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>
30 static const char *i915_fence_get_driver_name(struct dma_fence *fence)
35 static const char *i915_fence_get_timeline_name(struct dma_fence *fence)
37 return to_request(fence)->timeline->common->name;
40 static bool i915_fence_signaled(struct dma_fence *fence)
42 return i915_gem_request_completed(to_request(fence));
45 static bool i915_fence_enable_signaling(struct dma_fence *fence)
47 if (i915_fence_signaled(fence))
50 intel_engine_enable_signaling(to_request(fence));
54 static signed long i915_fence_wait(struct dma_fence *fence,
58 return i915_wait_request(to_request(fence), interruptible, timeout);
61 static void i915_fence_release(struct dma_fence *fence)
63 struct drm_i915_gem_request *req = to_request(fence);
65 kmem_cache_free(req->i915->requests, req);
68 const struct dma_fence_ops i915_fence_ops = {
69 .get_driver_name = i915_fence_get_driver_name,
70 .get_timeline_name = i915_fence_get_timeline_name,
71 .enable_signaling = i915_fence_enable_signaling,
72 .signaled = i915_fence_signaled,
73 .wait = i915_fence_wait,
74 .release = i915_fence_release,
77 int i915_gem_request_add_to_client(struct drm_i915_gem_request *req,
78 struct drm_file *file)
80 struct drm_i915_private *dev_private;
81 struct drm_i915_file_private *file_priv;
83 WARN_ON(!req || !file || req->file_priv);
91 dev_private = req->i915;
92 file_priv = file->driver_priv;
94 spin_lock(&file_priv->mm.lock);
95 req->file_priv = file_priv;
96 list_add_tail(&req->client_list, &file_priv->mm.request_list);
97 spin_unlock(&file_priv->mm.lock);
103 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
105 struct drm_i915_file_private *file_priv = request->file_priv;
110 spin_lock(&file_priv->mm.lock);
111 list_del(&request->client_list);
112 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(!RB_EMPTY_NODE(&pt->node));
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 RB_CLEAR_NODE(&pt->node);
189 pt->priority = INT_MIN;
192 void i915_gem_retire_noop(struct i915_gem_active *active,
193 struct drm_i915_gem_request *request)
195 /* Space left intentionally blank */
198 static void i915_gem_request_retire(struct drm_i915_gem_request *request)
200 struct i915_gem_active *active, *next;
202 lockdep_assert_held(&request->i915->drm.struct_mutex);
203 GEM_BUG_ON(!i915_sw_fence_signaled(&request->submit));
204 GEM_BUG_ON(!i915_sw_fence_signaled(&request->execute));
205 GEM_BUG_ON(!i915_gem_request_completed(request));
206 GEM_BUG_ON(!request->i915->gt.active_requests);
208 trace_i915_gem_request_retire(request);
210 spin_lock_irq(&request->engine->timeline->lock);
211 list_del_init(&request->link);
212 spin_unlock_irq(&request->engine->timeline->lock);
214 /* We know the GPU must have read the request to have
215 * sent us the seqno + interrupt, so use the position
216 * of tail of the request to update the last known position
219 * Note this requires that we are always called in request
222 list_del(&request->ring_link);
223 request->ring->last_retired_head = request->postfix;
224 if (!--request->i915->gt.active_requests) {
225 GEM_BUG_ON(!request->i915->gt.awake);
226 mod_delayed_work(request->i915->wq,
227 &request->i915->gt.idle_work,
228 msecs_to_jiffies(100));
231 /* Walk through the active list, calling retire on each. This allows
232 * objects to track their GPU activity and mark themselves as idle
233 * when their *last* active request is completed (updating state
234 * tracking lists for eviction, active references for GEM, etc).
236 * As the ->retire() may free the node, we decouple it first and
237 * pass along the auxiliary information (to avoid dereferencing
238 * the node after the callback).
240 list_for_each_entry_safe(active, next, &request->active_list, link) {
241 /* In microbenchmarks or focusing upon time inside the kernel,
242 * we may spend an inordinate amount of time simply handling
243 * the retirement of requests and processing their callbacks.
244 * Of which, this loop itself is particularly hot due to the
245 * cache misses when jumping around the list of i915_gem_active.
246 * So we try to keep this loop as streamlined as possible and
247 * also prefetch the next i915_gem_active to try and hide
248 * the likely cache miss.
252 INIT_LIST_HEAD(&active->link);
253 RCU_INIT_POINTER(active->request, NULL);
255 active->retire(active, request);
258 i915_gem_request_remove_from_client(request);
260 if (request->previous_context) {
261 if (i915.enable_execlists)
262 intel_lr_context_unpin(request->previous_context,
266 /* Retirement decays the ban score as it is a sign of ctx progress */
267 if (request->ctx->ban_score > 0)
268 request->ctx->ban_score--;
270 i915_gem_context_put(request->ctx);
272 dma_fence_signal(&request->fence);
274 i915_priotree_fini(request->i915, &request->priotree);
275 i915_gem_request_put(request);
278 void i915_gem_request_retire_upto(struct drm_i915_gem_request *req)
280 struct intel_engine_cs *engine = req->engine;
281 struct drm_i915_gem_request *tmp;
283 lockdep_assert_held(&req->i915->drm.struct_mutex);
284 GEM_BUG_ON(!i915_gem_request_completed(req));
286 if (list_empty(&req->link))
290 tmp = list_first_entry(&engine->timeline->requests,
293 i915_gem_request_retire(tmp);
294 } while (tmp != req);
297 static int i915_gem_check_wedge(struct drm_i915_private *dev_priv)
299 struct i915_gpu_error *error = &dev_priv->gpu_error;
301 if (i915_terminally_wedged(error))
304 if (i915_reset_in_progress(error)) {
305 /* Non-interruptible callers can't handle -EAGAIN, hence return
306 * -EIO unconditionally for these.
308 if (!dev_priv->mm.interruptible)
317 static int i915_gem_init_global_seqno(struct drm_i915_private *i915, u32 seqno)
319 struct i915_gem_timeline *timeline = &i915->gt.global_timeline;
320 struct intel_engine_cs *engine;
321 enum intel_engine_id id;
324 /* Carefully retire all requests without writing to the rings */
325 ret = i915_gem_wait_for_idle(i915,
326 I915_WAIT_INTERRUPTIBLE |
331 i915_gem_retire_requests(i915);
332 GEM_BUG_ON(i915->gt.active_requests > 1);
334 /* If the seqno wraps around, we need to clear the breadcrumb rbtree */
335 if (!i915_seqno_passed(seqno, atomic_read(&timeline->seqno))) {
336 while (intel_breadcrumbs_busy(i915))
337 cond_resched(); /* spin until threads are complete */
339 atomic_set(&timeline->seqno, seqno);
341 /* Finally reset hw state */
342 for_each_engine(engine, i915, id)
343 intel_engine_init_global_seqno(engine, seqno);
345 list_for_each_entry(timeline, &i915->gt.timelines, link) {
346 for_each_engine(engine, i915, id) {
347 struct intel_timeline *tl = &timeline->engine[id];
349 memset(tl->sync_seqno, 0, sizeof(tl->sync_seqno));
356 int i915_gem_set_global_seqno(struct drm_device *dev, u32 seqno)
358 struct drm_i915_private *dev_priv = to_i915(dev);
360 lockdep_assert_held(&dev_priv->drm.struct_mutex);
365 /* HWS page needs to be set less than what we
366 * will inject to ring
368 return i915_gem_init_global_seqno(dev_priv, seqno - 1);
371 static int reserve_global_seqno(struct drm_i915_private *i915)
373 u32 active_requests = ++i915->gt.active_requests;
374 u32 seqno = atomic_read(&i915->gt.global_timeline.seqno);
377 /* Reservation is fine until we need to wrap around */
378 if (likely(seqno + active_requests > seqno))
381 ret = i915_gem_init_global_seqno(i915, 0);
383 i915->gt.active_requests--;
390 static u32 __timeline_get_seqno(struct i915_gem_timeline *tl)
392 /* seqno only incremented under a mutex */
393 return ++tl->seqno.counter;
396 static u32 timeline_get_seqno(struct i915_gem_timeline *tl)
398 return atomic_inc_return(&tl->seqno);
401 void __i915_gem_request_submit(struct drm_i915_gem_request *request)
403 struct intel_engine_cs *engine = request->engine;
404 struct intel_timeline *timeline;
407 /* Transfer from per-context onto the global per-engine timeline */
408 timeline = engine->timeline;
409 GEM_BUG_ON(timeline == request->timeline);
410 assert_spin_locked(&timeline->lock);
412 seqno = timeline_get_seqno(timeline->common);
414 GEM_BUG_ON(i915_seqno_passed(intel_engine_get_seqno(engine), seqno));
416 GEM_BUG_ON(i915_seqno_passed(timeline->last_submitted_seqno, seqno));
417 request->previous_seqno = timeline->last_submitted_seqno;
418 timeline->last_submitted_seqno = seqno;
420 /* We may be recursing from the signal callback of another i915 fence */
421 spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
422 request->global_seqno = seqno;
423 if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
424 intel_engine_enable_signaling(request);
425 spin_unlock(&request->lock);
427 GEM_BUG_ON(!request->global_seqno);
428 engine->emit_breadcrumb(request,
429 request->ring->vaddr + request->postfix);
431 spin_lock(&request->timeline->lock);
432 list_move_tail(&request->link, &timeline->requests);
433 spin_unlock(&request->timeline->lock);
435 i915_sw_fence_commit(&request->execute);
438 void i915_gem_request_submit(struct drm_i915_gem_request *request)
440 struct intel_engine_cs *engine = request->engine;
443 /* Will be called from irq-context when using foreign fences. */
444 spin_lock_irqsave(&engine->timeline->lock, flags);
446 __i915_gem_request_submit(request);
448 spin_unlock_irqrestore(&engine->timeline->lock, flags);
451 static int __i915_sw_fence_call
452 submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
454 struct drm_i915_gem_request *request =
455 container_of(fence, typeof(*request), submit);
459 request->engine->submit_request(request);
463 i915_gem_request_put(request);
470 static int __i915_sw_fence_call
471 execute_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
473 struct drm_i915_gem_request *request =
474 container_of(fence, typeof(*request), execute);
481 i915_gem_request_put(request);
489 * i915_gem_request_alloc - allocate a request structure
491 * @engine: engine that we wish to issue the request on.
492 * @ctx: context that the request will be associated with.
493 * This can be NULL if the request is not directly related to
494 * any specific user context, in which case this function will
495 * choose an appropriate context to use.
497 * Returns a pointer to the allocated request if successful,
498 * or an error code if not.
500 struct drm_i915_gem_request *
501 i915_gem_request_alloc(struct intel_engine_cs *engine,
502 struct i915_gem_context *ctx)
504 struct drm_i915_private *dev_priv = engine->i915;
505 struct drm_i915_gem_request *req;
508 lockdep_assert_held(&dev_priv->drm.struct_mutex);
510 /* ABI: Before userspace accesses the GPU (e.g. execbuffer), report
511 * EIO if the GPU is already wedged, or EAGAIN to drop the struct_mutex
514 ret = i915_gem_check_wedge(dev_priv);
518 ret = reserve_global_seqno(dev_priv);
522 /* Move the oldest request to the slab-cache (if not in use!) */
523 req = list_first_entry_or_null(&engine->timeline->requests,
525 if (req && __i915_gem_request_completed(req))
526 i915_gem_request_retire(req);
528 /* Beware: Dragons be flying overhead.
530 * We use RCU to look up requests in flight. The lookups may
531 * race with the request being allocated from the slab freelist.
532 * That is the request we are writing to here, may be in the process
533 * of being read by __i915_gem_active_get_rcu(). As such,
534 * we have to be very careful when overwriting the contents. During
535 * the RCU lookup, we change chase the request->engine pointer,
536 * read the request->global_seqno and increment the reference count.
538 * The reference count is incremented atomically. If it is zero,
539 * the lookup knows the request is unallocated and complete. Otherwise,
540 * it is either still in use, or has been reallocated and reset
541 * with dma_fence_init(). This increment is safe for release as we
542 * check that the request we have a reference to and matches the active
545 * Before we increment the refcount, we chase the request->engine
546 * pointer. We must not call kmem_cache_zalloc() or else we set
547 * that pointer to NULL and cause a crash during the lookup. If
548 * we see the request is completed (based on the value of the
549 * old engine and seqno), the lookup is complete and reports NULL.
550 * If we decide the request is not completed (new engine or seqno),
551 * then we grab a reference and double check that it is still the
552 * active request - which it won't be and restart the lookup.
554 * Do not use kmem_cache_zalloc() here!
556 req = kmem_cache_alloc(dev_priv->requests, GFP_KERNEL);
562 req->timeline = i915_gem_context_lookup_timeline(ctx, engine);
563 GEM_BUG_ON(req->timeline == engine->timeline);
565 spin_lock_init(&req->lock);
566 dma_fence_init(&req->fence,
569 req->timeline->fence_context,
570 __timeline_get_seqno(req->timeline->common));
572 /* We bump the ref for the fence chain */
573 i915_sw_fence_init(&i915_gem_request_get(req)->submit, submit_notify);
574 i915_sw_fence_init(&i915_gem_request_get(req)->execute, execute_notify);
576 /* Ensure that the execute fence completes after the submit fence -
577 * as we complete the execute fence from within the submit fence
578 * callback, its completion would otherwise be visible first.
580 i915_sw_fence_await_sw_fence(&req->execute, &req->submit, &req->execq);
582 i915_priotree_init(&req->priotree);
584 INIT_LIST_HEAD(&req->active_list);
585 req->i915 = dev_priv;
586 req->engine = engine;
587 req->ctx = i915_gem_context_get(ctx);
589 /* No zalloc, must clear what we need by hand */
590 req->global_seqno = 0;
591 req->previous_context = NULL;
592 req->file_priv = NULL;
596 * Reserve space in the ring buffer for all the commands required to
597 * eventually emit this request. This is to guarantee that the
598 * i915_add_request() call can't fail. Note that the reserve may need
599 * to be redone if the request is not actually submitted straight
600 * away, e.g. because a GPU scheduler has deferred it.
602 req->reserved_space = MIN_SPACE_FOR_ADD_REQUEST;
603 GEM_BUG_ON(req->reserved_space < engine->emit_breadcrumb_sz);
605 if (i915.enable_execlists)
606 ret = intel_logical_ring_alloc_request_extras(req);
608 ret = intel_ring_alloc_request_extras(req);
612 /* Record the position of the start of the request so that
613 * should we detect the updated seqno part-way through the
614 * GPU processing the request, we never over-estimate the
615 * position of the head.
617 req->head = req->ring->tail;
622 /* Make sure we didn't add ourselves to external state before freeing */
623 GEM_BUG_ON(!list_empty(&req->active_list));
624 GEM_BUG_ON(!list_empty(&req->priotree.signalers_list));
625 GEM_BUG_ON(!list_empty(&req->priotree.waiters_list));
627 i915_gem_context_put(ctx);
628 kmem_cache_free(dev_priv->requests, req);
630 dev_priv->gt.active_requests--;
635 i915_gem_request_await_request(struct drm_i915_gem_request *to,
636 struct drm_i915_gem_request *from)
640 GEM_BUG_ON(to == from);
642 if (to->engine->schedule) {
643 ret = i915_priotree_add_dependency(to->i915,
650 if (to->timeline == from->timeline)
653 if (to->engine == from->engine) {
654 ret = i915_sw_fence_await_sw_fence_gfp(&to->submit,
657 return ret < 0 ? ret : 0;
660 if (!from->global_seqno) {
661 ret = i915_sw_fence_await_dma_fence(&to->submit,
664 return ret < 0 ? ret : 0;
667 if (from->global_seqno <= to->timeline->sync_seqno[from->engine->id])
670 trace_i915_gem_ring_sync_to(to, from);
671 if (!i915.semaphores) {
672 if (!i915_spin_request(from, TASK_INTERRUPTIBLE, 2)) {
673 ret = i915_sw_fence_await_dma_fence(&to->submit,
680 ret = to->engine->semaphore.sync_to(to, from);
685 to->timeline->sync_seqno[from->engine->id] = from->global_seqno;
690 i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
691 struct dma_fence *fence)
693 struct dma_fence_array *array;
697 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
700 if (dma_fence_is_i915(fence))
701 return i915_gem_request_await_request(req, to_request(fence));
703 if (!dma_fence_is_array(fence)) {
704 ret = i915_sw_fence_await_dma_fence(&req->submit,
705 fence, I915_FENCE_TIMEOUT,
707 return ret < 0 ? ret : 0;
710 /* Note that if the fence-array was created in signal-on-any mode,
711 * we should *not* decompose it into its individual fences. However,
712 * we don't currently store which mode the fence-array is operating
713 * in. Fortunately, the only user of signal-on-any is private to
714 * amdgpu and we should not see any incoming fence-array from
715 * sync-file being in signal-on-any mode.
718 array = to_dma_fence_array(fence);
719 for (i = 0; i < array->num_fences; i++) {
720 struct dma_fence *child = array->fences[i];
722 if (dma_fence_is_i915(child))
723 ret = i915_gem_request_await_request(req,
726 ret = i915_sw_fence_await_dma_fence(&req->submit,
727 child, I915_FENCE_TIMEOUT,
737 * i915_gem_request_await_object - set this request to (async) wait upon a bo
739 * @to: request we are wishing to use
740 * @obj: object which may be in use on another ring.
742 * This code is meant to abstract object synchronization with the GPU.
743 * Conceptually we serialise writes between engines inside the GPU.
744 * We only allow one engine to write into a buffer at any time, but
745 * multiple readers. To ensure each has a coherent view of memory, we must:
747 * - If there is an outstanding write request to the object, the new
748 * request must wait for it to complete (either CPU or in hw, requests
749 * on the same ring will be naturally ordered).
751 * - If we are a write request (pending_write_domain is set), the new
752 * request must wait for outstanding read requests to complete.
754 * Returns 0 if successful, else propagates up the lower layer error.
757 i915_gem_request_await_object(struct drm_i915_gem_request *to,
758 struct drm_i915_gem_object *obj,
761 struct dma_fence *excl;
765 struct dma_fence **shared;
766 unsigned int count, i;
768 ret = reservation_object_get_fences_rcu(obj->resv,
769 &excl, &count, &shared);
773 for (i = 0; i < count; i++) {
774 ret = i915_gem_request_await_dma_fence(to, shared[i]);
778 dma_fence_put(shared[i]);
781 for (; i < count; i++)
782 dma_fence_put(shared[i]);
785 excl = reservation_object_get_excl_rcu(obj->resv);
790 ret = i915_gem_request_await_dma_fence(to, excl);
798 static void i915_gem_mark_busy(const struct intel_engine_cs *engine)
800 struct drm_i915_private *dev_priv = engine->i915;
802 if (dev_priv->gt.awake)
805 GEM_BUG_ON(!dev_priv->gt.active_requests);
807 intel_runtime_pm_get_noresume(dev_priv);
808 dev_priv->gt.awake = true;
810 intel_enable_gt_powersave(dev_priv);
811 i915_update_gfx_val(dev_priv);
812 if (INTEL_GEN(dev_priv) >= 6)
813 gen6_rps_busy(dev_priv);
815 queue_delayed_work(dev_priv->wq,
816 &dev_priv->gt.retire_work,
817 round_jiffies_up_relative(HZ));
821 * NB: This function is not allowed to fail. Doing so would mean the the
822 * request is not being tracked for completion but the work itself is
823 * going to happen on the hardware. This would be a Bad Thing(tm).
825 void __i915_add_request(struct drm_i915_gem_request *request, bool flush_caches)
827 struct intel_engine_cs *engine = request->engine;
828 struct intel_ring *ring = request->ring;
829 struct intel_timeline *timeline = request->timeline;
830 struct drm_i915_gem_request *prev;
833 lockdep_assert_held(&request->i915->drm.struct_mutex);
834 trace_i915_gem_request_add(request);
837 * To ensure that this call will not fail, space for its emissions
838 * should already have been reserved in the ring buffer. Let the ring
839 * know that it is time to use that space up.
841 request->reserved_space = 0;
844 * Emit any outstanding flushes - execbuf can fail to emit the flush
845 * after having emitted the batchbuffer command. Hence we need to fix
846 * things up similar to emitting the lazy request. The difference here
847 * is that the flush _must_ happen before the next request, no matter
851 err = engine->emit_flush(request, EMIT_FLUSH);
853 /* Not allowed to fail! */
854 WARN(err, "engine->emit_flush() failed: %d!\n", err);
857 /* Record the position of the start of the breadcrumb so that
858 * should we detect the updated seqno part-way through the
859 * GPU processing the request, we never over-estimate the
860 * position of the ring's HEAD.
862 err = intel_ring_begin(request, engine->emit_breadcrumb_sz);
864 request->postfix = ring->tail;
865 ring->tail += engine->emit_breadcrumb_sz * sizeof(u32);
867 /* Seal the request and mark it as pending execution. Note that
868 * we may inspect this state, without holding any locks, during
869 * hangcheck. Hence we apply the barrier to ensure that we do not
870 * see a more recent value in the hws than we are tracking.
873 prev = i915_gem_active_raw(&timeline->last_request,
874 &request->i915->drm.struct_mutex);
876 i915_sw_fence_await_sw_fence(&request->submit, &prev->submit,
878 if (engine->schedule)
879 __i915_priotree_add_dependency(&request->priotree,
885 spin_lock_irq(&timeline->lock);
886 list_add_tail(&request->link, &timeline->requests);
887 spin_unlock_irq(&timeline->lock);
889 GEM_BUG_ON(i915_seqno_passed(timeline->last_submitted_seqno,
890 request->fence.seqno));
892 timeline->last_submitted_seqno = request->fence.seqno;
893 i915_gem_active_set(&timeline->last_request, request);
895 list_add_tail(&request->ring_link, &ring->request_list);
896 request->emitted_jiffies = jiffies;
898 i915_gem_mark_busy(engine);
900 /* Let the backend know a new request has arrived that may need
901 * to adjust the existing execution schedule due to a high priority
902 * request - i.e. we may want to preempt the current request in order
903 * to run a high priority dependency chain *before* we can execute this
906 * This is called before the request is ready to run so that we can
907 * decide whether to preempt the entire chain so that it is ready to
908 * run at the earliest possible convenience.
910 if (engine->schedule)
911 engine->schedule(request, request->ctx->priority);
914 i915_sw_fence_commit(&request->submit);
915 local_bh_enable(); /* Kick the execlists tasklet if just scheduled */
918 static void reset_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
922 spin_lock_irqsave(&q->lock, flags);
923 if (list_empty(&wait->task_list))
924 __add_wait_queue(q, wait);
925 spin_unlock_irqrestore(&q->lock, flags);
928 static unsigned long local_clock_us(unsigned int *cpu)
932 /* Cheaply and approximately convert from nanoseconds to microseconds.
933 * The result and subsequent calculations are also defined in the same
934 * approximate microseconds units. The principal source of timing
935 * error here is from the simple truncation.
937 * Note that local_clock() is only defined wrt to the current CPU;
938 * the comparisons are no longer valid if we switch CPUs. Instead of
939 * blocking preemption for the entire busywait, we can detect the CPU
940 * switch and use that as indicator of system load and a reason to
941 * stop busywaiting, see busywait_stop().
944 t = local_clock() >> 10;
950 static bool busywait_stop(unsigned long timeout, unsigned int cpu)
952 unsigned int this_cpu;
954 if (time_after(local_clock_us(&this_cpu), timeout))
957 return this_cpu != cpu;
960 bool __i915_spin_request(const struct drm_i915_gem_request *req,
961 int state, unsigned long timeout_us)
965 /* When waiting for high frequency requests, e.g. during synchronous
966 * rendering split between the CPU and GPU, the finite amount of time
967 * required to set up the irq and wait upon it limits the response
968 * rate. By busywaiting on the request completion for a short while we
969 * can service the high frequency waits as quick as possible. However,
970 * if it is a slow request, we want to sleep as quickly as possible.
971 * The tradeoff between waiting and sleeping is roughly the time it
972 * takes to sleep on a request, on the order of a microsecond.
975 timeout_us += local_clock_us(&cpu);
977 if (__i915_gem_request_completed(req))
980 if (signal_pending_state(state, current))
983 if (busywait_stop(timeout_us, cpu))
986 cpu_relax_lowlatency();
987 } while (!need_resched());
993 __i915_request_wait_for_execute(struct drm_i915_gem_request *request,
997 const int state = flags & I915_WAIT_INTERRUPTIBLE ?
998 TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
999 wait_queue_head_t *q = &request->i915->gpu_error.wait_queue;
1003 if (flags & I915_WAIT_LOCKED)
1004 add_wait_queue(q, &reset);
1007 prepare_to_wait(&request->execute.wait, &wait, state);
1009 if (i915_sw_fence_done(&request->execute))
1012 if (flags & I915_WAIT_LOCKED &&
1013 i915_reset_in_progress(&request->i915->gpu_error)) {
1014 __set_current_state(TASK_RUNNING);
1015 i915_reset(request->i915);
1016 reset_wait_queue(q, &reset);
1020 if (signal_pending_state(state, current)) {
1021 timeout = -ERESTARTSYS;
1025 timeout = io_schedule_timeout(timeout);
1027 finish_wait(&request->execute.wait, &wait);
1029 if (flags & I915_WAIT_LOCKED)
1030 remove_wait_queue(q, &reset);
1036 * i915_wait_request - wait until execution of request has finished
1037 * @req: the request to wait upon
1038 * @flags: how to wait
1039 * @timeout: how long to wait in jiffies
1041 * i915_wait_request() waits for the request to be completed, for a
1042 * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
1045 * If the caller holds the struct_mutex, the caller must pass I915_WAIT_LOCKED
1046 * in via the flags, and vice versa if the struct_mutex is not held, the caller
1047 * must not specify that the wait is locked.
1049 * Returns the remaining time (in jiffies) if the request completed, which may
1050 * be zero or -ETIME if the request is unfinished after the timeout expires.
1051 * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
1052 * pending before the request completes.
1054 long i915_wait_request(struct drm_i915_gem_request *req,
1058 const int state = flags & I915_WAIT_INTERRUPTIBLE ?
1059 TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
1061 struct intel_wait wait;
1064 #if IS_ENABLED(CONFIG_LOCKDEP)
1065 GEM_BUG_ON(debug_locks &&
1066 !!lockdep_is_held(&req->i915->drm.struct_mutex) !=
1067 !!(flags & I915_WAIT_LOCKED));
1069 GEM_BUG_ON(timeout < 0);
1071 if (i915_gem_request_completed(req))
1077 trace_i915_gem_request_wait_begin(req);
1079 if (!i915_sw_fence_done(&req->execute)) {
1080 timeout = __i915_request_wait_for_execute(req, flags, timeout);
1084 GEM_BUG_ON(!i915_sw_fence_done(&req->execute));
1086 GEM_BUG_ON(!i915_sw_fence_done(&req->submit));
1087 GEM_BUG_ON(!req->global_seqno);
1089 /* Optimistic short spin before touching IRQs */
1090 if (i915_spin_request(req, state, 5))
1093 set_current_state(state);
1094 if (flags & I915_WAIT_LOCKED)
1095 add_wait_queue(&req->i915->gpu_error.wait_queue, &reset);
1097 intel_wait_init(&wait, req->global_seqno);
1098 if (intel_engine_add_wait(req->engine, &wait))
1099 /* In order to check that we haven't missed the interrupt
1100 * as we enabled it, we need to kick ourselves to do a
1101 * coherent check on the seqno before we sleep.
1106 if (signal_pending_state(state, current)) {
1107 timeout = -ERESTARTSYS;
1116 timeout = io_schedule_timeout(timeout);
1118 if (intel_wait_complete(&wait))
1121 set_current_state(state);
1124 /* Carefully check if the request is complete, giving time
1125 * for the seqno to be visible following the interrupt.
1126 * We also have to check in case we are kicked by the GPU
1127 * reset in order to drop the struct_mutex.
1129 if (__i915_request_irq_complete(req))
1132 /* If the GPU is hung, and we hold the lock, reset the GPU
1133 * and then check for completion. On a full reset, the engine's
1134 * HW seqno will be advanced passed us and we are complete.
1135 * If we do a partial reset, we have to wait for the GPU to
1136 * resume and update the breadcrumb.
1138 * If we don't hold the mutex, we can just wait for the worker
1139 * to come along and update the breadcrumb (either directly
1140 * itself, or indirectly by recovering the GPU).
1142 if (flags & I915_WAIT_LOCKED &&
1143 i915_reset_in_progress(&req->i915->gpu_error)) {
1144 __set_current_state(TASK_RUNNING);
1145 i915_reset(req->i915);
1146 reset_wait_queue(&req->i915->gpu_error.wait_queue,
1151 /* Only spin if we know the GPU is processing this request */
1152 if (i915_spin_request(req, state, 2))
1156 intel_engine_remove_wait(req->engine, &wait);
1157 if (flags & I915_WAIT_LOCKED)
1158 remove_wait_queue(&req->i915->gpu_error.wait_queue, &reset);
1159 __set_current_state(TASK_RUNNING);
1162 trace_i915_gem_request_wait_end(req);
1167 static void engine_retire_requests(struct intel_engine_cs *engine)
1169 struct drm_i915_gem_request *request, *next;
1171 list_for_each_entry_safe(request, next,
1172 &engine->timeline->requests, link) {
1173 if (!__i915_gem_request_completed(request))
1176 i915_gem_request_retire(request);
1180 void i915_gem_retire_requests(struct drm_i915_private *dev_priv)
1182 struct intel_engine_cs *engine;
1183 enum intel_engine_id id;
1185 lockdep_assert_held(&dev_priv->drm.struct_mutex);
1187 if (!dev_priv->gt.active_requests)
1190 for_each_engine(engine, dev_priv, id)
1191 engine_retire_requests(engine);