2 * Copyright © 2016 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
28 #include "intel_ringbuffer.h"
29 #include "intel_frontbuffer.h"
31 #include <drm/drm_gem.h>
33 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
35 #include <linux/stackdepot.h>
37 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
39 unsigned long *entries;
40 unsigned int nr_entries;
43 if (!vma->node.stack) {
44 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
45 vma->node.start, vma->node.size, reason);
49 nr_entries = stack_depot_fetch(vma->node.stack, &entries);
50 stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
51 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
52 vma->node.start, vma->node.size, reason, buf);
57 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
63 static void obj_bump_mru(struct drm_i915_gem_object *obj)
65 struct drm_i915_private *i915 = to_i915(obj->base.dev);
67 spin_lock(&i915->mm.obj_lock);
69 list_move_tail(&obj->mm.link, &i915->mm.bound_list);
70 spin_unlock(&i915->mm.obj_lock);
72 obj->mm.dirty = true; /* be paranoid */
75 static void __i915_vma_retire(struct i915_active *ref)
77 struct i915_vma *vma = container_of(ref, typeof(*vma), active);
78 struct drm_i915_gem_object *obj = vma->obj;
80 GEM_BUG_ON(!i915_gem_object_is_active(obj));
81 if (--obj->active_count)
84 /* Prune the shared fence arrays iff completely idle (inc. external) */
85 if (reservation_object_trylock(obj->resv)) {
86 if (reservation_object_test_signaled_rcu(obj->resv, true))
87 reservation_object_add_excl_fence(obj->resv, NULL);
88 reservation_object_unlock(obj->resv);
92 * Bump our place on the bound list to keep it roughly in LRU order
93 * so that we don't steal from recently used but inactive objects
94 * (unless we are forced to ofc!)
98 if (i915_gem_object_has_active_reference(obj)) {
99 i915_gem_object_clear_active_reference(obj);
100 i915_gem_object_put(obj);
104 static struct i915_vma *
105 vma_create(struct drm_i915_gem_object *obj,
106 struct i915_address_space *vm,
107 const struct i915_ggtt_view *view)
109 struct i915_vma *vma;
110 struct rb_node *rb, **p;
112 /* The aliasing_ppgtt should never be used directly! */
113 GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->vm);
115 vma = kmem_cache_zalloc(vm->i915->vmas, GFP_KERNEL);
117 return ERR_PTR(-ENOMEM);
119 i915_active_init(vm->i915, &vma->active, __i915_vma_retire);
120 INIT_ACTIVE_REQUEST(&vma->last_fence);
123 vma->ops = &vm->vma_ops;
125 vma->resv = obj->resv;
126 vma->size = obj->base.size;
127 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
129 if (view && view->type != I915_GGTT_VIEW_NORMAL) {
130 vma->ggtt_view = *view;
131 if (view->type == I915_GGTT_VIEW_PARTIAL) {
132 GEM_BUG_ON(range_overflows_t(u64,
133 view->partial.offset,
135 obj->base.size >> PAGE_SHIFT));
136 vma->size = view->partial.size;
137 vma->size <<= PAGE_SHIFT;
138 GEM_BUG_ON(vma->size > obj->base.size);
139 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
140 vma->size = intel_rotation_info_size(&view->rotated);
141 vma->size <<= PAGE_SHIFT;
145 if (unlikely(vma->size > vm->total))
148 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
150 if (i915_is_ggtt(vm)) {
151 if (unlikely(overflows_type(vma->size, u32)))
154 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
155 i915_gem_object_get_tiling(obj),
156 i915_gem_object_get_stride(obj));
157 if (unlikely(vma->fence_size < vma->size || /* overflow */
158 vma->fence_size > vm->total))
161 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
163 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
164 i915_gem_object_get_tiling(obj),
165 i915_gem_object_get_stride(obj));
166 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
168 vma->flags |= I915_VMA_GGTT;
171 spin_lock(&obj->vma.lock);
174 p = &obj->vma.tree.rb_node;
176 struct i915_vma *pos;
180 pos = rb_entry(rb, struct i915_vma, obj_node);
183 * If the view already exists in the tree, another thread
184 * already created a matching vma, so return the older instance
185 * and dispose of ours.
187 cmp = i915_vma_compare(pos, vm, view);
189 spin_unlock(&obj->vma.lock);
190 kmem_cache_free(vm->i915->vmas, vma);
199 rb_link_node(&vma->obj_node, rb, p);
200 rb_insert_color(&vma->obj_node, &obj->vma.tree);
202 if (i915_vma_is_ggtt(vma))
204 * We put the GGTT vma at the start of the vma-list, followed
205 * by the ppGGTT vma. This allows us to break early when
206 * iterating over only the GGTT vma for an object, see
207 * for_each_ggtt_vma()
209 list_add(&vma->obj_link, &obj->vma.list);
211 list_add_tail(&vma->obj_link, &obj->vma.list);
213 spin_unlock(&obj->vma.lock);
215 mutex_lock(&vm->mutex);
216 list_add(&vma->vm_link, &vm->unbound_list);
217 mutex_unlock(&vm->mutex);
222 kmem_cache_free(vm->i915->vmas, vma);
223 return ERR_PTR(-E2BIG);
226 static struct i915_vma *
227 vma_lookup(struct drm_i915_gem_object *obj,
228 struct i915_address_space *vm,
229 const struct i915_ggtt_view *view)
233 rb = obj->vma.tree.rb_node;
235 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
238 cmp = i915_vma_compare(vma, vm, view);
252 * i915_vma_instance - return the singleton instance of the VMA
253 * @obj: parent &struct drm_i915_gem_object to be mapped
254 * @vm: address space in which the mapping is located
255 * @view: additional mapping requirements
257 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
258 * the same @view characteristics. If a match is not found, one is created.
259 * Once created, the VMA is kept until either the object is freed, or the
260 * address space is closed.
262 * Must be called with struct_mutex held.
264 * Returns the vma, or an error pointer.
267 i915_vma_instance(struct drm_i915_gem_object *obj,
268 struct i915_address_space *vm,
269 const struct i915_ggtt_view *view)
271 struct i915_vma *vma;
273 GEM_BUG_ON(view && !i915_is_ggtt(vm));
274 GEM_BUG_ON(vm->closed);
276 spin_lock(&obj->vma.lock);
277 vma = vma_lookup(obj, vm, view);
278 spin_unlock(&obj->vma.lock);
280 /* vma_create() will resolve the race if another creates the vma */
282 vma = vma_create(obj, vm, view);
284 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
289 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
291 * @cache_level: mapping cache level
292 * @flags: flags like global or local mapping
294 * DMA addresses are taken from the scatter-gather table of this object (or of
295 * this VMA in case of non-default GGTT views) and PTE entries set up.
296 * Note that DMA addresses are also the only part of the SG table we care about.
298 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
305 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
306 GEM_BUG_ON(vma->size > vma->node.size);
308 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
313 if (GEM_DEBUG_WARN_ON(!flags))
317 if (flags & PIN_GLOBAL)
318 bind_flags |= I915_VMA_GLOBAL_BIND;
319 if (flags & PIN_USER)
320 bind_flags |= I915_VMA_LOCAL_BIND;
322 vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
323 if (flags & PIN_UPDATE)
324 bind_flags |= vma_flags;
326 bind_flags &= ~vma_flags;
330 GEM_BUG_ON(!vma->pages);
332 trace_i915_vma_bind(vma, bind_flags);
333 ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
337 vma->flags |= bind_flags;
341 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
346 /* Access through the GTT requires the device to be awake. */
347 assert_rpm_wakelock_held(vma->vm->i915);
349 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
350 if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
355 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
356 GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
360 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
373 err = i915_vma_pin_fence(vma);
377 i915_vma_set_ggtt_write(vma);
381 __i915_vma_unpin(vma);
383 return IO_ERR_PTR(err);
386 void i915_vma_flush_writes(struct i915_vma *vma)
388 if (!i915_vma_has_ggtt_write(vma))
391 i915_gem_flush_ggtt_writes(vma->vm->i915);
393 i915_vma_unset_ggtt_write(vma);
396 void i915_vma_unpin_iomap(struct i915_vma *vma)
398 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
400 GEM_BUG_ON(vma->iomap == NULL);
402 i915_vma_flush_writes(vma);
404 i915_vma_unpin_fence(vma);
408 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
410 struct i915_vma *vma;
411 struct drm_i915_gem_object *obj;
413 vma = fetch_and_zero(p_vma);
423 if (flags & I915_VMA_RELEASE_MAP)
424 i915_gem_object_unpin_map(obj);
426 __i915_gem_object_release_unless_active(obj);
429 bool i915_vma_misplaced(const struct i915_vma *vma,
430 u64 size, u64 alignment, u64 flags)
432 if (!drm_mm_node_allocated(&vma->node))
435 if (vma->node.size < size)
438 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
439 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
442 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
445 if (flags & PIN_OFFSET_BIAS &&
446 vma->node.start < (flags & PIN_OFFSET_MASK))
449 if (flags & PIN_OFFSET_FIXED &&
450 vma->node.start != (flags & PIN_OFFSET_MASK))
456 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
458 bool mappable, fenceable;
460 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
461 GEM_BUG_ON(!vma->fence_size);
464 * Explicitly disable for rotated VMA since the display does not
465 * need the fence and the VMA is not accessible to other users.
467 if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
470 fenceable = (vma->node.size >= vma->fence_size &&
471 IS_ALIGNED(vma->node.start, vma->fence_alignment));
473 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
475 if (mappable && fenceable)
476 vma->flags |= I915_VMA_CAN_FENCE;
478 vma->flags &= ~I915_VMA_CAN_FENCE;
481 static bool color_differs(struct drm_mm_node *node, unsigned long color)
483 return node->allocated && node->color != color;
486 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
488 struct drm_mm_node *node = &vma->node;
489 struct drm_mm_node *other;
492 * On some machines we have to be careful when putting differing types
493 * of snoopable memory together to avoid the prefetcher crossing memory
494 * domains and dying. During vm initialisation, we decide whether or not
495 * these constraints apply and set the drm_mm.color_adjust
498 if (vma->vm->mm.color_adjust == NULL)
501 /* Only valid to be called on an already inserted vma */
502 GEM_BUG_ON(!drm_mm_node_allocated(node));
503 GEM_BUG_ON(list_empty(&node->node_list));
505 other = list_prev_entry(node, node_list);
506 if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
509 other = list_next_entry(node, node_list);
510 if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
516 static void assert_bind_count(const struct drm_i915_gem_object *obj)
519 * Combine the assertion that the object is bound and that we have
520 * pinned its pages. But we should never have bound the object
521 * more than we have pinned its pages. (For complete accuracy, we
522 * assume that no else is pinning the pages, but as a rough assertion
523 * that we will not run into problems later, this will do!)
525 GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
529 * i915_vma_insert - finds a slot for the vma in its address space
531 * @size: requested size in bytes (can be larger than the VMA)
532 * @alignment: required alignment
533 * @flags: mask of PIN_* flags to use
535 * First we try to allocate some free space that meets the requirements for
536 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
537 * preferrably the oldest idle entry to make room for the new VMA.
540 * 0 on success, negative error code otherwise.
543 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
545 struct drm_i915_private *dev_priv = vma->vm->i915;
546 unsigned int cache_level;
550 GEM_BUG_ON(i915_vma_is_closed(vma));
551 GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
552 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
554 size = max(size, vma->size);
555 alignment = max(alignment, vma->display_alignment);
556 if (flags & PIN_MAPPABLE) {
557 size = max_t(typeof(size), size, vma->fence_size);
558 alignment = max_t(typeof(alignment),
559 alignment, vma->fence_alignment);
562 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
563 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
564 GEM_BUG_ON(!is_power_of_2(alignment));
566 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
567 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
569 end = vma->vm->total;
570 if (flags & PIN_MAPPABLE)
571 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
572 if (flags & PIN_ZONE_4G)
573 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
574 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
576 /* If binding the object/GGTT view requires more space than the entire
577 * aperture has, reject it early before evicting everything in a vain
578 * attempt to find space.
581 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
582 size, flags & PIN_MAPPABLE ? "mappable" : "total",
588 ret = i915_gem_object_pin_pages(vma->obj);
592 cache_level = vma->obj->cache_level;
597 GEM_BUG_ON(vma->pages);
599 ret = vma->ops->set_pages(vma);
603 if (flags & PIN_OFFSET_FIXED) {
604 u64 offset = flags & PIN_OFFSET_MASK;
605 if (!IS_ALIGNED(offset, alignment) ||
606 range_overflows(offset, size, end)) {
611 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
612 size, offset, cache_level,
618 * We only support huge gtt pages through the 48b PPGTT,
619 * however we also don't want to force any alignment for
620 * objects which need to be tightly packed into the low 32bits.
622 * Note that we assume that GGTT are limited to 4GiB for the
623 * forseeable future. See also i915_ggtt_offset().
625 if (upper_32_bits(end - 1) &&
626 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
628 * We can't mix 64K and 4K PTEs in the same page-table
629 * (2M block), and so to avoid the ugliness and
630 * complexity of coloring we opt for just aligning 64K
634 rounddown_pow_of_two(vma->page_sizes.sg |
635 I915_GTT_PAGE_SIZE_2M);
638 * Check we don't expand for the limited Global GTT
639 * (mappable aperture is even more precious!). This
640 * also checks that we exclude the aliasing-ppgtt.
642 GEM_BUG_ON(i915_vma_is_ggtt(vma));
644 alignment = max(alignment, page_alignment);
646 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
647 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
650 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
651 size, alignment, cache_level,
656 GEM_BUG_ON(vma->node.start < start);
657 GEM_BUG_ON(vma->node.start + vma->node.size > end);
659 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
660 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, cache_level));
662 mutex_lock(&vma->vm->mutex);
663 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
664 mutex_unlock(&vma->vm->mutex);
667 struct drm_i915_gem_object *obj = vma->obj;
669 spin_lock(&dev_priv->mm.obj_lock);
670 list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
672 spin_unlock(&dev_priv->mm.obj_lock);
674 assert_bind_count(obj);
680 vma->ops->clear_pages(vma);
683 i915_gem_object_unpin_pages(vma->obj);
688 i915_vma_remove(struct i915_vma *vma)
690 struct drm_i915_private *i915 = vma->vm->i915;
692 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
693 GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
695 vma->ops->clear_pages(vma);
697 mutex_lock(&vma->vm->mutex);
698 drm_mm_remove_node(&vma->node);
699 list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
700 mutex_unlock(&vma->vm->mutex);
703 * Since the unbound list is global, only move to that list if
704 * no more VMAs exist.
707 struct drm_i915_gem_object *obj = vma->obj;
709 spin_lock(&i915->mm.obj_lock);
710 if (--obj->bind_count == 0)
711 list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
712 spin_unlock(&i915->mm.obj_lock);
715 * And finally now the object is completely decoupled from this
716 * vma, we can drop its hold on the backing storage and allow
717 * it to be reaped by the shrinker.
719 i915_gem_object_unpin_pages(obj);
720 assert_bind_count(obj);
724 int __i915_vma_do_pin(struct i915_vma *vma,
725 u64 size, u64 alignment, u64 flags)
727 const unsigned int bound = vma->flags;
730 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
731 GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
732 GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
734 if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
739 if ((bound & I915_VMA_BIND_MASK) == 0) {
740 ret = i915_vma_insert(vma, size, alignment, flags);
744 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
746 ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
750 GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
752 if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
753 __i915_vma_set_map_and_fenceable(vma);
755 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
759 if ((bound & I915_VMA_BIND_MASK) == 0) {
760 i915_vma_remove(vma);
761 GEM_BUG_ON(vma->pages);
762 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
765 __i915_vma_unpin(vma);
769 void i915_vma_close(struct i915_vma *vma)
771 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
773 GEM_BUG_ON(i915_vma_is_closed(vma));
774 vma->flags |= I915_VMA_CLOSED;
777 * We defer actually closing, unbinding and destroying the VMA until
778 * the next idle point, or if the object is freed in the meantime. By
779 * postponing the unbind, we allow for it to be resurrected by the
780 * client, avoiding the work required to rebind the VMA. This is
781 * advantageous for DRI, where the client/server pass objects
782 * between themselves, temporarily opening a local VMA to the
783 * object, and then closing it again. The same object is then reused
784 * on the next frame (or two, depending on the depth of the swap queue)
785 * causing us to rebind the VMA once more. This ends up being a lot
786 * of wasted work for the steady state.
788 list_add_tail(&vma->closed_link, &vma->vm->i915->gt.closed_vma);
791 void i915_vma_reopen(struct i915_vma *vma)
793 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
795 if (vma->flags & I915_VMA_CLOSED) {
796 vma->flags &= ~I915_VMA_CLOSED;
797 list_del(&vma->closed_link);
801 static void __i915_vma_destroy(struct i915_vma *vma)
803 struct drm_i915_private *i915 = vma->vm->i915;
805 GEM_BUG_ON(vma->node.allocated);
806 GEM_BUG_ON(vma->fence);
808 GEM_BUG_ON(i915_active_request_isset(&vma->last_fence));
810 mutex_lock(&vma->vm->mutex);
811 list_del(&vma->vm_link);
812 mutex_unlock(&vma->vm->mutex);
815 struct drm_i915_gem_object *obj = vma->obj;
817 spin_lock(&obj->vma.lock);
818 list_del(&vma->obj_link);
819 rb_erase(&vma->obj_node, &vma->obj->vma.tree);
820 spin_unlock(&obj->vma.lock);
823 i915_active_fini(&vma->active);
825 kmem_cache_free(i915->vmas, vma);
828 void i915_vma_destroy(struct i915_vma *vma)
830 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
832 GEM_BUG_ON(i915_vma_is_active(vma));
833 GEM_BUG_ON(i915_vma_is_pinned(vma));
835 if (i915_vma_is_closed(vma))
836 list_del(&vma->closed_link);
838 WARN_ON(i915_vma_unbind(vma));
839 __i915_vma_destroy(vma);
842 void i915_vma_parked(struct drm_i915_private *i915)
844 struct i915_vma *vma, *next;
846 list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
847 GEM_BUG_ON(!i915_vma_is_closed(vma));
848 i915_vma_destroy(vma);
851 GEM_BUG_ON(!list_empty(&i915->gt.closed_vma));
854 static void __i915_vma_iounmap(struct i915_vma *vma)
856 GEM_BUG_ON(i915_vma_is_pinned(vma));
858 if (vma->iomap == NULL)
861 io_mapping_unmap(vma->iomap);
865 void i915_vma_revoke_mmap(struct i915_vma *vma)
867 struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
870 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
872 if (!i915_vma_has_userfault(vma))
875 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
876 GEM_BUG_ON(!vma->obj->userfault_count);
878 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
879 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
880 drm_vma_node_offset_addr(node) + vma_offset,
884 i915_vma_unset_userfault(vma);
885 if (!--vma->obj->userfault_count)
886 list_del(&vma->obj->userfault_link);
889 static void export_fence(struct i915_vma *vma,
890 struct i915_request *rq,
893 struct reservation_object *resv = vma->resv;
896 * Ignore errors from failing to allocate the new fence, we can't
897 * handle an error right now. Worst case should be missed
898 * synchronisation leading to rendering corruption.
900 reservation_object_lock(resv, NULL);
901 if (flags & EXEC_OBJECT_WRITE)
902 reservation_object_add_excl_fence(resv, &rq->fence);
903 else if (reservation_object_reserve_shared(resv, 1) == 0)
904 reservation_object_add_shared_fence(resv, &rq->fence);
905 reservation_object_unlock(resv);
908 int i915_vma_move_to_active(struct i915_vma *vma,
909 struct i915_request *rq,
912 struct drm_i915_gem_object *obj = vma->obj;
914 lockdep_assert_held(&rq->i915->drm.struct_mutex);
915 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
918 * Add a reference if we're newly entering the active list.
919 * The order in which we add operations to the retirement queue is
920 * vital here: mark_active adds to the start of the callback list,
921 * such that subsequent callbacks are called first. Therefore we
922 * add the active reference first and queue for it to be dropped
925 if (!vma->active.count)
928 if (unlikely(i915_active_ref(&vma->active, rq->fence.context, rq))) {
929 if (!vma->active.count)
934 GEM_BUG_ON(!i915_vma_is_active(vma));
935 GEM_BUG_ON(!obj->active_count);
937 obj->write_domain = 0;
938 if (flags & EXEC_OBJECT_WRITE) {
939 obj->write_domain = I915_GEM_DOMAIN_RENDER;
941 if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
942 __i915_active_request_set(&obj->frontbuffer_write, rq);
944 obj->read_domains = 0;
946 obj->read_domains |= I915_GEM_GPU_DOMAINS;
948 if (flags & EXEC_OBJECT_NEEDS_FENCE)
949 __i915_active_request_set(&vma->last_fence, rq);
951 export_fence(vma, rq, flags);
955 int i915_vma_unbind(struct i915_vma *vma)
959 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
962 * First wait upon any activity as retiring the request may
963 * have side-effects such as unpinning or even unbinding this vma.
966 if (i915_vma_is_active(vma)) {
968 * When a closed VMA is retired, it is unbound - eek.
969 * In order to prevent it from being recursively closed,
970 * take a pin on the vma so that the second unbind is
973 * Even more scary is that the retire callback may free
974 * the object (last active vma). To prevent the explosion
975 * we defer the actual object free to a worker that can
976 * only proceed once it acquires the struct_mutex (which
977 * we currently hold, therefore it cannot free this object
978 * before we are finished).
982 ret = i915_active_wait(&vma->active);
986 ret = i915_active_request_retire(&vma->last_fence,
987 &vma->vm->i915->drm.struct_mutex);
989 __i915_vma_unpin(vma);
993 GEM_BUG_ON(i915_vma_is_active(vma));
995 if (i915_vma_is_pinned(vma)) {
996 vma_print_allocator(vma, "is pinned");
1000 if (!drm_mm_node_allocated(&vma->node))
1003 if (i915_vma_is_map_and_fenceable(vma)) {
1005 * Check that we have flushed all writes through the GGTT
1006 * before the unbind, other due to non-strict nature of those
1007 * indirect writes they may end up referencing the GGTT PTE
1010 i915_vma_flush_writes(vma);
1011 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1013 /* release the fence reg _after_ flushing */
1014 ret = i915_vma_put_fence(vma);
1018 /* Force a pagefault for domain tracking on next user access */
1019 i915_vma_revoke_mmap(vma);
1021 __i915_vma_iounmap(vma);
1022 vma->flags &= ~I915_VMA_CAN_FENCE;
1024 GEM_BUG_ON(vma->fence);
1025 GEM_BUG_ON(i915_vma_has_userfault(vma));
1027 if (likely(!vma->vm->closed)) {
1028 trace_i915_vma_unbind(vma);
1029 vma->ops->unbind_vma(vma);
1031 vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1033 i915_vma_remove(vma);
1038 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1039 #include "selftests/i915_vma.c"