2 * Copyright © 2008-2015 Intel Corporation
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:
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
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
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>
33 static const char *i915_fence_get_driver_name(struct dma_fence *fence)
38 static const char *i915_fence_get_timeline_name(struct dma_fence *fence)
40 /* The timeline struct (as part of the ppgtt underneath a context)
41 * may be freed when the request is no longer in use by the GPU.
42 * We could extend the life of a context to beyond that of all
43 * fences, possibly keeping the hw resource around indefinitely,
44 * or we just give them a false name. Since
45 * dma_fence_ops.get_timeline_name is a debug feature, the occasional
46 * lie seems justifiable.
48 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
51 return to_request(fence)->timeline->common->name;
54 static bool i915_fence_signaled(struct dma_fence *fence)
56 return i915_gem_request_completed(to_request(fence));
59 static bool i915_fence_enable_signaling(struct dma_fence *fence)
61 if (i915_fence_signaled(fence))
64 intel_engine_enable_signaling(to_request(fence), true);
65 return !i915_fence_signaled(fence);
68 static signed long i915_fence_wait(struct dma_fence *fence,
72 return i915_wait_request(to_request(fence), interruptible, timeout);
75 static void i915_fence_release(struct dma_fence *fence)
77 struct drm_i915_gem_request *req = to_request(fence);
79 /* The request is put onto a RCU freelist (i.e. the address
80 * is immediately reused), mark the fences as being freed now.
81 * Otherwise the debugobjects for the fences are only marked as
82 * freed when the slab cache itself is freed, and so we would get
83 * caught trying to reuse dead objects.
85 i915_sw_fence_fini(&req->submit);
87 kmem_cache_free(req->i915->requests, req);
90 const struct dma_fence_ops i915_fence_ops = {
91 .get_driver_name = i915_fence_get_driver_name,
92 .get_timeline_name = i915_fence_get_timeline_name,
93 .enable_signaling = i915_fence_enable_signaling,
94 .signaled = i915_fence_signaled,
95 .wait = i915_fence_wait,
96 .release = i915_fence_release,
100 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
102 struct drm_i915_file_private *file_priv;
104 file_priv = request->file_priv;
108 spin_lock(&file_priv->mm.lock);
109 if (request->file_priv) {
110 list_del(&request->client_link);
111 request->file_priv = NULL;
113 spin_unlock(&file_priv->mm.lock);
116 static struct i915_dependency *
117 i915_dependency_alloc(struct drm_i915_private *i915)
119 return kmem_cache_alloc(i915->dependencies, GFP_KERNEL);
123 i915_dependency_free(struct drm_i915_private *i915,
124 struct i915_dependency *dep)
126 kmem_cache_free(i915->dependencies, dep);
130 __i915_priotree_add_dependency(struct i915_priotree *pt,
131 struct i915_priotree *signal,
132 struct i915_dependency *dep,
135 INIT_LIST_HEAD(&dep->dfs_link);
136 list_add(&dep->wait_link, &signal->waiters_list);
137 list_add(&dep->signal_link, &pt->signalers_list);
138 dep->signaler = signal;
143 i915_priotree_add_dependency(struct drm_i915_private *i915,
144 struct i915_priotree *pt,
145 struct i915_priotree *signal)
147 struct i915_dependency *dep;
149 dep = i915_dependency_alloc(i915);
153 __i915_priotree_add_dependency(pt, signal, dep, I915_DEPENDENCY_ALLOC);
158 i915_priotree_fini(struct drm_i915_private *i915, struct i915_priotree *pt)
160 struct i915_dependency *dep, *next;
162 GEM_BUG_ON(!list_empty(&pt->link));
164 /* Everyone we depended upon (the fences we wait to be signaled)
165 * should retire before us and remove themselves from our list.
166 * However, retirement is run independently on each timeline and
167 * so we may be called out-of-order.
169 list_for_each_entry_safe(dep, next, &pt->signalers_list, signal_link) {
170 list_del(&dep->wait_link);
171 if (dep->flags & I915_DEPENDENCY_ALLOC)
172 i915_dependency_free(i915, dep);
175 /* Remove ourselves from everyone who depends upon us */
176 list_for_each_entry_safe(dep, next, &pt->waiters_list, wait_link) {
177 list_del(&dep->signal_link);
178 if (dep->flags & I915_DEPENDENCY_ALLOC)
179 i915_dependency_free(i915, dep);
184 i915_priotree_init(struct i915_priotree *pt)
186 INIT_LIST_HEAD(&pt->signalers_list);
187 INIT_LIST_HEAD(&pt->waiters_list);
188 INIT_LIST_HEAD(&pt->link);
189 pt->priority = INT_MIN;
192 static int reset_all_global_seqno(struct drm_i915_private *i915, u32 seqno)
194 struct intel_engine_cs *engine;
195 enum intel_engine_id id;
198 /* Carefully retire all requests without writing to the rings */
199 ret = i915_gem_wait_for_idle(i915,
200 I915_WAIT_INTERRUPTIBLE |
205 /* If the seqno wraps around, we need to clear the breadcrumb rbtree */
206 for_each_engine(engine, i915, id) {
207 struct i915_gem_timeline *timeline;
208 struct intel_timeline *tl = engine->timeline;
210 if (!i915_seqno_passed(seqno, tl->seqno)) {
211 /* spin until threads are complete */
212 while (intel_breadcrumbs_busy(engine))
216 /* Check we are idle before we fiddle with hw state! */
217 GEM_BUG_ON(!intel_engine_is_idle(engine));
218 GEM_BUG_ON(i915_gem_active_isset(&engine->timeline->last_request));
220 /* Finally reset hw state */
221 intel_engine_init_global_seqno(engine, seqno);
224 list_for_each_entry(timeline, &i915->gt.timelines, link)
225 memset(timeline->engine[id].global_sync, 0,
226 sizeof(timeline->engine[id].global_sync));
232 int i915_gem_set_global_seqno(struct drm_device *dev, u32 seqno)
234 struct drm_i915_private *dev_priv = to_i915(dev);
236 lockdep_assert_held(&dev_priv->drm.struct_mutex);
241 /* HWS page needs to be set less than what we
242 * will inject to ring
244 return reset_all_global_seqno(dev_priv, seqno - 1);
247 static void mark_busy(struct drm_i915_private *i915)
252 GEM_BUG_ON(!i915->gt.active_requests);
254 intel_runtime_pm_get_noresume(i915);
255 i915->gt.awake = true;
257 intel_enable_gt_powersave(i915);
258 i915_update_gfx_val(i915);
259 if (INTEL_GEN(i915) >= 6)
262 queue_delayed_work(i915->wq,
263 &i915->gt.retire_work,
264 round_jiffies_up_relative(HZ));
267 static int reserve_engine(struct intel_engine_cs *engine)
269 struct drm_i915_private *i915 = engine->i915;
270 u32 active = ++engine->timeline->inflight_seqnos;
271 u32 seqno = engine->timeline->seqno;
274 /* Reservation is fine until we need to wrap around */
275 if (unlikely(add_overflows(seqno, active))) {
276 ret = reset_all_global_seqno(i915, 0);
278 engine->timeline->inflight_seqnos--;
283 if (!i915->gt.active_requests++)
289 static void unreserve_engine(struct intel_engine_cs *engine)
291 struct drm_i915_private *i915 = engine->i915;
293 if (!--i915->gt.active_requests) {
294 /* Cancel the mark_busy() from our reserve_engine() */
295 GEM_BUG_ON(!i915->gt.awake);
296 mod_delayed_work(i915->wq,
298 msecs_to_jiffies(100));
301 GEM_BUG_ON(!engine->timeline->inflight_seqnos);
302 engine->timeline->inflight_seqnos--;
305 void i915_gem_retire_noop(struct i915_gem_active *active,
306 struct drm_i915_gem_request *request)
308 /* Space left intentionally blank */
311 static void advance_ring(struct drm_i915_gem_request *request)
315 /* We know the GPU must have read the request to have
316 * sent us the seqno + interrupt, so use the position
317 * of tail of the request to update the last known position
320 * Note this requires that we are always called in request
323 if (list_is_last(&request->ring_link, &request->ring->request_list)) {
324 /* We may race here with execlists resubmitting this request
325 * as we retire it. The resubmission will move the ring->tail
326 * forwards (to request->wa_tail). We either read the
327 * current value that was written to hw, or the value that
328 * is just about to be. Either works, if we miss the last two
329 * noops - they are safe to be replayed on a reset.
331 tail = READ_ONCE(request->ring->tail);
333 tail = request->postfix;
335 list_del(&request->ring_link);
337 request->ring->head = tail;
340 static void free_capture_list(struct drm_i915_gem_request *request)
342 struct i915_gem_capture_list *capture;
344 capture = request->capture_list;
346 struct i915_gem_capture_list *next = capture->next;
353 static void i915_gem_request_retire(struct drm_i915_gem_request *request)
355 struct intel_engine_cs *engine = request->engine;
356 struct i915_gem_active *active, *next;
358 lockdep_assert_held(&request->i915->drm.struct_mutex);
359 GEM_BUG_ON(!i915_sw_fence_signaled(&request->submit));
360 GEM_BUG_ON(!i915_gem_request_completed(request));
361 GEM_BUG_ON(!request->i915->gt.active_requests);
363 trace_i915_gem_request_retire(request);
365 spin_lock_irq(&engine->timeline->lock);
366 list_del_init(&request->link);
367 spin_unlock_irq(&engine->timeline->lock);
369 unreserve_engine(request->engine);
370 advance_ring(request);
372 free_capture_list(request);
374 /* Walk through the active list, calling retire on each. This allows
375 * objects to track their GPU activity and mark themselves as idle
376 * when their *last* active request is completed (updating state
377 * tracking lists for eviction, active references for GEM, etc).
379 * As the ->retire() may free the node, we decouple it first and
380 * pass along the auxiliary information (to avoid dereferencing
381 * the node after the callback).
383 list_for_each_entry_safe(active, next, &request->active_list, link) {
384 /* In microbenchmarks or focusing upon time inside the kernel,
385 * we may spend an inordinate amount of time simply handling
386 * the retirement of requests and processing their callbacks.
387 * Of which, this loop itself is particularly hot due to the
388 * cache misses when jumping around the list of i915_gem_active.
389 * So we try to keep this loop as streamlined as possible and
390 * also prefetch the next i915_gem_active to try and hide
391 * the likely cache miss.
395 INIT_LIST_HEAD(&active->link);
396 RCU_INIT_POINTER(active->request, NULL);
398 active->retire(active, request);
401 i915_gem_request_remove_from_client(request);
403 /* Retirement decays the ban score as it is a sign of ctx progress */
404 atomic_dec_if_positive(&request->ctx->ban_score);
406 /* The backing object for the context is done after switching to the
407 * *next* context. Therefore we cannot retire the previous context until
408 * the next context has already started running. However, since we
409 * cannot take the required locks at i915_gem_request_submit() we
410 * defer the unpinning of the active context to now, retirement of
411 * the subsequent request.
413 if (engine->last_retired_context)
414 engine->context_unpin(engine, engine->last_retired_context);
415 engine->last_retired_context = request->ctx;
417 spin_lock_irq(&request->lock);
418 if (request->waitboost)
419 atomic_dec(&request->i915->rps.num_waiters);
420 dma_fence_signal_locked(&request->fence);
421 spin_unlock_irq(&request->lock);
423 i915_priotree_fini(request->i915, &request->priotree);
424 i915_gem_request_put(request);
427 void i915_gem_request_retire_upto(struct drm_i915_gem_request *req)
429 struct intel_engine_cs *engine = req->engine;
430 struct drm_i915_gem_request *tmp;
432 lockdep_assert_held(&req->i915->drm.struct_mutex);
433 GEM_BUG_ON(!i915_gem_request_completed(req));
435 if (list_empty(&req->link))
439 tmp = list_first_entry(&engine->timeline->requests,
442 i915_gem_request_retire(tmp);
443 } while (tmp != req);
446 static u32 timeline_get_seqno(struct intel_timeline *tl)
451 void __i915_gem_request_submit(struct drm_i915_gem_request *request)
453 struct intel_engine_cs *engine = request->engine;
454 struct intel_timeline *timeline;
457 GEM_BUG_ON(!irqs_disabled());
458 lockdep_assert_held(&engine->timeline->lock);
460 trace_i915_gem_request_execute(request);
462 /* Transfer from per-context onto the global per-engine timeline */
463 timeline = engine->timeline;
464 GEM_BUG_ON(timeline == request->timeline);
466 seqno = timeline_get_seqno(timeline);
468 GEM_BUG_ON(i915_seqno_passed(intel_engine_get_seqno(engine), seqno));
470 /* We may be recursing from the signal callback of another i915 fence */
471 spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
472 request->global_seqno = seqno;
473 if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
474 intel_engine_enable_signaling(request, false);
475 spin_unlock(&request->lock);
477 engine->emit_breadcrumb(request,
478 request->ring->vaddr + request->postfix);
480 spin_lock(&request->timeline->lock);
481 list_move_tail(&request->link, &timeline->requests);
482 spin_unlock(&request->timeline->lock);
484 wake_up_all(&request->execute);
487 void i915_gem_request_submit(struct drm_i915_gem_request *request)
489 struct intel_engine_cs *engine = request->engine;
492 /* Will be called from irq-context when using foreign fences. */
493 spin_lock_irqsave(&engine->timeline->lock, flags);
495 __i915_gem_request_submit(request);
497 spin_unlock_irqrestore(&engine->timeline->lock, flags);
500 void __i915_gem_request_unsubmit(struct drm_i915_gem_request *request)
502 struct intel_engine_cs *engine = request->engine;
503 struct intel_timeline *timeline;
505 GEM_BUG_ON(!irqs_disabled());
506 lockdep_assert_held(&engine->timeline->lock);
508 /* Only unwind in reverse order, required so that the per-context list
509 * is kept in seqno/ring order.
511 GEM_BUG_ON(request->global_seqno != engine->timeline->seqno);
512 engine->timeline->seqno--;
514 /* We may be recursing from the signal callback of another i915 fence */
515 spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
516 request->global_seqno = 0;
517 if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
518 intel_engine_cancel_signaling(request);
519 spin_unlock(&request->lock);
521 /* Transfer back from the global per-engine timeline to per-context */
522 timeline = request->timeline;
523 GEM_BUG_ON(timeline == engine->timeline);
525 spin_lock(&timeline->lock);
526 list_move(&request->link, &timeline->requests);
527 spin_unlock(&timeline->lock);
529 /* We don't need to wake_up any waiters on request->execute, they
530 * will get woken by any other event or us re-adding this request
531 * to the engine timeline (__i915_gem_request_submit()). The waiters
532 * should be quite adapt at finding that the request now has a new
533 * global_seqno to the one they went to sleep on.
537 void i915_gem_request_unsubmit(struct drm_i915_gem_request *request)
539 struct intel_engine_cs *engine = request->engine;
542 /* Will be called from irq-context when using foreign fences. */
543 spin_lock_irqsave(&engine->timeline->lock, flags);
545 __i915_gem_request_unsubmit(request);
547 spin_unlock_irqrestore(&engine->timeline->lock, flags);
550 static int __i915_sw_fence_call
551 submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
553 struct drm_i915_gem_request *request =
554 container_of(fence, typeof(*request), submit);
558 trace_i915_gem_request_submit(request);
559 request->engine->submit_request(request);
563 i915_gem_request_put(request);
571 * i915_gem_request_alloc - allocate a request structure
573 * @engine: engine that we wish to issue the request on.
574 * @ctx: context that the request will be associated with.
576 * Returns a pointer to the allocated request if successful,
577 * or an error code if not.
579 struct drm_i915_gem_request *
580 i915_gem_request_alloc(struct intel_engine_cs *engine,
581 struct i915_gem_context *ctx)
583 struct drm_i915_private *dev_priv = engine->i915;
584 struct drm_i915_gem_request *req;
585 struct intel_ring *ring;
588 lockdep_assert_held(&dev_priv->drm.struct_mutex);
590 /* ABI: Before userspace accesses the GPU (e.g. execbuffer), report
591 * EIO if the GPU is already wedged.
593 if (i915_terminally_wedged(&dev_priv->gpu_error))
594 return ERR_PTR(-EIO);
596 /* Pinning the contexts may generate requests in order to acquire
597 * GGTT space, so do this first before we reserve a seqno for
600 ring = engine->context_pin(engine, ctx);
602 return ERR_CAST(ring);
605 ret = reserve_engine(engine);
609 /* Move the oldest request to the slab-cache (if not in use!) */
610 req = list_first_entry_or_null(&engine->timeline->requests,
612 if (req && i915_gem_request_completed(req))
613 i915_gem_request_retire(req);
615 /* Beware: Dragons be flying overhead.
617 * We use RCU to look up requests in flight. The lookups may
618 * race with the request being allocated from the slab freelist.
619 * That is the request we are writing to here, may be in the process
620 * of being read by __i915_gem_active_get_rcu(). As such,
621 * we have to be very careful when overwriting the contents. During
622 * the RCU lookup, we change chase the request->engine pointer,
623 * read the request->global_seqno and increment the reference count.
625 * The reference count is incremented atomically. If it is zero,
626 * the lookup knows the request is unallocated and complete. Otherwise,
627 * it is either still in use, or has been reallocated and reset
628 * with dma_fence_init(). This increment is safe for release as we
629 * check that the request we have a reference to and matches the active
632 * Before we increment the refcount, we chase the request->engine
633 * pointer. We must not call kmem_cache_zalloc() or else we set
634 * that pointer to NULL and cause a crash during the lookup. If
635 * we see the request is completed (based on the value of the
636 * old engine and seqno), the lookup is complete and reports NULL.
637 * If we decide the request is not completed (new engine or seqno),
638 * then we grab a reference and double check that it is still the
639 * active request - which it won't be and restart the lookup.
641 * Do not use kmem_cache_zalloc() here!
643 req = kmem_cache_alloc(dev_priv->requests, GFP_KERNEL);
649 req->timeline = i915_gem_context_lookup_timeline(ctx, engine);
650 GEM_BUG_ON(req->timeline == engine->timeline);
652 spin_lock_init(&req->lock);
653 dma_fence_init(&req->fence,
656 req->timeline->fence_context,
657 timeline_get_seqno(req->timeline));
659 /* We bump the ref for the fence chain */
660 i915_sw_fence_init(&i915_gem_request_get(req)->submit, submit_notify);
661 init_waitqueue_head(&req->execute);
663 i915_priotree_init(&req->priotree);
665 INIT_LIST_HEAD(&req->active_list);
666 req->i915 = dev_priv;
667 req->engine = engine;
671 /* No zalloc, must clear what we need by hand */
672 req->global_seqno = 0;
673 req->file_priv = NULL;
675 req->capture_list = NULL;
676 req->waitboost = false;
679 * Reserve space in the ring buffer for all the commands required to
680 * eventually emit this request. This is to guarantee that the
681 * i915_add_request() call can't fail. Note that the reserve may need
682 * to be redone if the request is not actually submitted straight
683 * away, e.g. because a GPU scheduler has deferred it.
685 req->reserved_space = MIN_SPACE_FOR_ADD_REQUEST;
686 GEM_BUG_ON(req->reserved_space < engine->emit_breadcrumb_sz);
688 ret = engine->request_alloc(req);
692 /* Record the position of the start of the request so that
693 * should we detect the updated seqno part-way through the
694 * GPU processing the request, we never over-estimate the
695 * position of the head.
697 req->head = req->ring->emit;
699 /* Check that we didn't interrupt ourselves with a new request */
700 GEM_BUG_ON(req->timeline->seqno != req->fence.seqno);
704 /* Make sure we didn't add ourselves to external state before freeing */
705 GEM_BUG_ON(!list_empty(&req->active_list));
706 GEM_BUG_ON(!list_empty(&req->priotree.signalers_list));
707 GEM_BUG_ON(!list_empty(&req->priotree.waiters_list));
709 kmem_cache_free(dev_priv->requests, req);
711 unreserve_engine(engine);
713 engine->context_unpin(engine, ctx);
718 i915_gem_request_await_request(struct drm_i915_gem_request *to,
719 struct drm_i915_gem_request *from)
723 GEM_BUG_ON(to == from);
724 GEM_BUG_ON(to->timeline == from->timeline);
726 if (i915_gem_request_completed(from))
729 if (to->engine->schedule) {
730 ret = i915_priotree_add_dependency(to->i915,
737 if (to->engine == from->engine) {
738 ret = i915_sw_fence_await_sw_fence_gfp(&to->submit,
741 return ret < 0 ? ret : 0;
744 if (to->engine->semaphore.sync_to) {
747 GEM_BUG_ON(!from->engine->semaphore.signal);
749 seqno = i915_gem_request_global_seqno(from);
751 goto await_dma_fence;
753 if (seqno <= to->timeline->global_sync[from->engine->id])
756 trace_i915_gem_ring_sync_to(to, from);
757 ret = to->engine->semaphore.sync_to(to, from);
761 to->timeline->global_sync[from->engine->id] = seqno;
766 ret = i915_sw_fence_await_dma_fence(&to->submit,
769 return ret < 0 ? ret : 0;
773 i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
774 struct dma_fence *fence)
776 struct dma_fence **child = &fence;
777 unsigned int nchild = 1;
780 /* Note that if the fence-array was created in signal-on-any mode,
781 * we should *not* decompose it into its individual fences. However,
782 * we don't currently store which mode the fence-array is operating
783 * in. Fortunately, the only user of signal-on-any is private to
784 * amdgpu and we should not see any incoming fence-array from
785 * sync-file being in signal-on-any mode.
787 if (dma_fence_is_array(fence)) {
788 struct dma_fence_array *array = to_dma_fence_array(fence);
790 child = array->fences;
791 nchild = array->num_fences;
797 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
801 * Requests on the same timeline are explicitly ordered, along
802 * with their dependencies, by i915_add_request() which ensures
803 * that requests are submitted in-order through each ring.
805 if (fence->context == req->fence.context)
808 /* Squash repeated waits to the same timelines */
809 if (fence->context != req->i915->mm.unordered_timeline &&
810 intel_timeline_sync_is_later(req->timeline, fence))
813 if (dma_fence_is_i915(fence))
814 ret = i915_gem_request_await_request(req,
817 ret = i915_sw_fence_await_dma_fence(&req->submit, fence,
823 /* Record the latest fence used against each timeline */
824 if (fence->context != req->i915->mm.unordered_timeline)
825 intel_timeline_sync_set(req->timeline, fence);
832 * i915_gem_request_await_object - set this request to (async) wait upon a bo
834 * @to: request we are wishing to use
835 * @obj: object which may be in use on another ring.
837 * This code is meant to abstract object synchronization with the GPU.
838 * Conceptually we serialise writes between engines inside the GPU.
839 * We only allow one engine to write into a buffer at any time, but
840 * multiple readers. To ensure each has a coherent view of memory, we must:
842 * - If there is an outstanding write request to the object, the new
843 * request must wait for it to complete (either CPU or in hw, requests
844 * on the same ring will be naturally ordered).
846 * - If we are a write request (pending_write_domain is set), the new
847 * request must wait for outstanding read requests to complete.
849 * Returns 0 if successful, else propagates up the lower layer error.
852 i915_gem_request_await_object(struct drm_i915_gem_request *to,
853 struct drm_i915_gem_object *obj,
856 struct dma_fence *excl;
860 struct dma_fence **shared;
861 unsigned int count, i;
863 ret = reservation_object_get_fences_rcu(obj->resv,
864 &excl, &count, &shared);
868 for (i = 0; i < count; i++) {
869 ret = i915_gem_request_await_dma_fence(to, shared[i]);
873 dma_fence_put(shared[i]);
876 for (; i < count; i++)
877 dma_fence_put(shared[i]);
880 excl = reservation_object_get_excl_rcu(obj->resv);
885 ret = i915_gem_request_await_dma_fence(to, excl);
894 * NB: This function is not allowed to fail. Doing so would mean the the
895 * request is not being tracked for completion but the work itself is
896 * going to happen on the hardware. This would be a Bad Thing(tm).
898 void __i915_add_request(struct drm_i915_gem_request *request, bool flush_caches)
900 struct intel_engine_cs *engine = request->engine;
901 struct intel_ring *ring = request->ring;
902 struct intel_timeline *timeline = request->timeline;
903 struct drm_i915_gem_request *prev;
907 lockdep_assert_held(&request->i915->drm.struct_mutex);
908 trace_i915_gem_request_add(request);
910 /* Make sure that no request gazumped us - if it was allocated after
911 * our i915_gem_request_alloc() and called __i915_add_request() before
912 * us, the timeline will hold its seqno which is later than ours.
914 GEM_BUG_ON(timeline->seqno != request->fence.seqno);
917 * To ensure that this call will not fail, space for its emissions
918 * should already have been reserved in the ring buffer. Let the ring
919 * know that it is time to use that space up.
921 request->reserved_space = 0;
924 * Emit any outstanding flushes - execbuf can fail to emit the flush
925 * after having emitted the batchbuffer command. Hence we need to fix
926 * things up similar to emitting the lazy request. The difference here
927 * is that the flush _must_ happen before the next request, no matter
931 err = engine->emit_flush(request, EMIT_FLUSH);
933 /* Not allowed to fail! */
934 WARN(err, "engine->emit_flush() failed: %d!\n", err);
937 /* Record the position of the start of the breadcrumb so that
938 * should we detect the updated seqno part-way through the
939 * GPU processing the request, we never over-estimate the
940 * position of the ring's HEAD.
942 cs = intel_ring_begin(request, engine->emit_breadcrumb_sz);
943 GEM_BUG_ON(IS_ERR(cs));
944 request->postfix = intel_ring_offset(request, cs);
946 /* Seal the request and mark it as pending execution. Note that
947 * we may inspect this state, without holding any locks, during
948 * hangcheck. Hence we apply the barrier to ensure that we do not
949 * see a more recent value in the hws than we are tracking.
952 prev = i915_gem_active_raw(&timeline->last_request,
953 &request->i915->drm.struct_mutex);
955 i915_sw_fence_await_sw_fence(&request->submit, &prev->submit,
957 if (engine->schedule)
958 __i915_priotree_add_dependency(&request->priotree,
964 spin_lock_irq(&timeline->lock);
965 list_add_tail(&request->link, &timeline->requests);
966 spin_unlock_irq(&timeline->lock);
968 GEM_BUG_ON(timeline->seqno != request->fence.seqno);
969 i915_gem_active_set(&timeline->last_request, request);
971 list_add_tail(&request->ring_link, &ring->request_list);
972 request->emitted_jiffies = jiffies;
974 /* Let the backend know a new request has arrived that may need
975 * to adjust the existing execution schedule due to a high priority
976 * request - i.e. we may want to preempt the current request in order
977 * to run a high priority dependency chain *before* we can execute this
980 * This is called before the request is ready to run so that we can
981 * decide whether to preempt the entire chain so that it is ready to
982 * run at the earliest possible convenience.
984 if (engine->schedule)
985 engine->schedule(request, request->ctx->priority);
988 i915_sw_fence_commit(&request->submit);
989 local_bh_enable(); /* Kick the execlists tasklet if just scheduled */
992 static unsigned long local_clock_us(unsigned int *cpu)
996 /* Cheaply and approximately convert from nanoseconds to microseconds.
997 * The result and subsequent calculations are also defined in the same
998 * approximate microseconds units. The principal source of timing
999 * error here is from the simple truncation.
1001 * Note that local_clock() is only defined wrt to the current CPU;
1002 * the comparisons are no longer valid if we switch CPUs. Instead of
1003 * blocking preemption for the entire busywait, we can detect the CPU
1004 * switch and use that as indicator of system load and a reason to
1005 * stop busywaiting, see busywait_stop().
1008 t = local_clock() >> 10;
1014 static bool busywait_stop(unsigned long timeout, unsigned int cpu)
1016 unsigned int this_cpu;
1018 if (time_after(local_clock_us(&this_cpu), timeout))
1021 return this_cpu != cpu;
1024 bool __i915_spin_request(const struct drm_i915_gem_request *req,
1025 u32 seqno, int state, unsigned long timeout_us)
1027 struct intel_engine_cs *engine = req->engine;
1028 unsigned int irq, cpu;
1030 /* When waiting for high frequency requests, e.g. during synchronous
1031 * rendering split between the CPU and GPU, the finite amount of time
1032 * required to set up the irq and wait upon it limits the response
1033 * rate. By busywaiting on the request completion for a short while we
1034 * can service the high frequency waits as quick as possible. However,
1035 * if it is a slow request, we want to sleep as quickly as possible.
1036 * The tradeoff between waiting and sleeping is roughly the time it
1037 * takes to sleep on a request, on the order of a microsecond.
1040 irq = atomic_read(&engine->irq_count);
1041 timeout_us += local_clock_us(&cpu);
1043 if (seqno != i915_gem_request_global_seqno(req))
1046 if (i915_seqno_passed(intel_engine_get_seqno(req->engine),
1050 /* Seqno are meant to be ordered *before* the interrupt. If
1051 * we see an interrupt without a corresponding seqno advance,
1052 * assume we won't see one in the near future but require
1053 * the engine->seqno_barrier() to fixup coherency.
1055 if (atomic_read(&engine->irq_count) != irq)
1058 if (signal_pending_state(state, current))
1061 if (busywait_stop(timeout_us, cpu))
1065 } while (!need_resched());
1070 static bool __i915_wait_request_check_and_reset(struct drm_i915_gem_request *request)
1072 if (likely(!i915_reset_handoff(&request->i915->gpu_error)))
1075 __set_current_state(TASK_RUNNING);
1076 i915_reset(request->i915, 0);
1081 * i915_wait_request - wait until execution of request has finished
1082 * @req: the request to wait upon
1083 * @flags: how to wait
1084 * @timeout: how long to wait in jiffies
1086 * i915_wait_request() waits for the request to be completed, for a
1087 * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
1090 * If the caller holds the struct_mutex, the caller must pass I915_WAIT_LOCKED
1091 * in via the flags, and vice versa if the struct_mutex is not held, the caller
1092 * must not specify that the wait is locked.
1094 * Returns the remaining time (in jiffies) if the request completed, which may
1095 * be zero or -ETIME if the request is unfinished after the timeout expires.
1096 * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
1097 * pending before the request completes.
1099 long i915_wait_request(struct drm_i915_gem_request *req,
1103 const int state = flags & I915_WAIT_INTERRUPTIBLE ?
1104 TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
1105 wait_queue_head_t *errq = &req->i915->gpu_error.wait_queue;
1106 DEFINE_WAIT_FUNC(reset, default_wake_function);
1107 DEFINE_WAIT_FUNC(exec, default_wake_function);
1108 struct intel_wait wait;
1111 #if IS_ENABLED(CONFIG_LOCKDEP)
1112 GEM_BUG_ON(debug_locks &&
1113 !!lockdep_is_held(&req->i915->drm.struct_mutex) !=
1114 !!(flags & I915_WAIT_LOCKED));
1116 GEM_BUG_ON(timeout < 0);
1118 if (i915_gem_request_completed(req))
1124 trace_i915_gem_request_wait_begin(req, flags);
1126 add_wait_queue(&req->execute, &exec);
1127 if (flags & I915_WAIT_LOCKED)
1128 add_wait_queue(errq, &reset);
1130 intel_wait_init(&wait, req);
1134 set_current_state(state);
1135 if (intel_wait_update_request(&wait, req))
1138 if (flags & I915_WAIT_LOCKED &&
1139 __i915_wait_request_check_and_reset(req))
1142 if (signal_pending_state(state, current)) {
1143 timeout = -ERESTARTSYS;
1152 timeout = io_schedule_timeout(timeout);
1155 GEM_BUG_ON(!intel_wait_has_seqno(&wait));
1156 GEM_BUG_ON(!i915_sw_fence_signaled(&req->submit));
1158 /* Optimistic short spin before touching IRQs */
1159 if (i915_spin_request(req, state, 5))
1162 set_current_state(state);
1163 if (intel_engine_add_wait(req->engine, &wait))
1164 /* In order to check that we haven't missed the interrupt
1165 * as we enabled it, we need to kick ourselves to do a
1166 * coherent check on the seqno before we sleep.
1170 if (flags & I915_WAIT_LOCKED)
1171 __i915_wait_request_check_and_reset(req);
1174 if (signal_pending_state(state, current)) {
1175 timeout = -ERESTARTSYS;
1184 timeout = io_schedule_timeout(timeout);
1186 if (intel_wait_complete(&wait) &&
1187 intel_wait_check_request(&wait, req))
1190 set_current_state(state);
1193 /* Carefully check if the request is complete, giving time
1194 * for the seqno to be visible following the interrupt.
1195 * We also have to check in case we are kicked by the GPU
1196 * reset in order to drop the struct_mutex.
1198 if (__i915_request_irq_complete(req))
1201 /* If the GPU is hung, and we hold the lock, reset the GPU
1202 * and then check for completion. On a full reset, the engine's
1203 * HW seqno will be advanced passed us and we are complete.
1204 * If we do a partial reset, we have to wait for the GPU to
1205 * resume and update the breadcrumb.
1207 * If we don't hold the mutex, we can just wait for the worker
1208 * to come along and update the breadcrumb (either directly
1209 * itself, or indirectly by recovering the GPU).
1211 if (flags & I915_WAIT_LOCKED &&
1212 __i915_wait_request_check_and_reset(req))
1215 /* Only spin if we know the GPU is processing this request */
1216 if (i915_spin_request(req, state, 2))
1219 if (!intel_wait_check_request(&wait, req)) {
1220 intel_engine_remove_wait(req->engine, &wait);
1225 intel_engine_remove_wait(req->engine, &wait);
1227 __set_current_state(TASK_RUNNING);
1228 if (flags & I915_WAIT_LOCKED)
1229 remove_wait_queue(errq, &reset);
1230 remove_wait_queue(&req->execute, &exec);
1231 trace_i915_gem_request_wait_end(req);
1236 static void engine_retire_requests(struct intel_engine_cs *engine)
1238 struct drm_i915_gem_request *request, *next;
1239 u32 seqno = intel_engine_get_seqno(engine);
1242 spin_lock_irq(&engine->timeline->lock);
1243 list_for_each_entry_safe(request, next,
1244 &engine->timeline->requests, link) {
1245 if (!i915_seqno_passed(seqno, request->global_seqno))
1248 list_move_tail(&request->link, &retire);
1250 spin_unlock_irq(&engine->timeline->lock);
1252 list_for_each_entry_safe(request, next, &retire, link)
1253 i915_gem_request_retire(request);
1256 void i915_gem_retire_requests(struct drm_i915_private *dev_priv)
1258 struct intel_engine_cs *engine;
1259 enum intel_engine_id id;
1261 lockdep_assert_held(&dev_priv->drm.struct_mutex);
1263 if (!dev_priv->gt.active_requests)
1266 for_each_engine(engine, dev_priv, id)
1267 engine_retire_requests(engine);
1270 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1271 #include "selftests/mock_request.c"
1272 #include "selftests/i915_gem_request.c"