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 
33 static const char *i915_fence_get_driver_name(struct dma_fence *fence)
34 {
35 	return "i915";
36 }
37 
38 static 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->name;
53 }
54 
55 static bool i915_fence_signaled(struct dma_fence *fence)
56 {
57 	return i915_request_completed(to_request(fence));
58 }
59 
60 static bool i915_fence_enable_signaling(struct dma_fence *fence)
61 {
62 	return intel_engine_enable_signaling(to_request(fence), true);
63 }
64 
65 static signed long i915_fence_wait(struct dma_fence *fence,
66 				   bool interruptible,
67 				   signed long timeout)
68 {
69 	return i915_request_wait(to_request(fence), interruptible, timeout);
70 }
71 
72 static void i915_fence_release(struct dma_fence *fence)
73 {
74 	struct i915_request *rq = to_request(fence);
75 
76 	/*
77 	 * The request is put onto a RCU freelist (i.e. the address
78 	 * is immediately reused), mark the fences as being freed now.
79 	 * Otherwise the debugobjects for the fences are only marked as
80 	 * freed when the slab cache itself is freed, and so we would get
81 	 * caught trying to reuse dead objects.
82 	 */
83 	i915_sw_fence_fini(&rq->submit);
84 
85 	kmem_cache_free(rq->i915->requests, rq);
86 }
87 
88 const struct dma_fence_ops i915_fence_ops = {
89 	.get_driver_name = i915_fence_get_driver_name,
90 	.get_timeline_name = i915_fence_get_timeline_name,
91 	.enable_signaling = i915_fence_enable_signaling,
92 	.signaled = i915_fence_signaled,
93 	.wait = i915_fence_wait,
94 	.release = i915_fence_release,
95 };
96 
97 static inline void
98 i915_request_remove_from_client(struct i915_request *request)
99 {
100 	struct drm_i915_file_private *file_priv;
101 
102 	file_priv = request->file_priv;
103 	if (!file_priv)
104 		return;
105 
106 	spin_lock(&file_priv->mm.lock);
107 	if (request->file_priv) {
108 		list_del(&request->client_link);
109 		request->file_priv = NULL;
110 	}
111 	spin_unlock(&file_priv->mm.lock);
112 }
113 
114 static struct i915_dependency *
115 i915_dependency_alloc(struct drm_i915_private *i915)
116 {
117 	return kmem_cache_alloc(i915->dependencies, GFP_KERNEL);
118 }
119 
120 static void
121 i915_dependency_free(struct drm_i915_private *i915,
122 		     struct i915_dependency *dep)
123 {
124 	kmem_cache_free(i915->dependencies, dep);
125 }
126 
127 static void
128 __i915_sched_node_add_dependency(struct i915_sched_node *node,
129 				 struct i915_sched_node *signal,
130 				 struct i915_dependency *dep,
131 				 unsigned long flags)
132 {
133 	INIT_LIST_HEAD(&dep->dfs_link);
134 	list_add(&dep->wait_link, &signal->waiters_list);
135 	list_add(&dep->signal_link, &node->signalers_list);
136 	dep->signaler = signal;
137 	dep->flags = flags;
138 }
139 
140 static int
141 i915_sched_node_add_dependency(struct drm_i915_private *i915,
142 			       struct i915_sched_node *node,
143 			       struct i915_sched_node *signal)
144 {
145 	struct i915_dependency *dep;
146 
147 	dep = i915_dependency_alloc(i915);
148 	if (!dep)
149 		return -ENOMEM;
150 
151 	__i915_sched_node_add_dependency(node, signal, dep,
152 					 I915_DEPENDENCY_ALLOC);
153 	return 0;
154 }
155 
156 static void
157 i915_sched_node_fini(struct drm_i915_private *i915,
158 		     struct i915_sched_node *node)
159 {
160 	struct i915_dependency *dep, *tmp;
161 
162 	GEM_BUG_ON(!list_empty(&node->link));
163 
164 	/*
165 	 * Everyone we depended upon (the fences we wait to be signaled)
166 	 * should retire before us and remove themselves from our list.
167 	 * However, retirement is run independently on each timeline and
168 	 * so we may be called out-of-order.
169 	 */
170 	list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
171 		GEM_BUG_ON(!i915_sched_node_signaled(dep->signaler));
172 		GEM_BUG_ON(!list_empty(&dep->dfs_link));
173 
174 		list_del(&dep->wait_link);
175 		if (dep->flags & I915_DEPENDENCY_ALLOC)
176 			i915_dependency_free(i915, dep);
177 	}
178 
179 	/* Remove ourselves from everyone who depends upon us */
180 	list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
181 		GEM_BUG_ON(dep->signaler != node);
182 		GEM_BUG_ON(!list_empty(&dep->dfs_link));
183 
184 		list_del(&dep->signal_link);
185 		if (dep->flags & I915_DEPENDENCY_ALLOC)
186 			i915_dependency_free(i915, dep);
187 	}
188 }
189 
190 static void
191 i915_sched_node_init(struct i915_sched_node *node)
192 {
193 	INIT_LIST_HEAD(&node->signalers_list);
194 	INIT_LIST_HEAD(&node->waiters_list);
195 	INIT_LIST_HEAD(&node->link);
196 	node->attr.priority = I915_PRIORITY_INVALID;
197 }
198 
199 static int reset_all_global_seqno(struct drm_i915_private *i915, u32 seqno)
200 {
201 	struct intel_engine_cs *engine;
202 	struct i915_timeline *timeline;
203 	enum intel_engine_id id;
204 	int ret;
205 
206 	/* Carefully retire all requests without writing to the rings */
207 	ret = i915_gem_wait_for_idle(i915,
208 				     I915_WAIT_INTERRUPTIBLE |
209 				     I915_WAIT_LOCKED,
210 				     MAX_SCHEDULE_TIMEOUT);
211 	if (ret)
212 		return ret;
213 
214 	GEM_BUG_ON(i915->gt.active_requests);
215 
216 	/* If the seqno wraps around, we need to clear the breadcrumb rbtree */
217 	for_each_engine(engine, i915, id) {
218 		GEM_TRACE("%s seqno %d (current %d) -> %d\n",
219 			  engine->name,
220 			  engine->timeline.seqno,
221 			  intel_engine_get_seqno(engine),
222 			  seqno);
223 
224 		if (!i915_seqno_passed(seqno, engine->timeline.seqno)) {
225 			/* Flush any waiters before we reuse the seqno */
226 			intel_engine_disarm_breadcrumbs(engine);
227 			intel_engine_init_hangcheck(engine);
228 			GEM_BUG_ON(!list_empty(&engine->breadcrumbs.signals));
229 		}
230 
231 		/* Check we are idle before we fiddle with hw state! */
232 		GEM_BUG_ON(!intel_engine_is_idle(engine));
233 		GEM_BUG_ON(i915_gem_active_isset(&engine->timeline.last_request));
234 
235 		/* Finally reset hw state */
236 		intel_engine_init_global_seqno(engine, seqno);
237 		engine->timeline.seqno = seqno;
238 	}
239 
240 	list_for_each_entry(timeline, &i915->gt.timelines, link)
241 		memset(timeline->global_sync, 0, sizeof(timeline->global_sync));
242 
243 	i915->gt.request_serial = seqno;
244 
245 	return 0;
246 }
247 
248 int i915_gem_set_global_seqno(struct drm_device *dev, u32 seqno)
249 {
250 	struct drm_i915_private *i915 = to_i915(dev);
251 
252 	lockdep_assert_held(&i915->drm.struct_mutex);
253 
254 	if (seqno == 0)
255 		return -EINVAL;
256 
257 	/* HWS page needs to be set less than what we will inject to ring */
258 	return reset_all_global_seqno(i915, seqno - 1);
259 }
260 
261 static int reserve_gt(struct drm_i915_private *i915)
262 {
263 	int ret;
264 
265 	/*
266 	 * Reservation is fine until we may need to wrap around
267 	 *
268 	 * By incrementing the serial for every request, we know that no
269 	 * individual engine may exceed that serial (as each is reset to 0
270 	 * on any wrap). This protects even the most pessimistic of migrations
271 	 * of every request from all engines onto just one.
272 	 */
273 	while (unlikely(++i915->gt.request_serial == 0)) {
274 		ret = reset_all_global_seqno(i915, 0);
275 		if (ret) {
276 			i915->gt.request_serial--;
277 			return ret;
278 		}
279 	}
280 
281 	if (!i915->gt.active_requests++)
282 		i915_gem_unpark(i915);
283 
284 	return 0;
285 }
286 
287 static void unreserve_gt(struct drm_i915_private *i915)
288 {
289 	GEM_BUG_ON(!i915->gt.active_requests);
290 	if (!--i915->gt.active_requests)
291 		i915_gem_park(i915);
292 }
293 
294 void i915_gem_retire_noop(struct i915_gem_active *active,
295 			  struct i915_request *request)
296 {
297 	/* Space left intentionally blank */
298 }
299 
300 static void advance_ring(struct i915_request *request)
301 {
302 	struct intel_ring *ring = request->ring;
303 	unsigned int tail;
304 
305 	/*
306 	 * We know the GPU must have read the request to have
307 	 * sent us the seqno + interrupt, so use the position
308 	 * of tail of the request to update the last known position
309 	 * of the GPU head.
310 	 *
311 	 * Note this requires that we are always called in request
312 	 * completion order.
313 	 */
314 	GEM_BUG_ON(!list_is_first(&request->ring_link, &ring->request_list));
315 	if (list_is_last(&request->ring_link, &ring->request_list)) {
316 		/*
317 		 * We may race here with execlists resubmitting this request
318 		 * as we retire it. The resubmission will move the ring->tail
319 		 * forwards (to request->wa_tail). We either read the
320 		 * current value that was written to hw, or the value that
321 		 * is just about to be. Either works, if we miss the last two
322 		 * noops - they are safe to be replayed on a reset.
323 		 */
324 		GEM_TRACE("marking %s as inactive\n", ring->timeline->name);
325 		tail = READ_ONCE(request->tail);
326 		list_del(&ring->active_link);
327 	} else {
328 		tail = request->postfix;
329 	}
330 	list_del_init(&request->ring_link);
331 
332 	ring->head = tail;
333 }
334 
335 static void free_capture_list(struct i915_request *request)
336 {
337 	struct i915_capture_list *capture;
338 
339 	capture = request->capture_list;
340 	while (capture) {
341 		struct i915_capture_list *next = capture->next;
342 
343 		kfree(capture);
344 		capture = next;
345 	}
346 }
347 
348 static void __retire_engine_request(struct intel_engine_cs *engine,
349 				    struct i915_request *rq)
350 {
351 	GEM_TRACE("%s(%s) fence %llx:%d, global=%d, current %d\n",
352 		  __func__, engine->name,
353 		  rq->fence.context, rq->fence.seqno,
354 		  rq->global_seqno,
355 		  intel_engine_get_seqno(engine));
356 
357 	GEM_BUG_ON(!i915_request_completed(rq));
358 
359 	local_irq_disable();
360 
361 	spin_lock(&engine->timeline.lock);
362 	GEM_BUG_ON(!list_is_first(&rq->link, &engine->timeline.requests));
363 	list_del_init(&rq->link);
364 	spin_unlock(&engine->timeline.lock);
365 
366 	spin_lock(&rq->lock);
367 	if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags))
368 		dma_fence_signal_locked(&rq->fence);
369 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &rq->fence.flags))
370 		intel_engine_cancel_signaling(rq);
371 	if (rq->waitboost) {
372 		GEM_BUG_ON(!atomic_read(&rq->i915->gt_pm.rps.num_waiters));
373 		atomic_dec(&rq->i915->gt_pm.rps.num_waiters);
374 	}
375 	spin_unlock(&rq->lock);
376 
377 	local_irq_enable();
378 
379 	/*
380 	 * The backing object for the context is done after switching to the
381 	 * *next* context. Therefore we cannot retire the previous context until
382 	 * the next context has already started running. However, since we
383 	 * cannot take the required locks at i915_request_submit() we
384 	 * defer the unpinning of the active context to now, retirement of
385 	 * the subsequent request.
386 	 */
387 	if (engine->last_retired_context)
388 		intel_context_unpin(engine->last_retired_context);
389 	engine->last_retired_context = rq->hw_context;
390 }
391 
392 static void __retire_engine_upto(struct intel_engine_cs *engine,
393 				 struct i915_request *rq)
394 {
395 	struct i915_request *tmp;
396 
397 	if (list_empty(&rq->link))
398 		return;
399 
400 	do {
401 		tmp = list_first_entry(&engine->timeline.requests,
402 				       typeof(*tmp), link);
403 
404 		GEM_BUG_ON(tmp->engine != engine);
405 		__retire_engine_request(engine, tmp);
406 	} while (tmp != rq);
407 }
408 
409 static void i915_request_retire(struct i915_request *request)
410 {
411 	struct i915_gem_active *active, *next;
412 
413 	GEM_TRACE("%s fence %llx:%d, global=%d, current %d\n",
414 		  request->engine->name,
415 		  request->fence.context, request->fence.seqno,
416 		  request->global_seqno,
417 		  intel_engine_get_seqno(request->engine));
418 
419 	lockdep_assert_held(&request->i915->drm.struct_mutex);
420 	GEM_BUG_ON(!i915_sw_fence_signaled(&request->submit));
421 	GEM_BUG_ON(!i915_request_completed(request));
422 
423 	trace_i915_request_retire(request);
424 
425 	advance_ring(request);
426 	free_capture_list(request);
427 
428 	/*
429 	 * Walk through the active list, calling retire on each. This allows
430 	 * objects to track their GPU activity and mark themselves as idle
431 	 * when their *last* active request is completed (updating state
432 	 * tracking lists for eviction, active references for GEM, etc).
433 	 *
434 	 * As the ->retire() may free the node, we decouple it first and
435 	 * pass along the auxiliary information (to avoid dereferencing
436 	 * the node after the callback).
437 	 */
438 	list_for_each_entry_safe(active, next, &request->active_list, link) {
439 		/*
440 		 * In microbenchmarks or focusing upon time inside the kernel,
441 		 * we may spend an inordinate amount of time simply handling
442 		 * the retirement of requests and processing their callbacks.
443 		 * Of which, this loop itself is particularly hot due to the
444 		 * cache misses when jumping around the list of i915_gem_active.
445 		 * So we try to keep this loop as streamlined as possible and
446 		 * also prefetch the next i915_gem_active to try and hide
447 		 * the likely cache miss.
448 		 */
449 		prefetchw(next);
450 
451 		INIT_LIST_HEAD(&active->link);
452 		RCU_INIT_POINTER(active->request, NULL);
453 
454 		active->retire(active, request);
455 	}
456 
457 	i915_request_remove_from_client(request);
458 
459 	/* Retirement decays the ban score as it is a sign of ctx progress */
460 	atomic_dec_if_positive(&request->gem_context->ban_score);
461 	intel_context_unpin(request->hw_context);
462 
463 	__retire_engine_upto(request->engine, request);
464 
465 	unreserve_gt(request->i915);
466 
467 	i915_sched_node_fini(request->i915, &request->sched);
468 	i915_request_put(request);
469 }
470 
471 void i915_request_retire_upto(struct i915_request *rq)
472 {
473 	struct intel_ring *ring = rq->ring;
474 	struct i915_request *tmp;
475 
476 	GEM_TRACE("%s fence %llx:%d, global=%d, current %d\n",
477 		  rq->engine->name,
478 		  rq->fence.context, rq->fence.seqno,
479 		  rq->global_seqno,
480 		  intel_engine_get_seqno(rq->engine));
481 
482 	lockdep_assert_held(&rq->i915->drm.struct_mutex);
483 	GEM_BUG_ON(!i915_request_completed(rq));
484 
485 	if (list_empty(&rq->ring_link))
486 		return;
487 
488 	do {
489 		tmp = list_first_entry(&ring->request_list,
490 				       typeof(*tmp), ring_link);
491 
492 		i915_request_retire(tmp);
493 	} while (tmp != rq);
494 }
495 
496 static u32 timeline_get_seqno(struct i915_timeline *tl)
497 {
498 	return ++tl->seqno;
499 }
500 
501 static void move_to_timeline(struct i915_request *request,
502 			     struct i915_timeline *timeline)
503 {
504 	GEM_BUG_ON(request->timeline == &request->engine->timeline);
505 	lockdep_assert_held(&request->engine->timeline.lock);
506 
507 	spin_lock(&request->timeline->lock);
508 	list_move_tail(&request->link, &timeline->requests);
509 	spin_unlock(&request->timeline->lock);
510 }
511 
512 void __i915_request_submit(struct i915_request *request)
513 {
514 	struct intel_engine_cs *engine = request->engine;
515 	u32 seqno;
516 
517 	GEM_TRACE("%s fence %llx:%d -> global=%d, current %d\n",
518 		  engine->name,
519 		  request->fence.context, request->fence.seqno,
520 		  engine->timeline.seqno + 1,
521 		  intel_engine_get_seqno(engine));
522 
523 	GEM_BUG_ON(!irqs_disabled());
524 	lockdep_assert_held(&engine->timeline.lock);
525 
526 	GEM_BUG_ON(request->global_seqno);
527 
528 	seqno = timeline_get_seqno(&engine->timeline);
529 	GEM_BUG_ON(!seqno);
530 	GEM_BUG_ON(intel_engine_signaled(engine, seqno));
531 
532 	/* We may be recursing from the signal callback of another i915 fence */
533 	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
534 	request->global_seqno = seqno;
535 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
536 		intel_engine_enable_signaling(request, false);
537 	spin_unlock(&request->lock);
538 
539 	engine->emit_breadcrumb(request,
540 				request->ring->vaddr + request->postfix);
541 
542 	/* Transfer from per-context onto the global per-engine timeline */
543 	move_to_timeline(request, &engine->timeline);
544 
545 	trace_i915_request_execute(request);
546 
547 	wake_up_all(&request->execute);
548 }
549 
550 void i915_request_submit(struct i915_request *request)
551 {
552 	struct intel_engine_cs *engine = request->engine;
553 	unsigned long flags;
554 
555 	/* Will be called from irq-context when using foreign fences. */
556 	spin_lock_irqsave(&engine->timeline.lock, flags);
557 
558 	__i915_request_submit(request);
559 
560 	spin_unlock_irqrestore(&engine->timeline.lock, flags);
561 }
562 
563 void __i915_request_unsubmit(struct i915_request *request)
564 {
565 	struct intel_engine_cs *engine = request->engine;
566 
567 	GEM_TRACE("%s fence %llx:%d <- global=%d, current %d\n",
568 		  engine->name,
569 		  request->fence.context, request->fence.seqno,
570 		  request->global_seqno,
571 		  intel_engine_get_seqno(engine));
572 
573 	GEM_BUG_ON(!irqs_disabled());
574 	lockdep_assert_held(&engine->timeline.lock);
575 
576 	/*
577 	 * Only unwind in reverse order, required so that the per-context list
578 	 * is kept in seqno/ring order.
579 	 */
580 	GEM_BUG_ON(!request->global_seqno);
581 	GEM_BUG_ON(request->global_seqno != engine->timeline.seqno);
582 	GEM_BUG_ON(intel_engine_has_completed(engine, request->global_seqno));
583 	engine->timeline.seqno--;
584 
585 	/* We may be recursing from the signal callback of another i915 fence */
586 	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
587 	request->global_seqno = 0;
588 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
589 		intel_engine_cancel_signaling(request);
590 	spin_unlock(&request->lock);
591 
592 	/* Transfer back from the global per-engine timeline to per-context */
593 	move_to_timeline(request, request->timeline);
594 
595 	/*
596 	 * We don't need to wake_up any waiters on request->execute, they
597 	 * will get woken by any other event or us re-adding this request
598 	 * to the engine timeline (__i915_request_submit()). The waiters
599 	 * should be quite adapt at finding that the request now has a new
600 	 * global_seqno to the one they went to sleep on.
601 	 */
602 }
603 
604 void i915_request_unsubmit(struct i915_request *request)
605 {
606 	struct intel_engine_cs *engine = request->engine;
607 	unsigned long flags;
608 
609 	/* Will be called from irq-context when using foreign fences. */
610 	spin_lock_irqsave(&engine->timeline.lock, flags);
611 
612 	__i915_request_unsubmit(request);
613 
614 	spin_unlock_irqrestore(&engine->timeline.lock, flags);
615 }
616 
617 static int __i915_sw_fence_call
618 submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
619 {
620 	struct i915_request *request =
621 		container_of(fence, typeof(*request), submit);
622 
623 	switch (state) {
624 	case FENCE_COMPLETE:
625 		trace_i915_request_submit(request);
626 		/*
627 		 * We need to serialize use of the submit_request() callback
628 		 * with its hotplugging performed during an emergency
629 		 * i915_gem_set_wedged().  We use the RCU mechanism to mark the
630 		 * critical section in order to force i915_gem_set_wedged() to
631 		 * wait until the submit_request() is completed before
632 		 * proceeding.
633 		 */
634 		rcu_read_lock();
635 		request->engine->submit_request(request);
636 		rcu_read_unlock();
637 		break;
638 
639 	case FENCE_FREE:
640 		i915_request_put(request);
641 		break;
642 	}
643 
644 	return NOTIFY_DONE;
645 }
646 
647 /**
648  * i915_request_alloc - allocate a request structure
649  *
650  * @engine: engine that we wish to issue the request on.
651  * @ctx: context that the request will be associated with.
652  *
653  * Returns a pointer to the allocated request if successful,
654  * or an error code if not.
655  */
656 struct i915_request *
657 i915_request_alloc(struct intel_engine_cs *engine, struct i915_gem_context *ctx)
658 {
659 	struct drm_i915_private *i915 = engine->i915;
660 	struct i915_request *rq;
661 	struct intel_context *ce;
662 	int ret;
663 
664 	lockdep_assert_held(&i915->drm.struct_mutex);
665 
666 	/*
667 	 * Preempt contexts are reserved for exclusive use to inject a
668 	 * preemption context switch. They are never to be used for any trivial
669 	 * request!
670 	 */
671 	GEM_BUG_ON(ctx == i915->preempt_context);
672 
673 	/*
674 	 * ABI: Before userspace accesses the GPU (e.g. execbuffer), report
675 	 * EIO if the GPU is already wedged.
676 	 */
677 	if (i915_terminally_wedged(&i915->gpu_error))
678 		return ERR_PTR(-EIO);
679 
680 	/*
681 	 * Pinning the contexts may generate requests in order to acquire
682 	 * GGTT space, so do this first before we reserve a seqno for
683 	 * ourselves.
684 	 */
685 	ce = intel_context_pin(ctx, engine);
686 	if (IS_ERR(ce))
687 		return ERR_CAST(ce);
688 
689 	ret = reserve_gt(i915);
690 	if (ret)
691 		goto err_unpin;
692 
693 	ret = intel_ring_wait_for_space(ce->ring, MIN_SPACE_FOR_ADD_REQUEST);
694 	if (ret)
695 		goto err_unreserve;
696 
697 	/* Move our oldest request to the slab-cache (if not in use!) */
698 	rq = list_first_entry(&ce->ring->request_list, typeof(*rq), ring_link);
699 	if (!list_is_last(&rq->ring_link, &ce->ring->request_list) &&
700 	    i915_request_completed(rq))
701 		i915_request_retire(rq);
702 
703 	/*
704 	 * Beware: Dragons be flying overhead.
705 	 *
706 	 * We use RCU to look up requests in flight. The lookups may
707 	 * race with the request being allocated from the slab freelist.
708 	 * That is the request we are writing to here, may be in the process
709 	 * of being read by __i915_gem_active_get_rcu(). As such,
710 	 * we have to be very careful when overwriting the contents. During
711 	 * the RCU lookup, we change chase the request->engine pointer,
712 	 * read the request->global_seqno and increment the reference count.
713 	 *
714 	 * The reference count is incremented atomically. If it is zero,
715 	 * the lookup knows the request is unallocated and complete. Otherwise,
716 	 * it is either still in use, or has been reallocated and reset
717 	 * with dma_fence_init(). This increment is safe for release as we
718 	 * check that the request we have a reference to and matches the active
719 	 * request.
720 	 *
721 	 * Before we increment the refcount, we chase the request->engine
722 	 * pointer. We must not call kmem_cache_zalloc() or else we set
723 	 * that pointer to NULL and cause a crash during the lookup. If
724 	 * we see the request is completed (based on the value of the
725 	 * old engine and seqno), the lookup is complete and reports NULL.
726 	 * If we decide the request is not completed (new engine or seqno),
727 	 * then we grab a reference and double check that it is still the
728 	 * active request - which it won't be and restart the lookup.
729 	 *
730 	 * Do not use kmem_cache_zalloc() here!
731 	 */
732 	rq = kmem_cache_alloc(i915->requests,
733 			      GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
734 	if (unlikely(!rq)) {
735 		i915_retire_requests(i915);
736 
737 		/* Ratelimit ourselves to prevent oom from malicious clients */
738 		rq = i915_gem_active_raw(&ce->ring->timeline->last_request,
739 					 &i915->drm.struct_mutex);
740 		if (rq)
741 			cond_synchronize_rcu(rq->rcustate);
742 
743 		/*
744 		 * We've forced the client to stall and catch up with whatever
745 		 * backlog there might have been. As we are assuming that we
746 		 * caused the mempressure, now is an opportune time to
747 		 * recover as much memory from the request pool as is possible.
748 		 * Having already penalized the client to stall, we spend
749 		 * a little extra time to re-optimise page allocation.
750 		 */
751 		kmem_cache_shrink(i915->requests);
752 		rcu_barrier(); /* Recover the TYPESAFE_BY_RCU pages */
753 
754 		rq = kmem_cache_alloc(i915->requests, GFP_KERNEL);
755 		if (!rq) {
756 			ret = -ENOMEM;
757 			goto err_unreserve;
758 		}
759 	}
760 
761 	rq->rcustate = get_state_synchronize_rcu();
762 
763 	INIT_LIST_HEAD(&rq->active_list);
764 	rq->i915 = i915;
765 	rq->engine = engine;
766 	rq->gem_context = ctx;
767 	rq->hw_context = ce;
768 	rq->ring = ce->ring;
769 	rq->timeline = ce->ring->timeline;
770 	GEM_BUG_ON(rq->timeline == &engine->timeline);
771 
772 	spin_lock_init(&rq->lock);
773 	dma_fence_init(&rq->fence,
774 		       &i915_fence_ops,
775 		       &rq->lock,
776 		       rq->timeline->fence_context,
777 		       timeline_get_seqno(rq->timeline));
778 
779 	/* We bump the ref for the fence chain */
780 	i915_sw_fence_init(&i915_request_get(rq)->submit, submit_notify);
781 	init_waitqueue_head(&rq->execute);
782 
783 	i915_sched_node_init(&rq->sched);
784 
785 	/* No zalloc, must clear what we need by hand */
786 	rq->global_seqno = 0;
787 	rq->signaling.wait.seqno = 0;
788 	rq->file_priv = NULL;
789 	rq->batch = NULL;
790 	rq->capture_list = NULL;
791 	rq->waitboost = false;
792 
793 	/*
794 	 * Reserve space in the ring buffer for all the commands required to
795 	 * eventually emit this request. This is to guarantee that the
796 	 * i915_request_add() call can't fail. Note that the reserve may need
797 	 * to be redone if the request is not actually submitted straight
798 	 * away, e.g. because a GPU scheduler has deferred it.
799 	 */
800 	rq->reserved_space = MIN_SPACE_FOR_ADD_REQUEST;
801 	GEM_BUG_ON(rq->reserved_space < engine->emit_breadcrumb_sz);
802 
803 	/*
804 	 * Record the position of the start of the request so that
805 	 * should we detect the updated seqno part-way through the
806 	 * GPU processing the request, we never over-estimate the
807 	 * position of the head.
808 	 */
809 	rq->head = rq->ring->emit;
810 
811 	/* Unconditionally invalidate GPU caches and TLBs. */
812 	ret = engine->emit_flush(rq, EMIT_INVALIDATE);
813 	if (ret)
814 		goto err_unwind;
815 
816 	ret = engine->request_alloc(rq);
817 	if (ret)
818 		goto err_unwind;
819 
820 	/* Keep a second pin for the dual retirement along engine and ring */
821 	__intel_context_pin(ce);
822 
823 	rq->infix = rq->ring->emit; /* end of header; start of user payload */
824 
825 	/* Check that we didn't interrupt ourselves with a new request */
826 	GEM_BUG_ON(rq->timeline->seqno != rq->fence.seqno);
827 	return rq;
828 
829 err_unwind:
830 	ce->ring->emit = rq->head;
831 
832 	/* Make sure we didn't add ourselves to external state before freeing */
833 	GEM_BUG_ON(!list_empty(&rq->active_list));
834 	GEM_BUG_ON(!list_empty(&rq->sched.signalers_list));
835 	GEM_BUG_ON(!list_empty(&rq->sched.waiters_list));
836 
837 	kmem_cache_free(i915->requests, rq);
838 err_unreserve:
839 	unreserve_gt(i915);
840 err_unpin:
841 	intel_context_unpin(ce);
842 	return ERR_PTR(ret);
843 }
844 
845 static int
846 i915_request_await_request(struct i915_request *to, struct i915_request *from)
847 {
848 	int ret;
849 
850 	GEM_BUG_ON(to == from);
851 	GEM_BUG_ON(to->timeline == from->timeline);
852 
853 	if (i915_request_completed(from))
854 		return 0;
855 
856 	if (to->engine->schedule) {
857 		ret = i915_sched_node_add_dependency(to->i915,
858 						     &to->sched,
859 						     &from->sched);
860 		if (ret < 0)
861 			return ret;
862 	}
863 
864 	if (to->engine == from->engine) {
865 		ret = i915_sw_fence_await_sw_fence_gfp(&to->submit,
866 						       &from->submit,
867 						       I915_FENCE_GFP);
868 		return ret < 0 ? ret : 0;
869 	}
870 
871 	if (to->engine->semaphore.sync_to) {
872 		u32 seqno;
873 
874 		GEM_BUG_ON(!from->engine->semaphore.signal);
875 
876 		seqno = i915_request_global_seqno(from);
877 		if (!seqno)
878 			goto await_dma_fence;
879 
880 		if (seqno <= to->timeline->global_sync[from->engine->id])
881 			return 0;
882 
883 		trace_i915_gem_ring_sync_to(to, from);
884 		ret = to->engine->semaphore.sync_to(to, from);
885 		if (ret)
886 			return ret;
887 
888 		to->timeline->global_sync[from->engine->id] = seqno;
889 		return 0;
890 	}
891 
892 await_dma_fence:
893 	ret = i915_sw_fence_await_dma_fence(&to->submit,
894 					    &from->fence, 0,
895 					    I915_FENCE_GFP);
896 	return ret < 0 ? ret : 0;
897 }
898 
899 int
900 i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence)
901 {
902 	struct dma_fence **child = &fence;
903 	unsigned int nchild = 1;
904 	int ret;
905 
906 	/*
907 	 * Note that if the fence-array was created in signal-on-any mode,
908 	 * we should *not* decompose it into its individual fences. However,
909 	 * we don't currently store which mode the fence-array is operating
910 	 * in. Fortunately, the only user of signal-on-any is private to
911 	 * amdgpu and we should not see any incoming fence-array from
912 	 * sync-file being in signal-on-any mode.
913 	 */
914 	if (dma_fence_is_array(fence)) {
915 		struct dma_fence_array *array = to_dma_fence_array(fence);
916 
917 		child = array->fences;
918 		nchild = array->num_fences;
919 		GEM_BUG_ON(!nchild);
920 	}
921 
922 	do {
923 		fence = *child++;
924 		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
925 			continue;
926 
927 		/*
928 		 * Requests on the same timeline are explicitly ordered, along
929 		 * with their dependencies, by i915_request_add() which ensures
930 		 * that requests are submitted in-order through each ring.
931 		 */
932 		if (fence->context == rq->fence.context)
933 			continue;
934 
935 		/* Squash repeated waits to the same timelines */
936 		if (fence->context != rq->i915->mm.unordered_timeline &&
937 		    i915_timeline_sync_is_later(rq->timeline, fence))
938 			continue;
939 
940 		if (dma_fence_is_i915(fence))
941 			ret = i915_request_await_request(rq, to_request(fence));
942 		else
943 			ret = i915_sw_fence_await_dma_fence(&rq->submit, fence,
944 							    I915_FENCE_TIMEOUT,
945 							    I915_FENCE_GFP);
946 		if (ret < 0)
947 			return ret;
948 
949 		/* Record the latest fence used against each timeline */
950 		if (fence->context != rq->i915->mm.unordered_timeline)
951 			i915_timeline_sync_set(rq->timeline, fence);
952 	} while (--nchild);
953 
954 	return 0;
955 }
956 
957 /**
958  * i915_request_await_object - set this request to (async) wait upon a bo
959  * @to: request we are wishing to use
960  * @obj: object which may be in use on another ring.
961  * @write: whether the wait is on behalf of a writer
962  *
963  * This code is meant to abstract object synchronization with the GPU.
964  * Conceptually we serialise writes between engines inside the GPU.
965  * We only allow one engine to write into a buffer at any time, but
966  * multiple readers. To ensure each has a coherent view of memory, we must:
967  *
968  * - If there is an outstanding write request to the object, the new
969  *   request must wait for it to complete (either CPU or in hw, requests
970  *   on the same ring will be naturally ordered).
971  *
972  * - If we are a write request (pending_write_domain is set), the new
973  *   request must wait for outstanding read requests to complete.
974  *
975  * Returns 0 if successful, else propagates up the lower layer error.
976  */
977 int
978 i915_request_await_object(struct i915_request *to,
979 			  struct drm_i915_gem_object *obj,
980 			  bool write)
981 {
982 	struct dma_fence *excl;
983 	int ret = 0;
984 
985 	if (write) {
986 		struct dma_fence **shared;
987 		unsigned int count, i;
988 
989 		ret = reservation_object_get_fences_rcu(obj->resv,
990 							&excl, &count, &shared);
991 		if (ret)
992 			return ret;
993 
994 		for (i = 0; i < count; i++) {
995 			ret = i915_request_await_dma_fence(to, shared[i]);
996 			if (ret)
997 				break;
998 
999 			dma_fence_put(shared[i]);
1000 		}
1001 
1002 		for (; i < count; i++)
1003 			dma_fence_put(shared[i]);
1004 		kfree(shared);
1005 	} else {
1006 		excl = reservation_object_get_excl_rcu(obj->resv);
1007 	}
1008 
1009 	if (excl) {
1010 		if (ret == 0)
1011 			ret = i915_request_await_dma_fence(to, excl);
1012 
1013 		dma_fence_put(excl);
1014 	}
1015 
1016 	return ret;
1017 }
1018 
1019 void i915_request_skip(struct i915_request *rq, int error)
1020 {
1021 	void *vaddr = rq->ring->vaddr;
1022 	u32 head;
1023 
1024 	GEM_BUG_ON(!IS_ERR_VALUE((long)error));
1025 	dma_fence_set_error(&rq->fence, error);
1026 
1027 	/*
1028 	 * As this request likely depends on state from the lost
1029 	 * context, clear out all the user operations leaving the
1030 	 * breadcrumb at the end (so we get the fence notifications).
1031 	 */
1032 	head = rq->infix;
1033 	if (rq->postfix < head) {
1034 		memset(vaddr + head, 0, rq->ring->size - head);
1035 		head = 0;
1036 	}
1037 	memset(vaddr + head, 0, rq->postfix - head);
1038 }
1039 
1040 /*
1041  * NB: This function is not allowed to fail. Doing so would mean the the
1042  * request is not being tracked for completion but the work itself is
1043  * going to happen on the hardware. This would be a Bad Thing(tm).
1044  */
1045 void i915_request_add(struct i915_request *request)
1046 {
1047 	struct intel_engine_cs *engine = request->engine;
1048 	struct i915_timeline *timeline = request->timeline;
1049 	struct intel_ring *ring = request->ring;
1050 	struct i915_request *prev;
1051 	u32 *cs;
1052 
1053 	GEM_TRACE("%s fence %llx:%d\n",
1054 		  engine->name, request->fence.context, request->fence.seqno);
1055 
1056 	lockdep_assert_held(&request->i915->drm.struct_mutex);
1057 	trace_i915_request_add(request);
1058 
1059 	/*
1060 	 * Make sure that no request gazumped us - if it was allocated after
1061 	 * our i915_request_alloc() and called __i915_request_add() before
1062 	 * us, the timeline will hold its seqno which is later than ours.
1063 	 */
1064 	GEM_BUG_ON(timeline->seqno != request->fence.seqno);
1065 
1066 	/*
1067 	 * To ensure that this call will not fail, space for its emissions
1068 	 * should already have been reserved in the ring buffer. Let the ring
1069 	 * know that it is time to use that space up.
1070 	 */
1071 	request->reserved_space = 0;
1072 	engine->emit_flush(request, EMIT_FLUSH);
1073 
1074 	/*
1075 	 * Record the position of the start of the breadcrumb so that
1076 	 * should we detect the updated seqno part-way through the
1077 	 * GPU processing the request, we never over-estimate the
1078 	 * position of the ring's HEAD.
1079 	 */
1080 	cs = intel_ring_begin(request, engine->emit_breadcrumb_sz);
1081 	GEM_BUG_ON(IS_ERR(cs));
1082 	request->postfix = intel_ring_offset(request, cs);
1083 
1084 	/*
1085 	 * Seal the request and mark it as pending execution. Note that
1086 	 * we may inspect this state, without holding any locks, during
1087 	 * hangcheck. Hence we apply the barrier to ensure that we do not
1088 	 * see a more recent value in the hws than we are tracking.
1089 	 */
1090 
1091 	prev = i915_gem_active_raw(&timeline->last_request,
1092 				   &request->i915->drm.struct_mutex);
1093 	if (prev && !i915_request_completed(prev)) {
1094 		i915_sw_fence_await_sw_fence(&request->submit, &prev->submit,
1095 					     &request->submitq);
1096 		if (engine->schedule)
1097 			__i915_sched_node_add_dependency(&request->sched,
1098 							 &prev->sched,
1099 							 &request->dep,
1100 							 0);
1101 	}
1102 
1103 	spin_lock_irq(&timeline->lock);
1104 	list_add_tail(&request->link, &timeline->requests);
1105 	spin_unlock_irq(&timeline->lock);
1106 
1107 	GEM_BUG_ON(timeline->seqno != request->fence.seqno);
1108 	i915_gem_active_set(&timeline->last_request, request);
1109 
1110 	list_add_tail(&request->ring_link, &ring->request_list);
1111 	if (list_is_first(&request->ring_link, &ring->request_list)) {
1112 		GEM_TRACE("marking %s as active\n", ring->timeline->name);
1113 		list_add(&ring->active_link, &request->i915->gt.active_rings);
1114 	}
1115 	request->emitted_jiffies = jiffies;
1116 
1117 	/*
1118 	 * Let the backend know a new request has arrived that may need
1119 	 * to adjust the existing execution schedule due to a high priority
1120 	 * request - i.e. we may want to preempt the current request in order
1121 	 * to run a high priority dependency chain *before* we can execute this
1122 	 * request.
1123 	 *
1124 	 * This is called before the request is ready to run so that we can
1125 	 * decide whether to preempt the entire chain so that it is ready to
1126 	 * run at the earliest possible convenience.
1127 	 */
1128 	local_bh_disable();
1129 	rcu_read_lock(); /* RCU serialisation for set-wedged protection */
1130 	if (engine->schedule)
1131 		engine->schedule(request, &request->gem_context->sched);
1132 	rcu_read_unlock();
1133 	i915_sw_fence_commit(&request->submit);
1134 	local_bh_enable(); /* Kick the execlists tasklet if just scheduled */
1135 
1136 	/*
1137 	 * In typical scenarios, we do not expect the previous request on
1138 	 * the timeline to be still tracked by timeline->last_request if it
1139 	 * has been completed. If the completed request is still here, that
1140 	 * implies that request retirement is a long way behind submission,
1141 	 * suggesting that we haven't been retiring frequently enough from
1142 	 * the combination of retire-before-alloc, waiters and the background
1143 	 * retirement worker. So if the last request on this timeline was
1144 	 * already completed, do a catch up pass, flushing the retirement queue
1145 	 * up to this client. Since we have now moved the heaviest operations
1146 	 * during retirement onto secondary workers, such as freeing objects
1147 	 * or contexts, retiring a bunch of requests is mostly list management
1148 	 * (and cache misses), and so we should not be overly penalizing this
1149 	 * client by performing excess work, though we may still performing
1150 	 * work on behalf of others -- but instead we should benefit from
1151 	 * improved resource management. (Well, that's the theory at least.)
1152 	 */
1153 	if (prev && i915_request_completed(prev))
1154 		i915_request_retire_upto(prev);
1155 }
1156 
1157 static unsigned long local_clock_us(unsigned int *cpu)
1158 {
1159 	unsigned long t;
1160 
1161 	/*
1162 	 * Cheaply and approximately convert from nanoseconds to microseconds.
1163 	 * The result and subsequent calculations are also defined in the same
1164 	 * approximate microseconds units. The principal source of timing
1165 	 * error here is from the simple truncation.
1166 	 *
1167 	 * Note that local_clock() is only defined wrt to the current CPU;
1168 	 * the comparisons are no longer valid if we switch CPUs. Instead of
1169 	 * blocking preemption for the entire busywait, we can detect the CPU
1170 	 * switch and use that as indicator of system load and a reason to
1171 	 * stop busywaiting, see busywait_stop().
1172 	 */
1173 	*cpu = get_cpu();
1174 	t = local_clock() >> 10;
1175 	put_cpu();
1176 
1177 	return t;
1178 }
1179 
1180 static bool busywait_stop(unsigned long timeout, unsigned int cpu)
1181 {
1182 	unsigned int this_cpu;
1183 
1184 	if (time_after(local_clock_us(&this_cpu), timeout))
1185 		return true;
1186 
1187 	return this_cpu != cpu;
1188 }
1189 
1190 static bool __i915_spin_request(const struct i915_request *rq,
1191 				u32 seqno, int state, unsigned long timeout_us)
1192 {
1193 	struct intel_engine_cs *engine = rq->engine;
1194 	unsigned int irq, cpu;
1195 
1196 	GEM_BUG_ON(!seqno);
1197 
1198 	/*
1199 	 * Only wait for the request if we know it is likely to complete.
1200 	 *
1201 	 * We don't track the timestamps around requests, nor the average
1202 	 * request length, so we do not have a good indicator that this
1203 	 * request will complete within the timeout. What we do know is the
1204 	 * order in which requests are executed by the engine and so we can
1205 	 * tell if the request has started. If the request hasn't started yet,
1206 	 * it is a fair assumption that it will not complete within our
1207 	 * relatively short timeout.
1208 	 */
1209 	if (!intel_engine_has_started(engine, seqno))
1210 		return false;
1211 
1212 	/*
1213 	 * When waiting for high frequency requests, e.g. during synchronous
1214 	 * rendering split between the CPU and GPU, the finite amount of time
1215 	 * required to set up the irq and wait upon it limits the response
1216 	 * rate. By busywaiting on the request completion for a short while we
1217 	 * can service the high frequency waits as quick as possible. However,
1218 	 * if it is a slow request, we want to sleep as quickly as possible.
1219 	 * The tradeoff between waiting and sleeping is roughly the time it
1220 	 * takes to sleep on a request, on the order of a microsecond.
1221 	 */
1222 
1223 	irq = READ_ONCE(engine->breadcrumbs.irq_count);
1224 	timeout_us += local_clock_us(&cpu);
1225 	do {
1226 		if (intel_engine_has_completed(engine, seqno))
1227 			return seqno == i915_request_global_seqno(rq);
1228 
1229 		/*
1230 		 * Seqno are meant to be ordered *before* the interrupt. If
1231 		 * we see an interrupt without a corresponding seqno advance,
1232 		 * assume we won't see one in the near future but require
1233 		 * the engine->seqno_barrier() to fixup coherency.
1234 		 */
1235 		if (READ_ONCE(engine->breadcrumbs.irq_count) != irq)
1236 			break;
1237 
1238 		if (signal_pending_state(state, current))
1239 			break;
1240 
1241 		if (busywait_stop(timeout_us, cpu))
1242 			break;
1243 
1244 		cpu_relax();
1245 	} while (!need_resched());
1246 
1247 	return false;
1248 }
1249 
1250 static bool __i915_wait_request_check_and_reset(struct i915_request *request)
1251 {
1252 	struct i915_gpu_error *error = &request->i915->gpu_error;
1253 
1254 	if (likely(!i915_reset_handoff(error)))
1255 		return false;
1256 
1257 	__set_current_state(TASK_RUNNING);
1258 	i915_reset(request->i915, error->stalled_mask, error->reason);
1259 	return true;
1260 }
1261 
1262 /**
1263  * i915_request_wait - wait until execution of request has finished
1264  * @rq: the request to wait upon
1265  * @flags: how to wait
1266  * @timeout: how long to wait in jiffies
1267  *
1268  * i915_request_wait() waits for the request to be completed, for a
1269  * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
1270  * unbounded wait).
1271  *
1272  * If the caller holds the struct_mutex, the caller must pass I915_WAIT_LOCKED
1273  * in via the flags, and vice versa if the struct_mutex is not held, the caller
1274  * must not specify that the wait is locked.
1275  *
1276  * Returns the remaining time (in jiffies) if the request completed, which may
1277  * be zero or -ETIME if the request is unfinished after the timeout expires.
1278  * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
1279  * pending before the request completes.
1280  */
1281 long i915_request_wait(struct i915_request *rq,
1282 		       unsigned int flags,
1283 		       long timeout)
1284 {
1285 	const int state = flags & I915_WAIT_INTERRUPTIBLE ?
1286 		TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
1287 	wait_queue_head_t *errq = &rq->i915->gpu_error.wait_queue;
1288 	DEFINE_WAIT_FUNC(reset, default_wake_function);
1289 	DEFINE_WAIT_FUNC(exec, default_wake_function);
1290 	struct intel_wait wait;
1291 
1292 	might_sleep();
1293 #if IS_ENABLED(CONFIG_LOCKDEP)
1294 	GEM_BUG_ON(debug_locks &&
1295 		   !!lockdep_is_held(&rq->i915->drm.struct_mutex) !=
1296 		   !!(flags & I915_WAIT_LOCKED));
1297 #endif
1298 	GEM_BUG_ON(timeout < 0);
1299 
1300 	if (i915_request_completed(rq))
1301 		return timeout;
1302 
1303 	if (!timeout)
1304 		return -ETIME;
1305 
1306 	trace_i915_request_wait_begin(rq, flags);
1307 
1308 	add_wait_queue(&rq->execute, &exec);
1309 	if (flags & I915_WAIT_LOCKED)
1310 		add_wait_queue(errq, &reset);
1311 
1312 	intel_wait_init(&wait);
1313 
1314 restart:
1315 	do {
1316 		set_current_state(state);
1317 		if (intel_wait_update_request(&wait, rq))
1318 			break;
1319 
1320 		if (flags & I915_WAIT_LOCKED &&
1321 		    __i915_wait_request_check_and_reset(rq))
1322 			continue;
1323 
1324 		if (signal_pending_state(state, current)) {
1325 			timeout = -ERESTARTSYS;
1326 			goto complete;
1327 		}
1328 
1329 		if (!timeout) {
1330 			timeout = -ETIME;
1331 			goto complete;
1332 		}
1333 
1334 		timeout = io_schedule_timeout(timeout);
1335 	} while (1);
1336 
1337 	GEM_BUG_ON(!intel_wait_has_seqno(&wait));
1338 	GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit));
1339 
1340 	/* Optimistic short spin before touching IRQs */
1341 	if (__i915_spin_request(rq, wait.seqno, state, 5))
1342 		goto complete;
1343 
1344 	set_current_state(state);
1345 	if (intel_engine_add_wait(rq->engine, &wait))
1346 		/*
1347 		 * In order to check that we haven't missed the interrupt
1348 		 * as we enabled it, we need to kick ourselves to do a
1349 		 * coherent check on the seqno before we sleep.
1350 		 */
1351 		goto wakeup;
1352 
1353 	if (flags & I915_WAIT_LOCKED)
1354 		__i915_wait_request_check_and_reset(rq);
1355 
1356 	for (;;) {
1357 		if (signal_pending_state(state, current)) {
1358 			timeout = -ERESTARTSYS;
1359 			break;
1360 		}
1361 
1362 		if (!timeout) {
1363 			timeout = -ETIME;
1364 			break;
1365 		}
1366 
1367 		timeout = io_schedule_timeout(timeout);
1368 
1369 		if (intel_wait_complete(&wait) &&
1370 		    intel_wait_check_request(&wait, rq))
1371 			break;
1372 
1373 		set_current_state(state);
1374 
1375 wakeup:
1376 		/*
1377 		 * Carefully check if the request is complete, giving time
1378 		 * for the seqno to be visible following the interrupt.
1379 		 * We also have to check in case we are kicked by the GPU
1380 		 * reset in order to drop the struct_mutex.
1381 		 */
1382 		if (__i915_request_irq_complete(rq))
1383 			break;
1384 
1385 		/*
1386 		 * If the GPU is hung, and we hold the lock, reset the GPU
1387 		 * and then check for completion. On a full reset, the engine's
1388 		 * HW seqno will be advanced passed us and we are complete.
1389 		 * If we do a partial reset, we have to wait for the GPU to
1390 		 * resume and update the breadcrumb.
1391 		 *
1392 		 * If we don't hold the mutex, we can just wait for the worker
1393 		 * to come along and update the breadcrumb (either directly
1394 		 * itself, or indirectly by recovering the GPU).
1395 		 */
1396 		if (flags & I915_WAIT_LOCKED &&
1397 		    __i915_wait_request_check_and_reset(rq))
1398 			continue;
1399 
1400 		/* Only spin if we know the GPU is processing this request */
1401 		if (__i915_spin_request(rq, wait.seqno, state, 2))
1402 			break;
1403 
1404 		if (!intel_wait_check_request(&wait, rq)) {
1405 			intel_engine_remove_wait(rq->engine, &wait);
1406 			goto restart;
1407 		}
1408 	}
1409 
1410 	intel_engine_remove_wait(rq->engine, &wait);
1411 complete:
1412 	__set_current_state(TASK_RUNNING);
1413 	if (flags & I915_WAIT_LOCKED)
1414 		remove_wait_queue(errq, &reset);
1415 	remove_wait_queue(&rq->execute, &exec);
1416 	trace_i915_request_wait_end(rq);
1417 
1418 	return timeout;
1419 }
1420 
1421 static void ring_retire_requests(struct intel_ring *ring)
1422 {
1423 	struct i915_request *request, *next;
1424 
1425 	list_for_each_entry_safe(request, next,
1426 				 &ring->request_list, ring_link) {
1427 		if (!i915_request_completed(request))
1428 			break;
1429 
1430 		i915_request_retire(request);
1431 	}
1432 }
1433 
1434 void i915_retire_requests(struct drm_i915_private *i915)
1435 {
1436 	struct intel_ring *ring, *tmp;
1437 
1438 	lockdep_assert_held(&i915->drm.struct_mutex);
1439 
1440 	if (!i915->gt.active_requests)
1441 		return;
1442 
1443 	list_for_each_entry_safe(ring, tmp, &i915->gt.active_rings, active_link)
1444 		ring_retire_requests(ring);
1445 }
1446 
1447 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1448 #include "selftests/mock_request.c"
1449 #include "selftests/i915_request.c"
1450 #endif
1451