focusing upon time inside the kernel,
 417		 * we may spend an inordinate amount of time simply handling
 418		 * the retirement of requests and processing their callbacks.
 419		 * Of which, this loop itself is particularly hot due to the
 420		 * cache misses when jumping around the list of i915_gem_active.
 421		 * So we try to keep this loop as streamlined as possible and
 422		 * also prefetch the next i915_gem_active to try and hide
 423		 * the likely cache miss.
 424		 */
 425		prefetchw(next);
 426
 427		INIT_LIST_HEAD(&active->link);
 428		RCU_INIT_POINTER(active->request, NULL);
 429
 430		active->retire(active, request);
 431	}
 432
 433	i915_request_remove_from_client(request);
 434
 435	/* Retirement decays the ban score as it is a sign of ctx progress */
 436	atomic_dec_if_positive(&request->ctx->ban_score);
 437
 438	/*
 439	 * The backing object for the context is done after switching to the
 440	 * *next* context. Therefore we cannot retire the previous context until
 441	 * the next context has already started running. However, since we
 442	 * cannot take the required locks at i915_request_submit() we
 443	 * defer the unpinning of the active context to now, retirement of
 444	 * the subsequent request.
 445	 */
 446	if (engine->last_retired_context)
 447		engine->context_unpin(engine, engine->last_retired_context);
 448	engine->last_retired_context = request->ctx;
 449
 450	spin_lock_irq(&request->lock);
 451	if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &request->fence.flags))
 452		dma_fence_signal_locked(&request->fence);
 453	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
 454		intel_engine_cancel_signaling(request);
 455	if (request->waitboost) {
 456		GEM_BUG_ON(!atomic_read(&request->i915->gt_pm.rps.num_waiters));
 457		atomic_dec(&request->i915->gt_pm.rps.num_waiters);
 458	}
 459	spin_unlock_irq(&request->lock);
 460
 461	i915_priotree_fini(request->i915, &request->priotree);
 462	i915_request_put(request);
 463}
 464
 465void i915_request_retire_upto(struct i915_request *rq)
 466{
 467	struct intel_engine_cs *engine = rq->engine;
 468	struct i915_request *tmp;
 469
 470	lockdep_assert_held(&rq->i915->drm.struct_mutex);
 471	GEM_BUG_ON(!i915_request_completed(rq));
 472
 473	if (list_empty(&rq->link))
 474		return;
 475
 476	do {
 477		tmp = list_first_entry(&engine->timeline->requests,
 478				       typeof(*tmp), link);
 479
 480		i915_request_retire(tmp);
 481	} while (tmp != rq);
 482}
 483
 484static u32 timeline_get_seqno(struct intel_timeline *tl)
 485{
 486	return ++tl->seqno;
 487}
 488
 489void __i915_request_submit(struct i915_request *request)
 490{
 491	struct intel_engine_cs *engine = request->engine;
 492	struct intel_timeline *timeline;
 493	u32 seqno;
 494
 495	GEM_BUG_ON(!irqs_disabled());
 496	lockdep_assert_held(&engine->timeline->lock);
 497
 498	/* Transfer from per-context onto the global per-engine timeline */
 499	timeline = engine->timeline;
 500	GEM_BUG_ON(timeline == request->timeline);
 501	GEM_BUG_ON(request->global_seqno);
 502
 503	seqno = timeline_get_seqno(timeline);
 504	GEM_BUG_ON(!seqno);
 505	GEM_BUG_ON(i915_seqno_passed(intel_engine_get_seqno(engine), seqno));
 506
 507	/* We may be recursing from the signal callback of another i915 fence */
 508	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
 509	request->global_seqno = seqno;
 510	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
 511		intel_engine_enable_signaling(request, false);
 512	spin_unlock(&request->lock);
 513
 514	engine->emit_breadcrumb(request,
 515				request->ring->vaddr + request->postfix);
 516
 517	spin_lock(&request->timeline->lock);
 518	list_move_tail(&request->link, &timeline->requests);
 519	spin_unlock(&request->timeline->lock);
 520
 521	trace_i915_request_execute(request);
 522
 523	wake_up_all(&request->execute);
 524}
 525
 526void i915_request_submit(struct i915_request *request)
 527{
 528	struct intel_engine_cs *engine = request->engine;
 529	unsigned long flags;
 530
 531	/* Will be called from irq-context when using foreign fences. */
 532	spin_lock_irqsave(&engine->timeline->lock, flags);
 533
 534	__i915_request_submit(request);
 535
 536	spin_unlock_irqrestore(&engine->timeline->lock, flags);
 537}
 538
 539void __i915_request_unsubmit(struct i915_request *request)
 540{
 541	struct intel_engine_cs *engine = request->engine;
 542	struct intel_timeline *timeline;
 543
 544	GEM_BUG_ON(!irqs_disabled());
 545	lockdep_assert_held(&engine->timeline->lock);
 546
 547	/*
 548	 * Only unwind in reverse order, required so that the per-context list
 549	 * is kept in seqno/ring order.
 550	 */
 551	GEM_BUG_ON(!request->global_seqno);
 552	GEM_BUG_ON(request->global_seqno != engine->timeline->seqno);
 553	GEM_BUG_ON(i915_seqno_passed(intel_engine_get_seqno(engine),
 554				     request->global_seqno));
 555	engine->timeline->seqno--;
 556
 557	/* We may be recursing from the signal callback of another i915 fence */
 558	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
 559	request->global_seqno = 0;
 560	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
 561		intel_engine_cancel_signaling(request);
 562	spin_unlock(&request->lock);
 563
 564	/* Transfer back from the global per-engine timeline to per-context */
 565	timeline = request->timeline;
 566	GEM_BUG_ON(timeline == engine->timeline);
 567
 568	spin_lock(&timeline->lock);
 569	list_move(&request->link, &timeline->requests);
 570	spin_unlock(&timeline->lock);
 571
 572	/*
 573	 * We don't need to wake_up any waiters on request->execute, they
 574	 * will get woken by any other event or us re-adding this request
 575	 * to the engine timeline (__i915_request_submit()). The waiters
 576	 * should be quite adapt at finding that the request now has a new
 577	 * global_seqno to the one they went to sleep on.
 578	 */
 579}
 580
 581void i915_request_unsubmit(struct i915_request *request)
 582{
 583	struct intel_engine_cs *engine = request->engine;
 584	unsigned long flags;
 585
 586	/* Will be called from irq-context when using foreign fences. */
 587	spin_lock_irqsave(&engine->timeline->lock, flags);
 588
 589	__i915_request_unsubmit(request);
 590
 591	spin_unlock_irqrestore(&engine->timeline->lock, flags);
 592}
 593
 594static int __i915_sw_fence_call
 595submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
 596{
 597	struct i915_request *request =
 598		container_of(fence, typeof(*request), submit);
 599
 600	switch (state) {
 601	case FENCE_COMPLETE:
 602		trace_i915_request_submit(request);
 603		/*
 604		 * We need to serialize use of the submit_request() callback
 605		 * with its hotplugging performed during an emergency
 606		 * i915_gem_set_wedged().  We use the RCU mechanism to mark the
 607		 * critical section in order to force i915_gem_set_wedged() to
 608		 * wait until the submit_request() is completed before
 609		 * proceeding.
 610		 */
 611		rcu_read_lock();
 612		request->engine->submit_request(request);
 613		rcu_read_unlock();
 614		break;
 615
 616	case FENCE_FREE:
 617		i915_request_put(request);
 618		break;
 619	}
 620
 621	return NOTIFY_DONE;
 622}
 623
 624/**
 625 * i915_request_alloc - allocate a request structure
 626 *
 627 * @engine: engine that we wish to issue the request on.
 628 * @ctx: context that the request will be associated with.
 629 *
 630 * Returns a pointer to the allocated request if successful,
 631 * or an error code if not.
 632 */
 633struct i915_request *
 634i915_request_alloc(struct intel_engine_cs *engine, struct i915_gem_context *ctx)
 635{
 636	struct drm_i915_private *i915 = engine->i915;
 637	struct i915_request *rq;
 638	struct intel_ring *ring;
 639	int ret;
 640
 641	lockdep_assert_held(&i915->drm.struct_mutex);
 642
 643	/*
 644	 * Preempt contexts are reserved for exclusive use to inject a
 645	 * preemption context switch. They are never to be used for any trivial
 646	 * request!
 647	 */
 648	GEM_BUG_ON(ctx == i915->preempt_context);
 649
 650	/*
 651	 * ABI: Before userspace accesses the GPU (e.g. execbuffer), report
 652	 * EIO if the GPU is already wedged.
 653	 */
 654	if (i915_terminally_wedged(&i915->gpu_error))
 655		return ERR_PTR(-EIO);
 656
 657	/*
 658	 * Pinning the contexts may generate requests in order to acquire
 659	 * GGTT space, so do this first before we reserve a seqno for
 660	 * ourselves.
 661	 */
 662	ring = engine->context_pin(engine, ctx);
 663	if (IS_ERR(ring))
 664		return ERR_CAST(ring);
 665	GEM_BUG_ON(!ring);
 666
 667	ret = reserve_engine(engine);
 668	if (ret)
 669		goto err_unpin;
 670
 671	ret = intel_ring_wait_for_space(ring, MIN_SPACE_FOR_ADD_REQUEST);
 672	if (ret)
 673		goto err_unreserve;
 674
 675	/* Move the oldest request to the slab-cache (if not in use!) */
 676	rq = list_first_entry_or_null(&engine->timeline->requests,
 677				      typeof(*rq), link);
 678	if (rq && i915_request_completed(rq))
 679		i915_request_retire(rq);
 680
 681	/*
 682	 * Beware: Dragons be flying overhead.
 683	 *
 684	 * We use RCU to look up requests in flight. The lookups may
 685	 * race with the request being allocated from the slab freelist.
 686	 * That is the request we are writing to here, may be in the process
 687	 * of being read by __i915_gem_active_get_rcu(). As such,
 688	 * we have to be very careful when overwriting the contents. During
 689	 * the RCU lookup, we change chase the request->engine pointer,
 690	 * read the request->global_seqno and increment the reference count.
 691	 *
 692	 * The reference count is incremented atomically. If it is zero,
 693	 * the lookup knows the request is unallocated and complete. Otherwise,
 694	 * it is either still in use, or has been reallocated and reset
 695	 * with dma_fence_init(). This increment is safe for release as we
 696	 * check that the request we have a reference to and matches the active
 697	 * request.
 698	 *
 699	 * Before we increment the refcount, we chase the request->engine
 700	 * pointer. We must not call kmem_cache_zalloc() or else we set
 701	 * that pointer to NULL and cause a crash during the lookup. If
 702	 * we see the request is completed (based on the value of the
 703	 * old engine and seqno), the lookup is complete and reports NULL.
 704	 * If we decide the request is not completed (new engine or seqno),
 705	 * then we grab a reference and double check that it is still the
 706	 * active request - which it won't be and restart the lookup.
 707	 *
 708	 * Do not use kmem_cache_zalloc() here!
 709	 */
 710	rq = kmem_cache_alloc(i915->requests,
 711			      GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
 712	if (unlikely(!rq)) {
 713		/* Ratelimit ourselves to prevent oom from malicious clients */
 714		ret = i915_gem_wait_for_idle(i915,
 715					     I915_WAIT_LOCKED |
 716					     I915_WAIT_INTERRUPTIBLE);
 717		if (ret)
 718			goto err_unreserve;
 719
 720		/*
 721		 * We've forced the client to stall and catch up with whatever
 722		 * backlog there might have been. As we are assuming that we
 723		 * caused the mempressure, now is an opportune time to
 724		 * recover as much memory from the request pool as is possible.
 725		 * Having already penalized the client to stall, we spend
 726		 * a little extra time to re-optimise page allocation.
 727		 */
 728		kmem_cache_shrink(i915->requests);
 729		rcu_barrier(); /* Recover the TYPESAFE_BY_RCU pages */
 730
 731		rq = kmem_cache_alloc(i915->requests, GFP_KERNEL);
 732		if (!rq) {
 733			ret = -ENOMEM;
 734			goto err_unreserve;
 735		}
 736	}
 737
 738	rq->timeline = i915_gem_context_lookup_timeline(ctx, engine);
 739	GEM_BUG_ON(rq->timeline == engine->timeline);
 740
 741	spin_lock_init(&rq->lock);
 742	dma_fence_init(&rq->fence,
 743		       &i915_fence_ops,
 744		       &rq->lock,
 745		       rq->timeline->fence_context,
 746		       timeline_get_seqno(rq->timeline));
 747
 748	/* We bump the ref for the fence chain */
 749	i915_sw_fence_init(&i915_request_get(rq)->submit, submit_notify);
 750	init_waitqueue_head(&rq->execute);
 751
 752	i915_priotree_init(&rq->priotree);
 753
 754	INIT_LIST_HEAD(&rq->active_list);
 755	rq->i915 = i915;
 756	rq->engine = engine;
 757	rq->ctx = ctx;
 758	rq->ring = ring;
 759
 760	/* No zalloc, must clear what we need by hand */
 761	rq->global_seqno = 0;
 762	rq->signaling.wait.seqno = 0;
 763	rq->file_priv = NULL;
 764	rq->batch = NULL;
 765	rq->capture_list = NULL;
 766	rq->waitboost = false;
 767
 768	/*
 769	 * Reserve space in the ring buffer for all the commands required to
 770	 * eventually emit this request. This is to guarantee that the
 771	 * i915_request_add() call can't fail. Note that the reserve may need
 772	 * to be redone if the request is not actually submitted straight
 773	 * away, e.g. because a GPU scheduler has deferred it.
 774	 */
 775	rq->reserved_space = MIN_SPACE_FOR_ADD_REQUEST;
 776	GEM_BUG_ON(rq->reserved_space < engine->emit_breadcrumb_sz);
 777
 778	/*
 779	 * Record the position of the start of the request so that
 780	 * should we detect the updated seqno part-way through the
 781	 * GPU processing the request, we never over-estimate the
 782	 * position of the head.
 783	 */
 784	rq->head = rq->ring->emit;
 785
 786	/* Unconditionally invalidate GPU caches and TLBs. */
 787	ret = engine->emit_flush(rq, EMIT_INVALIDATE);
 788	if (ret)
 789		goto err_unwind;
 790
 791	ret = engine->request_alloc(rq);
 792	if (ret)
 793		goto err_unwind;
 794
 795	/* Check that we didn't interrupt ourselves with a new request */
 796	GEM_BUG_ON(rq->timeline->seqno != rq->fence.seqno);
 797	return rq;
 798
 799err_unwind:
 800	rq->ring->emit = rq->head;
 801
 802	/* Make sure we didn't add ourselves to external state before freeing */
 803	GEM_BUG_ON(!list_empty(&rq->active_list));
 804	GEM_BUG_ON(!list_empty(&rq->priotree.signalers_list));
 805	GEM_BUG_ON(!list_empty(&rq->priotree.waiters_list));
 806
 807	kmem_cache_free(i915->requests, rq);
 808err_unreserve:
 809	unreserve_engine(engine);
 810err_unpin:
 811	engine->context_unpin(engine, ctx);
 812	return ERR_PTR(ret);
 813}
 814
 815static int
 816i915_request_await_request(struct i915_request *to, struct i915_request *from)
 817{
 818	int ret;
 819
 820	GEM_BUG_ON(to == from);
 821	GEM_BUG_ON(to->timeline == from->timeline);
 822
 823	if (i915_request_completed(from))
 824		return 0;
 825
 826	if (to->engine->schedule) {
 827		ret = i915_priotree_add_dependency(to->i915,
 828						   &to->priotree,
 829						   &from->priotree);
 830		if (ret < 0)
 831			return ret;
 832	}
 833
 834	if (to->engine == from->engine) {
 835		ret = i915_sw_fence_await_sw_fence_gfp(&to->submit,
 836						       &from->submit,
 837						       I915_FENCE_GFP);
 838		return ret < 0 ? ret : 0;
 839	}
 840
 841	if (to->engine->semaphore.sync_to) {
 842		u32 seqno;
 843
 844		GEM_BUG_ON(!from->engine->semaphore.signal);
 845
 846		seqno = i915_request_global_seqno(from);
 847		if (!seqno)
 848			goto await_dma_fence;
 849
 850		if (seqno <= to->timeline->global_sync[from->engine->id])
 851			return 0;
 852
 853		trace_i915_gem_ring_sync_to(to, from);
 854		ret = to->engine->semaphore.sync_to(to, from);
 855		if (ret)
 856			return ret;
 857
 858		to->timeline->global_sync[from->engine->id] = seqno;
 859		return 0;
 860	}
 861
 862await_dma_fence:
 863	ret = i915_sw_fence_await_dma_fence(&to->submit,
 864					    &from->fence, 0,
 865					    I915_FENCE_GFP);
 866	return ret < 0 ? ret : 0;
 867}
 868
 869int
 870i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence)
 871{
 872	struct dma_fence **child = &fence;
 873	unsigned int nchild = 1;
 874	int ret;
 875
 876	/*
 877	 * Note that if the fence-array was created in signal-on-any mode,
 878	 * we should *not* decompose it into its individual fences. However,
 879	 * we don't currently store which mode the fence-array is operating
 880	 * in. Fortunately, the only user of signal-on-any is private to
 881	 * amdgpu and we should not see any incoming fence-array from
 882	 * sync-file being in signal-on-any mode.
 883	 */
 884	if (dma_fence_is_array(fence)) {
 885		struct dma_fence_array *array = to_dma_fence_array(fence);
 886
 887		child = array->fences;
 888		nchild = array->num_fences;
 889		GEM_BUG_ON(!nchild);
 890	}
 891
 892	do {
 893		fence = *child++;
 894		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
 895			continue;
 896
 897		/*
 898		 * Requests on the same timeline are explicitly ordered, along
 899		 * with their dependencies, by i915_request_add() which ensures
 900		 * that requests are submitted in-order through each ring.
 901		 */
 902		if (fence->context == rq->fence.context)
 903			continue;
 904
 905		/* Squash repeated waits to the same timelines */
 906		if (fence->context != rq->i915->mm.unordered_timeline &&
 907		    intel_timeline_sync_is_later(rq->timeline, fence))
 908			continue;
 909
 910		if (dma_fence_is_i915(fence))
 911			ret = i915_request_await_request(rq, to_request(fence));
 912		else
 913			ret = i915_sw_fence_await_dma_fence(&rq->submit, fence,
 914							    I915_FENCE_TIMEOUT,
 915							    I915_FENCE_GFP);
 916		if (ret < 0)
 917			return ret;
 918
 919		/* Record the latest fence used against each timeline */
 920		if (fence->context != rq->i915->mm.unordered_timeline)
 921			intel_timeline_sync_set(rq->timeline, fence);
 922	} while (--nchild);
 923
 924	return 0;
 925}
 926
 927/**
 928 * i915_request_await_object - set this request to (async) wait upon a bo
 929 * @to: request we are wishing to use
 930 * @obj: object which may be in use on another ring.
 931 * @write: whether the wait is on behalf of a writer
 932 *
 933 * This code is meant to abstract object synchronization with the GPU.
 934 * Conceptually we serialise writes between engines inside the GPU.
 935 * We only allow one engine to write into a buffer at any time, but
 936 * multiple readers. To ensure each has a coherent view of memory, we must:
 937 *
 938 * - If there is an outstanding write request to the object, the new
 939 *   request must wait for it to complete (either CPU or in hw, requests
 940 *   on the same ring will be naturally ordered).
 941 *
 942 * - If we are a write request (pending_write_domain is set), the new
 943 *   request must wait for outstanding read requests to complete.
 944 *
 945 * Returns 0 if successful, else propagates up the lower layer error.
 946 */
 947int
 948i915_request_await_object(struct i915_request *to,
 949			  struct drm_i915_gem_object *obj,
 950			  bool write)
 951{
 952	struct dma_fence *excl;
 953	int ret = 0;
 954
 955	if (write) {
 956		struct dma_fence **shared;
 957		unsigned int count, i;
 958
 959		ret = reservation_object_get_fences_rcu(obj->resv,
 960							&excl, &count, &shared);
 961		if (ret)
 962			return ret;
 963
 964		for (i = 0; i < count; i++) {
 965			ret = i915_request_await_dma_fence(to, shared[i]);
 966			if (ret)
 967				break;
 968
 969			dma_fence_put(shared[i]);
 970		}
 971
 972		for (; i < count; i++)
 973			dma_fence_put(shared[i]);
 974		kfree(shared);
 975	} else {
 976		excl = reservation_object_get_excl_rcu(obj->resv);
 977	}
 978
 979	if (excl) {
 980		if (ret == 0)
 981			ret = i915_request_await_dma_fence(to, excl);
 982
 983		dma_fence_put(excl);
 984	}
 985
 986	return ret;
 987}
 988
 989/*
 990 * NB: This function is not allowed to fail. Doing so would mean the the
 991 * request is not being tracked for completion but the work itself is
 992 * going to happen on the hardware. This would be a Bad Thing(tm).
 993 */
 994void __i915_request_add(struct i915_request *request, bool flush_caches)
 995{
 996	struct intel_engine_cs *engine = request->engine;
 997	struct intel_ring *ring = request->ring;
 998	struct intel_timeline *timeline = request->timeline;
 999	struct i915_request *prev;
1000	u32 *cs;
1001	int err;
1002
1003	lockdep_assert_held(&request->i915->drm.struct_mutex);
1004	trace_i915_request_add(request);
1005
1006	/*
1007	 * Make sure that no request gazumped us - if it was allocated after
1008	 * our i915_request_alloc() and called __i915_request_add() before
1009	 * us, the timeline will hold its seqno which is later than ours.
1010	 */
1011	GEM_BUG_ON(timeline->seqno != request->fence.seqno);
1012
1013	/*
1014	 * To ensure that this call will not fail, space for its emissions
1015	 * should already have been reserved in the ring buffer. Let the ring
1016	 * know that it is time to use that space up.
1017	 */
1018	request->reserved_space = 0;
1019
1020	/*
1021	 * Emit any outstanding flushes - execbuf can fail to emit the flush
1022	 * after having emitted the batchbuffer command. Hence we need to fix
1023	 * things up similar to emitting the lazy request. The difference here
1024	 * is that the flush _must_ happen before the next request, no matter
1025	 * what.
1026	 */
1027	if (flush_caches) {
1028		err = engine->emit_flush(request, EMIT_FLUSH);
1029
1030		/* Not allowed to fail! */
1031		WARN(err, "engine->emit_flush() failed: %d!\n", err);
1032	}
1033
1034	/*
1035	 * Record the position of the start of the breadcrumb so that
1036	 * should we detect the updated seqno part-way through the
1037	 * GPU processing the request, we never over-estimate the
1038	 * position of the ring's HEAD.
1039	 */
1040	cs = intel_ring_begin(request, engine->emit_breadcrumb_sz);
1041	GEM_BUG_ON(IS_ERR(cs));
1042	request->postfix = intel_ring_offset(request, cs);
1043
1044	/*
1045	 * Seal the request and mark it as pending execution. Note that
1046	 * we may inspect this state, without holding any locks, during
1047	 * hangcheck. Hence we apply the barrier to ensure that we do not
1048	 * see a more recent value in the hws than we are tracking.
1049	 */
1050
1051	prev = i915_gem_active_raw(&timeline->last_request,
1052				   &request->i915->drm.struct_mutex);
1053	if (prev && !i915_request_completed(prev)) {
1054		i915_sw_fence_await_sw_fence(&request->submit, &prev->submit,
1055					     &request->submitq);
1056		if (engine->schedule)
1057			__i915_priotree_add_dependency(&request->priotree,
1058						       &prev->priotree,
1059						       &request->dep,
1060						       0);
1061	}
1062
1063	spin_lock_irq(&timeline->lock);
1064	list_add_tail(&request->link, &timeline->requests);
1065	spin_unlock_irq(&timeline->lock);
1066
1067	GEM_BUG_ON(timeline->seqno != request->fence.seqno);
1068	i915_gem_active_set(&timeline->last_request, request);
1069
1070	list_add_tail(&request->ring_link, &ring->request_list);
1071	request->emitted_jiffies = jiffies;
1072
1073	/*
1074	 * Let the backend know a new request has arrived that may need
1075	 * to adjust the existing execution schedule due to a high priority
1076	 * request - i.e. we may want to preempt the current request in order
1077	 * to run a high priority dependency chain *before* we can execute this
1078	 * request.
1079	 *
1080	 * This is called before the request is ready to run so that we can
1081	 * decide whether to preempt the entire chain so that it is ready to
1082	 * run at the earliest possible convenience.
1083	 */
1084	rcu_read_lock();
1085	if (engine->schedule)
1086		engine->schedule(request, request->ctx->priority);
1087	rcu_read_unlock();
1088
1089	local_bh_disable();
1090	i915_sw_fence_commit(&request->submit);
1091	local_bh_enable(); /* Kick the execlists tasklet if just scheduled */
1092
1093	/*
1094	 * In typical scenarios, we do not expect the previous request on
1095	 * the timeline to be still tracked by timeline->last_request if it
1096	 * has been completed. If the completed request is still here, that
1097	 * implies that request retirement is a long way behind submission,
1098	 * suggesting that we haven't been retiring frequently enough from
1099	 * the combination of retire-before-alloc, waiters and the background
1100	 * retirement worker. So if the last request on this timeline was
1101	 * already completed, do a catch up pass, flushing the retirement queue
1102	 * up to this client. Since we have now moved the heaviest operations
1103	 * during retirement onto secondary workers, such as freeing objects
1104	 * or contexts, retiring a bunch of requests is mostly list management
1105	 * (and cache misses), and so we should not be overly penalizing this
1106	 * client by performing excess work, though we may still performing
1107	 * work on behalf of others -- but instead we should benefit from
1108	 * improved resource management. (Well, that's the theory at least.)
1109	 */
1110	if (prev && i915_request_completed(prev))
1111		i915_request_retire_upto(prev);
1112}
1113
1114static unsigned long local_clock_us(unsigned int *cpu)
1115{
1116	unsigned long t;
1117
1118	/*
1119	 * Cheaply and approximately convert from nanoseconds to microseconds.
1120	 * The result and subsequent calculations are also defined in the same
1121	 * approximate microseconds units. The principal source of timing
1122	 * error here is from the simple truncation.
1123	 *
1124	 * Note that local_clock() is only defined wrt to the current CPU;
1125	 * the comparisons are no longer valid if we switch CPUs. Instead of
1126	 * blocking preemption for the entire busywait, we can detect the CPU
1127	 * switch and use that as indicator of system load and a reason to
1128	 * stop busywaiting, see busywait_stop().
1129	 */
1130	*cpu = get_cpu();
1131	t = local_clock() >> 10;
1132	put_cpu();
1133
1134	return t;
1135}
1136
1137static bool busywait_stop(unsigned long timeout, unsigned int cpu)
1138{
1139	unsigned int this_cpu;
1140
1141	if (time_after(local_clock_us(&this_cpu), timeout))
1142		return true;
1143
1144	return this_cpu != cpu;
1145}
1146
1147static bool __i915_spin_request(const struct i915_request *rq,
1148				u32 seqno, int state, unsigned long timeout_us)
1149{
1150	struct intel_engine_cs *engine = rq->engine;
1151	unsigned int irq, cpu;
1152
1153	GEM_BUG_ON(!seqno);
1154
1155	/*
1156	 * Only wait for the request if we know it is likely to complete.
1157	 *
1158	 * We don't track the timestamps around requests, nor the average
1159	 * request length, so we do not have a good indicator that this
1160	 * request will complete within the timeout. What we do know is the
1161	 * order in which requests are executed by the engine and so we can
1162	 * tell if the request has started. If the request hasn't started yet,
1163	 * it is a fair assumption that it will not complete within our
1164	 * relatively short timeout.
1165	 */
1166	if (!i915_seqno_passed(intel_engine_get_seqno(engine), seqno - 1))
1167		return false;
1168
1169	/*
1170	 * When waiting for high frequency requests, e.g. during synchronous
1171	 * rendering split between the CPU and GPU, the finite amount of time
1172	 * required to set up the irq and wait upon it limits the response
1173	 * rate. By busywaiting on the request completion for a short while we
1174	 * can service the high frequency waits as quick as possible. However,
1175	 * if it is a slow request, we want to sleep as quickly as possible.
1176	 * The tradeoff between waiting and sleeping is roughly the time it
1177	 * takes to sleep on a request, on the order of a microsecond.
1178	 */
1179
1180	irq = atomic_read(&engine->irq_count);
1181	timeout_us += local_clock_us(&cpu);
1182	do {
1183		if (i915_seqno_passed(intel_engine_get_seqno(engine), seqno))
1184			return seqno == i915_request_global_seqno(rq);
1185
1186		/*
1187		 * Seqno are meant to be ordered *before* the interrupt. If
1188		 * we see an interrupt without a corresponding seqno advance,
1189		 * assume we won't see one in the near future but require
1190		 * the engine->seqno_barrier() to fixup coherency.
1191		 */
1192		if (atomic_read(&engine->irq_count) != irq)
1193			break;
1194
1195		if (signal_pending_state(state, current))
1196			break;
1197
1198		if (busywait_stop(timeout_us, cpu))
1199			break;
1200
1201		cpu_relax();
1202	} while (!need_resched());
1203
1204	return false;
1205}
1206
1207static bool __i915_wait_request_check_and_reset(struct i915_request *request)
1208{
1209	if (likely(!i915_reset_handoff(&request->i915->gpu_error)))
1210		return false;
1211
1212	__set_current_state(TASK_RUNNING);
1213	i915_reset(request->i915, 0);
1214	return true;
1215}
1216
1217/**
1218 * i915_request_wait - wait until execution of request has finished
1219 * @rq: the request to wait upon
1220 * @flags: how to wait
1221 * @timeout: how long to wait in jiffies
1222 *
1223 * i915_request_wait() waits for the request to be completed, for a
1224 * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
1225 * unbounded wait).
1226 *
1227 * If the caller holds the struct_mutex, the caller must pass I915_WAIT_LOCKED
1228 * in via the flags, and vice versa if the struct_mutex is not held, the caller
1229 * must not specify that the wait is locked.
1230 *
1231 * Returns the remaining time (in jiffies) if the request completed, which may
1232 * be zero or -ETIME if the request is unfinished after the timeout expires.
1233 * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
1234 * pending before the request completes.
1235 */
1236long i915_request_wait(struct i915_request *rq,
1237		       unsigned int flags,
1238		       long timeout)
1239{
1240	const int state = flags & I915_WAIT_INTERRUPTIBLE ?
1241		TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
1242	wait_queue_head_t *errq = &rq->i915->gpu_error.wait_queue;
1243	DEFINE_WAIT_FUNC(reset, default_wake_function);
1244	DEFINE_WAIT_FUNC(exec, default_wake_function);
1245	struct intel_wait wait;
1246
1247	might_sleep();
1248#if IS_ENABLED(CONFIG_LOCKDEP)
1249	GEM_BUG_ON(debug_locks &&
1250		   !!lockdep_is_held(&rq->i915->drm.struct_mutex) !=
1251		   !!(flags & I915_WAIT_LOCKED));
1252#endif
1253	GEM_BUG_ON(timeout < 0);
1254
1255	if (i915_request_completed(rq))
1256		return timeout;
1257
1258	if (!timeout)
1259		return -ETIME;
1260
1261	trace_i915_request_wait_begin(rq, flags);
1262
1263	add_wait_queue(&rq->execute, &exec);
1264	if (flags & I915_WAIT_LOCKED)
1265		add_wait_queue(errq, &reset);
1266
1267	intel_wait_init(&wait, rq);
1268
1269restart:
1270	do {
1271		set_current_state(state);
1272		if (intel_wait_update_request(&wait, rq))
1273			break;
1274
1275		if (flags & I915_WAIT_LOCKED &&
1276		    __i915_wait_request_check_and_reset(rq))
1277			continue;
1278
1279		if (signal_pending_state(state, current)) {
1280			timeout = -ERESTARTSYS;
1281			goto complete;
1282		}
1283
1284		if (!timeout) {
1285			timeout = -ETIME;
1286			goto complete;
1287		}
1288
1289		timeout = io_schedule_timeout(timeout);
1290	} while (1);
1291
1292	GEM_BUG_ON(!intel_wait_has_seqno(&wait));
1293	GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit));
1294
1295	/* Optimistic short spin before touching IRQs */
1296	if (__i915_spin_request(rq, wait.seqno, state, 5))
1297		goto complete;
1298
1299	set_current_state(state);
1300	if (intel_engine_add_wait(rq->engine, &wait))
1301		/*
1302		 * In order to check that we haven't missed the interrupt
1303		 * as we enabled it, we need to kick ourselves to do a
1304		 * coherent check on the seqno before we sleep.
1305		 */
1306		goto wakeup;
1307
1308	if (flags & I915_WAIT_LOCKED)
1309		__i915_wait_request_check_and_reset(rq);
1310
1311	for (;;) {
1312		if (signal_pending_state(state, current)) {
1313			timeout = -ERESTARTSYS;
1314			break;
1315		}
1316
1317		if (!timeout) {
1318			timeout = -ETIME;
1319			break;
1320		}
1321
1322		timeout = io_schedule_timeout(timeout);
1323
1324		if (intel_wait_complete(&wait) &&
1325		    intel_wait_check_request(&wait, rq))
1326			break;
1327
1328		set_current_state(state);
1329
1330wakeup:
1331		/*
1332		 * Carefully check if the request is complete, giving time
1333		 * for the seqno to be visible following the interrupt.
1334		 * We also have to check in case we are kicked by the GPU
1335		 * reset in order to drop the struct_mutex.
1336		 */
1337		if (__i915_request_irq_complete(rq))
1338			break;
1339
1340		/*
1341		 * If the GPU is hung, and we hold the lock, reset the GPU
1342		 * and then check for completion. On a full reset, the engine's
1343		 * HW seqno will be advanced passed us and we are complete.
1344		 * If we do a partial reset, we have to wait for the GPU to
1345		 * resume and update the breadcrumb.
1346		 *
1347		 * If we don't hold the mutex, we can just wait for the worker
1348		 * to come along and update the breadcrumb (either directly
1349		 * itself, or indirectly by recovering the GPU).
1350		 */
1351		if (flags & I915_WAIT_LOCKED &&
1352		    __i915_wait_request_check_and_reset(rq))
1353			continue;
1354
1355		/* Only spin if we know the GPU is processing this request */
1356		if (__i915_spin_request(rq, wait.seqno, state, 2))
1357			break;
1358
1359		if (!intel_wait_check_request(&wait, rq)) {
1360			intel_engine_remove_wait(rq->engine, &wait);
1361			goto restart;
1362		}
1363	}
1364
1365	intel_engine_remove_wait(rq->engine, &wait);
1366complete:
1367	__set_current_state(TASK_RUNNING);
1368	if (flags & I915_WAIT_LOCKED)
1369		remove_wait_queue(errq, &reset);
1370	remove_wait_queue(&rq->execute, &exec);
1371	trace_i915_request_wait_end(rq);
1372
1373	return timeout;
1374}
1375
1376static void engine_retire_requests(struct intel_engine_cs *engine)
1377{
1378	struct i915_request *request, *next;
1379	u32 seqno = intel_engine_get_seqno(engine);
1380	LIST_HEAD(retire);
1381
1382	spin_lock_irq(&engine->timeline->lock);
1383	list_for_each_entry_safe(request, next,
1384				 &engine->timeline->requests, link) {
1385		if (!i915_seqno_passed(seqno, request->global_seqno))
1386			break;
1387
1388		list_move_tail(&request->link, &retire);
1389	}
1390	spin_unlock_irq(&engine->timeline->lock);
1391
1392	list_for_each_entry_safe(request, next, &retire, link)
1393		i915_request_retire(request);
1394}
1395
1396void i915_retire_requests(struct drm_i915_private *i915)
1397{
1398	struct intel_engine_cs *engine;
1399	enum intel_engine_id id;
1400
1401	lockdep_assert_held(&i915->drm.struct_mutex);
1402
1403	if (!i915->gt.active_requests)
1404		return;
1405
1406	for_each_engine(engine, i915, id)
1407		engine_retire_requests(engine);
1408}
1409
1410#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1411#include "selftests/mock_request.c"
1412#include "selftests/i915_request.c"
1413#endif