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
25 #include <linux/sched/mm.h>
26 #include <drm/drm_gem.h>
28 #include "display/intel_frontbuffer.h"
30 #include "gem/i915_gem_lmem.h"
31 #include "gt/intel_engine.h"
32 #include "gt/intel_engine_heartbeat.h"
33 #include "gt/intel_gt.h"
34 #include "gt/intel_gt_requests.h"
37 #include "i915_sw_fence_work.h"
38 #include "i915_trace.h"
41 static struct kmem_cache *slab_vmas;
43 struct i915_vma *i915_vma_alloc(void)
45 return kmem_cache_zalloc(slab_vmas, GFP_KERNEL);
48 void i915_vma_free(struct i915_vma *vma)
50 return kmem_cache_free(slab_vmas, vma);
53 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
55 #include <linux/stackdepot.h>
57 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
61 if (!vma->node.stack) {
62 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
63 vma->node.start, vma->node.size, reason);
67 stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0);
68 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
69 vma->node.start, vma->node.size, reason, buf);
74 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
80 static inline struct i915_vma *active_to_vma(struct i915_active *ref)
82 return container_of(ref, typeof(struct i915_vma), active);
85 static int __i915_vma_active(struct i915_active *ref)
87 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
90 static void __i915_vma_retire(struct i915_active *ref)
92 i915_vma_put(active_to_vma(ref));
95 static struct i915_vma *
96 vma_create(struct drm_i915_gem_object *obj,
97 struct i915_address_space *vm,
98 const struct i915_ggtt_view *view)
100 struct i915_vma *pos = ERR_PTR(-E2BIG);
101 struct i915_vma *vma;
102 struct rb_node *rb, **p;
104 /* The aliasing_ppgtt should never be used directly! */
105 GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
107 vma = i915_vma_alloc();
109 return ERR_PTR(-ENOMEM);
111 kref_init(&vma->ref);
112 mutex_init(&vma->pages_mutex);
113 vma->vm = i915_vm_get(vm);
114 vma->ops = &vm->vma_ops;
116 vma->resv = obj->base.resv;
117 vma->size = obj->base.size;
118 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
120 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0);
122 /* Declare ourselves safe for use inside shrinkers */
123 if (IS_ENABLED(CONFIG_LOCKDEP)) {
124 fs_reclaim_acquire(GFP_KERNEL);
125 might_lock(&vma->active.mutex);
126 fs_reclaim_release(GFP_KERNEL);
129 INIT_LIST_HEAD(&vma->closed_link);
131 if (view && view->type != I915_GGTT_VIEW_NORMAL) {
132 vma->ggtt_view = *view;
133 if (view->type == I915_GGTT_VIEW_PARTIAL) {
134 GEM_BUG_ON(range_overflows_t(u64,
135 view->partial.offset,
137 obj->base.size >> PAGE_SHIFT));
138 vma->size = view->partial.size;
139 vma->size <<= PAGE_SHIFT;
140 GEM_BUG_ON(vma->size > obj->base.size);
141 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
142 vma->size = intel_rotation_info_size(&view->rotated);
143 vma->size <<= PAGE_SHIFT;
144 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
145 vma->size = intel_remapped_info_size(&view->remapped);
146 vma->size <<= PAGE_SHIFT;
150 if (unlikely(vma->size > vm->total))
153 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
155 spin_lock(&obj->vma.lock);
157 if (i915_is_ggtt(vm)) {
158 if (unlikely(overflows_type(vma->size, u32)))
161 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
162 i915_gem_object_get_tiling(obj),
163 i915_gem_object_get_stride(obj));
164 if (unlikely(vma->fence_size < vma->size || /* overflow */
165 vma->fence_size > vm->total))
168 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
170 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
171 i915_gem_object_get_tiling(obj),
172 i915_gem_object_get_stride(obj));
173 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
175 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
179 p = &obj->vma.tree.rb_node;
184 pos = rb_entry(rb, struct i915_vma, obj_node);
187 * If the view already exists in the tree, another thread
188 * already created a matching vma, so return the older instance
189 * and dispose of ours.
191 cmp = i915_vma_compare(pos, vm, view);
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);
218 spin_unlock(&obj->vma.lock);
225 static struct i915_vma *
226 i915_vma_lookup(struct drm_i915_gem_object *obj,
227 struct i915_address_space *vm,
228 const struct i915_ggtt_view *view)
232 rb = obj->vma.tree.rb_node;
234 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
237 cmp = i915_vma_compare(vma, vm, view);
251 * i915_vma_instance - return the singleton instance of the VMA
252 * @obj: parent &struct drm_i915_gem_object to be mapped
253 * @vm: address space in which the mapping is located
254 * @view: additional mapping requirements
256 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
257 * the same @view characteristics. If a match is not found, one is created.
258 * Once created, the VMA is kept until either the object is freed, or the
259 * address space is closed.
261 * Returns the vma, or an error pointer.
264 i915_vma_instance(struct drm_i915_gem_object *obj,
265 struct i915_address_space *vm,
266 const struct i915_ggtt_view *view)
268 struct i915_vma *vma;
270 GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm));
271 GEM_BUG_ON(!atomic_read(&vm->open));
273 spin_lock(&obj->vma.lock);
274 vma = i915_vma_lookup(obj, vm, view);
275 spin_unlock(&obj->vma.lock);
277 /* vma_create() will resolve the race if another creates the vma */
279 vma = vma_create(obj, vm, view);
281 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
285 struct i915_vma_work {
286 struct dma_fence_work base;
287 struct i915_address_space *vm;
288 struct i915_vm_pt_stash stash;
289 struct i915_vma *vma;
290 struct drm_i915_gem_object *pinned;
291 struct i915_sw_dma_fence_cb cb;
292 enum i915_cache_level cache_level;
296 static void __vma_bind(struct dma_fence_work *work)
298 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
299 struct i915_vma *vma = vw->vma;
301 vma->ops->bind_vma(vw->vm, &vw->stash,
302 vma, vw->cache_level, vw->flags);
305 static void __vma_release(struct dma_fence_work *work)
307 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
310 __i915_gem_object_unpin_pages(vw->pinned);
311 i915_gem_object_put(vw->pinned);
314 i915_vm_free_pt_stash(vw->vm, &vw->stash);
318 static const struct dma_fence_work_ops bind_ops = {
321 .release = __vma_release,
324 struct i915_vma_work *i915_vma_work(void)
326 struct i915_vma_work *vw;
328 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
332 dma_fence_work_init(&vw->base, &bind_ops);
333 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
338 int i915_vma_wait_for_bind(struct i915_vma *vma)
342 if (rcu_access_pointer(vma->active.excl.fence)) {
343 struct dma_fence *fence;
346 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
349 err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
350 dma_fence_put(fence);
358 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
360 * @cache_level: mapping cache level
361 * @flags: flags like global or local mapping
362 * @work: preallocated worker for allocating and binding the PTE
364 * DMA addresses are taken from the scatter-gather table of this object (or of
365 * this VMA in case of non-default GGTT views) and PTE entries set up.
366 * Note that DMA addresses are also the only part of the SG table we care about.
368 int i915_vma_bind(struct i915_vma *vma,
369 enum i915_cache_level cache_level,
371 struct i915_vma_work *work)
376 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
377 GEM_BUG_ON(vma->size > vma->node.size);
379 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
384 if (GEM_DEBUG_WARN_ON(!flags))
388 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
390 vma_flags = atomic_read(&vma->flags);
391 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
393 bind_flags &= ~vma_flags;
397 GEM_BUG_ON(!vma->pages);
399 trace_i915_vma_bind(vma, bind_flags);
400 if (work && bind_flags & vma->vm->bind_async_flags) {
401 struct dma_fence *prev;
404 work->cache_level = cache_level;
405 work->flags = bind_flags;
408 * Note we only want to chain up to the migration fence on
409 * the pages (not the object itself). As we don't track that,
410 * yet, we have to use the exclusive fence instead.
412 * Also note that we do not want to track the async vma as
413 * part of the obj->resv->excl_fence as it only affects
414 * execution and not content or object's backing store lifetime.
416 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
418 __i915_sw_fence_await_dma_fence(&work->base.chain,
424 work->base.dma.error = 0; /* enable the queue_work() */
427 __i915_gem_object_pin_pages(vma->obj);
428 work->pinned = i915_gem_object_get(vma->obj);
431 vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
434 atomic_or(bind_flags, &vma->flags);
438 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
443 if (!i915_gem_object_is_lmem(vma->obj)) {
444 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
450 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
451 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
453 ptr = READ_ONCE(vma->iomap);
456 * TODO: consider just using i915_gem_object_pin_map() for lmem
457 * instead, which already supports mapping non-contiguous chunks
458 * of pages, that way we can also drop the
459 * I915_BO_ALLOC_CONTIGUOUS when allocating the object.
461 if (i915_gem_object_is_lmem(vma->obj))
462 ptr = i915_gem_object_lmem_io_map(vma->obj, 0,
463 vma->obj->base.size);
465 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
473 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
474 io_mapping_unmap(ptr);
481 err = i915_vma_pin_fence(vma);
485 i915_vma_set_ggtt_write(vma);
487 /* NB Access through the GTT requires the device to be awake. */
491 __i915_vma_unpin(vma);
493 return IO_ERR_PTR(err);
496 void i915_vma_flush_writes(struct i915_vma *vma)
498 if (i915_vma_unset_ggtt_write(vma))
499 intel_gt_flush_ggtt_writes(vma->vm->gt);
502 void i915_vma_unpin_iomap(struct i915_vma *vma)
504 GEM_BUG_ON(vma->iomap == NULL);
506 i915_vma_flush_writes(vma);
508 i915_vma_unpin_fence(vma);
512 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
514 struct i915_vma *vma;
515 struct drm_i915_gem_object *obj;
517 vma = fetch_and_zero(p_vma);
526 if (flags & I915_VMA_RELEASE_MAP)
527 i915_gem_object_unpin_map(obj);
529 i915_gem_object_put(obj);
532 bool i915_vma_misplaced(const struct i915_vma *vma,
533 u64 size, u64 alignment, u64 flags)
535 if (!drm_mm_node_allocated(&vma->node))
538 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
541 if (vma->node.size < size)
544 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
545 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
548 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
551 if (flags & PIN_OFFSET_BIAS &&
552 vma->node.start < (flags & PIN_OFFSET_MASK))
555 if (flags & PIN_OFFSET_FIXED &&
556 vma->node.start != (flags & PIN_OFFSET_MASK))
562 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
564 bool mappable, fenceable;
566 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
567 GEM_BUG_ON(!vma->fence_size);
569 fenceable = (vma->node.size >= vma->fence_size &&
570 IS_ALIGNED(vma->node.start, vma->fence_alignment));
572 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
574 if (mappable && fenceable)
575 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
577 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
580 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
582 struct drm_mm_node *node = &vma->node;
583 struct drm_mm_node *other;
586 * On some machines we have to be careful when putting differing types
587 * of snoopable memory together to avoid the prefetcher crossing memory
588 * domains and dying. During vm initialisation, we decide whether or not
589 * these constraints apply and set the drm_mm.color_adjust
592 if (!i915_vm_has_cache_coloring(vma->vm))
595 /* Only valid to be called on an already inserted vma */
596 GEM_BUG_ON(!drm_mm_node_allocated(node));
597 GEM_BUG_ON(list_empty(&node->node_list));
599 other = list_prev_entry(node, node_list);
600 if (i915_node_color_differs(other, color) &&
601 !drm_mm_hole_follows(other))
604 other = list_next_entry(node, node_list);
605 if (i915_node_color_differs(other, color) &&
606 !drm_mm_hole_follows(node))
613 * i915_vma_insert - finds a slot for the vma in its address space
615 * @size: requested size in bytes (can be larger than the VMA)
616 * @alignment: required alignment
617 * @flags: mask of PIN_* flags to use
619 * First we try to allocate some free space that meets the requirements for
620 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
621 * preferrably the oldest idle entry to make room for the new VMA.
624 * 0 on success, negative error code otherwise.
627 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
633 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
634 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
636 size = max(size, vma->size);
637 alignment = max(alignment, vma->display_alignment);
638 if (flags & PIN_MAPPABLE) {
639 size = max_t(typeof(size), size, vma->fence_size);
640 alignment = max_t(typeof(alignment),
641 alignment, vma->fence_alignment);
644 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
645 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
646 GEM_BUG_ON(!is_power_of_2(alignment));
648 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
649 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
651 end = vma->vm->total;
652 if (flags & PIN_MAPPABLE)
653 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
654 if (flags & PIN_ZONE_4G)
655 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
656 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
658 /* If binding the object/GGTT view requires more space than the entire
659 * aperture has, reject it early before evicting everything in a vain
660 * attempt to find space.
663 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
664 size, flags & PIN_MAPPABLE ? "mappable" : "total",
670 if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
671 color = vma->obj->cache_level;
673 if (flags & PIN_OFFSET_FIXED) {
674 u64 offset = flags & PIN_OFFSET_MASK;
675 if (!IS_ALIGNED(offset, alignment) ||
676 range_overflows(offset, size, end))
679 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
686 * We only support huge gtt pages through the 48b PPGTT,
687 * however we also don't want to force any alignment for
688 * objects which need to be tightly packed into the low 32bits.
690 * Note that we assume that GGTT are limited to 4GiB for the
691 * forseeable future. See also i915_ggtt_offset().
693 if (upper_32_bits(end - 1) &&
694 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
696 * We can't mix 64K and 4K PTEs in the same page-table
697 * (2M block), and so to avoid the ugliness and
698 * complexity of coloring we opt for just aligning 64K
702 rounddown_pow_of_two(vma->page_sizes.sg |
703 I915_GTT_PAGE_SIZE_2M);
706 * Check we don't expand for the limited Global GTT
707 * (mappable aperture is even more precious!). This
708 * also checks that we exclude the aliasing-ppgtt.
710 GEM_BUG_ON(i915_vma_is_ggtt(vma));
712 alignment = max(alignment, page_alignment);
714 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
715 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
718 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
719 size, alignment, color,
724 GEM_BUG_ON(vma->node.start < start);
725 GEM_BUG_ON(vma->node.start + vma->node.size > end);
727 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
728 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
730 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
736 i915_vma_detach(struct i915_vma *vma)
738 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
739 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
742 * And finally now the object is completely decoupled from this
743 * vma, we can drop its hold on the backing storage and allow
744 * it to be reaped by the shrinker.
746 list_del(&vma->vm_link);
749 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
754 bound = atomic_read(&vma->flags);
756 if (unlikely(flags & ~bound))
759 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
762 if (!(bound & I915_VMA_PIN_MASK))
765 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
766 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
772 * If pin_count==0, but we are bound, check under the lock to avoid
773 * racing with a concurrent i915_vma_unbind().
775 mutex_lock(&vma->vm->mutex);
777 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
782 if (unlikely(flags & ~bound)) {
786 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
787 mutex_unlock(&vma->vm->mutex);
792 static int vma_get_pages(struct i915_vma *vma)
795 bool pinned_pages = false;
797 if (atomic_add_unless(&vma->pages_count, 1, 0))
801 err = i915_gem_object_pin_pages(vma->obj);
807 /* Allocations ahoy! */
808 if (mutex_lock_interruptible(&vma->pages_mutex)) {
813 if (!atomic_read(&vma->pages_count)) {
814 err = vma->ops->set_pages(vma);
817 pinned_pages = false;
819 atomic_inc(&vma->pages_count);
822 mutex_unlock(&vma->pages_mutex);
825 __i915_gem_object_unpin_pages(vma->obj);
830 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
832 /* We allocate under vma_get_pages, so beware the shrinker */
833 mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
834 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
835 if (atomic_sub_return(count, &vma->pages_count) == 0) {
836 vma->ops->clear_pages(vma);
837 GEM_BUG_ON(vma->pages);
839 i915_gem_object_unpin_pages(vma->obj);
841 mutex_unlock(&vma->pages_mutex);
844 static void vma_put_pages(struct i915_vma *vma)
846 if (atomic_add_unless(&vma->pages_count, -1, 1))
849 __vma_put_pages(vma, 1);
852 static void vma_unbind_pages(struct i915_vma *vma)
856 lockdep_assert_held(&vma->vm->mutex);
858 /* The upper portion of pages_count is the number of bindings */
859 count = atomic_read(&vma->pages_count);
860 count >>= I915_VMA_PAGES_BIAS;
863 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
866 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
867 u64 size, u64 alignment, u64 flags)
869 struct i915_vma_work *work = NULL;
870 intel_wakeref_t wakeref = 0;
874 #ifdef CONFIG_PROVE_LOCKING
875 if (debug_locks && !WARN_ON(!ww) && vma->resv)
876 assert_vma_held(vma);
879 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
880 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
882 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
884 /* First try and grab the pin without rebinding the vma */
885 if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
888 err = vma_get_pages(vma);
892 if (flags & PIN_GLOBAL)
893 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
895 if (flags & vma->vm->bind_async_flags) {
897 err = i915_vm_lock_objects(vma->vm, ww);
901 work = i915_vma_work();
907 work->vm = i915_vm_get(vma->vm);
909 /* Allocate enough page directories to used PTE */
910 if (vma->vm->allocate_va_range) {
911 err = i915_vm_alloc_pt_stash(vma->vm,
917 err = i915_vm_map_pt_stash(vma->vm, &work->stash);
924 * Differentiate between user/kernel vma inside the aliasing-ppgtt.
926 * We conflate the Global GTT with the user's vma when using the
927 * aliasing-ppgtt, but it is still vitally important to try and
928 * keep the use cases distinct. For example, userptr objects are
929 * not allowed inside the Global GTT as that will cause lock
930 * inversions when we have to evict them the mmu_notifier callbacks -
931 * but they are allowed to be part of the user ppGTT which can never
932 * be mapped. As such we try to give the distinct users of the same
933 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
934 * and i915_ppgtt separate].
936 * NB this may cause us to mask real lock inversions -- while the
937 * code is safe today, lockdep may not be able to spot future
940 err = mutex_lock_interruptible_nested(&vma->vm->mutex,
941 !(flags & PIN_GLOBAL));
945 /* No more allocations allowed now we hold vm->mutex */
947 if (unlikely(i915_vma_is_closed(vma))) {
952 bound = atomic_read(&vma->flags);
953 if (unlikely(bound & I915_VMA_ERROR)) {
958 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
959 err = -EAGAIN; /* pins are meant to be fairly temporary */
963 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
968 err = i915_active_acquire(&vma->active);
972 if (!(bound & I915_VMA_BIND_MASK)) {
973 err = i915_vma_insert(vma, size, alignment, flags);
977 if (i915_is_ggtt(vma->vm))
978 __i915_vma_set_map_and_fenceable(vma);
981 GEM_BUG_ON(!vma->pages);
982 err = i915_vma_bind(vma,
983 vma->obj ? vma->obj->cache_level : 0,
988 /* There should only be at most 2 active bindings (user, global) */
989 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
990 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
991 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
994 GEM_BUG_ON(!i915_vma_is_pinned(vma));
995 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
996 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
999 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
1000 i915_vma_detach(vma);
1001 drm_mm_remove_node(&vma->node);
1004 i915_active_release(&vma->active);
1006 mutex_unlock(&vma->vm->mutex);
1009 dma_fence_work_commit_imm(&work->base);
1012 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
1017 static void flush_idle_contexts(struct intel_gt *gt)
1019 struct intel_engine_cs *engine;
1020 enum intel_engine_id id;
1022 for_each_engine(engine, gt, id)
1023 intel_engine_flush_barriers(engine);
1025 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1028 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1029 u32 align, unsigned int flags)
1031 struct i915_address_space *vm = vma->vm;
1034 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1036 #ifdef CONFIG_LOCKDEP
1037 WARN_ON(!ww && vma->resv && dma_resv_held(vma->resv));
1042 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);
1044 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1045 if (err != -ENOSPC) {
1047 err = i915_vma_wait_for_bind(vma);
1049 i915_vma_unpin(vma);
1054 /* Unlike i915_vma_pin, we don't take no for an answer! */
1055 flush_idle_contexts(vm->gt);
1056 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1057 i915_gem_evict_vm(vm);
1058 mutex_unlock(&vm->mutex);
1063 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1066 * We defer actually closing, unbinding and destroying the VMA until
1067 * the next idle point, or if the object is freed in the meantime. By
1068 * postponing the unbind, we allow for it to be resurrected by the
1069 * client, avoiding the work required to rebind the VMA. This is
1070 * advantageous for DRI, where the client/server pass objects
1071 * between themselves, temporarily opening a local VMA to the
1072 * object, and then closing it again. The same object is then reused
1073 * on the next frame (or two, depending on the depth of the swap queue)
1074 * causing us to rebind the VMA once more. This ends up being a lot
1075 * of wasted work for the steady state.
1077 GEM_BUG_ON(i915_vma_is_closed(vma));
1078 list_add(&vma->closed_link, >->closed_vma);
1081 void i915_vma_close(struct i915_vma *vma)
1083 struct intel_gt *gt = vma->vm->gt;
1084 unsigned long flags;
1086 if (i915_vma_is_ggtt(vma))
1089 GEM_BUG_ON(!atomic_read(&vma->open_count));
1090 if (atomic_dec_and_lock_irqsave(&vma->open_count,
1093 __vma_close(vma, gt);
1094 spin_unlock_irqrestore(>->closed_lock, flags);
1098 static void __i915_vma_remove_closed(struct i915_vma *vma)
1100 struct intel_gt *gt = vma->vm->gt;
1102 spin_lock_irq(>->closed_lock);
1103 list_del_init(&vma->closed_link);
1104 spin_unlock_irq(>->closed_lock);
1107 void i915_vma_reopen(struct i915_vma *vma)
1109 if (i915_vma_is_closed(vma))
1110 __i915_vma_remove_closed(vma);
1113 void i915_vma_release(struct kref *ref)
1115 struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1117 if (drm_mm_node_allocated(&vma->node)) {
1118 mutex_lock(&vma->vm->mutex);
1119 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1120 WARN_ON(__i915_vma_unbind(vma));
1121 mutex_unlock(&vma->vm->mutex);
1122 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1124 GEM_BUG_ON(i915_vma_is_active(vma));
1127 struct drm_i915_gem_object *obj = vma->obj;
1129 spin_lock(&obj->vma.lock);
1130 list_del(&vma->obj_link);
1131 if (!RB_EMPTY_NODE(&vma->obj_node))
1132 rb_erase(&vma->obj_node, &obj->vma.tree);
1133 spin_unlock(&obj->vma.lock);
1136 __i915_vma_remove_closed(vma);
1137 i915_vm_put(vma->vm);
1139 i915_active_fini(&vma->active);
1143 void i915_vma_parked(struct intel_gt *gt)
1145 struct i915_vma *vma, *next;
1148 spin_lock_irq(>->closed_lock);
1149 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1150 struct drm_i915_gem_object *obj = vma->obj;
1151 struct i915_address_space *vm = vma->vm;
1153 /* XXX All to avoid keeping a reference on i915_vma itself */
1155 if (!kref_get_unless_zero(&obj->base.refcount))
1158 if (!i915_vm_tryopen(vm)) {
1159 i915_gem_object_put(obj);
1163 list_move(&vma->closed_link, &closed);
1165 spin_unlock_irq(>->closed_lock);
1167 /* As the GT is held idle, no vma can be reopened as we destroy them */
1168 list_for_each_entry_safe(vma, next, &closed, closed_link) {
1169 struct drm_i915_gem_object *obj = vma->obj;
1170 struct i915_address_space *vm = vma->vm;
1172 INIT_LIST_HEAD(&vma->closed_link);
1173 __i915_vma_put(vma);
1175 i915_gem_object_put(obj);
1180 static void __i915_vma_iounmap(struct i915_vma *vma)
1182 GEM_BUG_ON(i915_vma_is_pinned(vma));
1184 if (vma->iomap == NULL)
1187 io_mapping_unmap(vma->iomap);
1191 void i915_vma_revoke_mmap(struct i915_vma *vma)
1193 struct drm_vma_offset_node *node;
1196 if (!i915_vma_has_userfault(vma))
1199 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1200 GEM_BUG_ON(!vma->obj->userfault_count);
1202 node = &vma->mmo->vma_node;
1203 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1204 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1205 drm_vma_node_offset_addr(node) + vma_offset,
1209 i915_vma_unset_userfault(vma);
1210 if (!--vma->obj->userfault_count)
1211 list_del(&vma->obj->userfault_link);
1215 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1217 return __i915_request_await_exclusive(rq, &vma->active);
1220 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1224 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1226 /* Wait for the vma to be bound before we start! */
1227 err = __i915_request_await_bind(rq, vma);
1231 return i915_active_add_request(&vma->active, rq);
1234 int _i915_vma_move_to_active(struct i915_vma *vma,
1235 struct i915_request *rq,
1236 struct dma_fence *fence,
1239 struct drm_i915_gem_object *obj = vma->obj;
1242 assert_object_held(obj);
1244 err = __i915_vma_move_to_active(vma, rq);
1248 if (flags & EXEC_OBJECT_WRITE) {
1249 struct intel_frontbuffer *front;
1251 front = __intel_frontbuffer_get(obj);
1252 if (unlikely(front)) {
1253 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1254 i915_active_add_request(&front->write, rq);
1255 intel_frontbuffer_put(front);
1259 dma_resv_add_excl_fence(vma->resv, fence);
1260 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1261 obj->read_domains = 0;
1264 if (!(flags & __EXEC_OBJECT_NO_RESERVE)) {
1265 err = dma_resv_reserve_shared(vma->resv, 1);
1271 dma_resv_add_shared_fence(vma->resv, fence);
1272 obj->write_domain = 0;
1276 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1277 i915_active_add_request(&vma->fence->active, rq);
1279 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1280 obj->mm.dirty = true;
1282 GEM_BUG_ON(!i915_vma_is_active(vma));
1286 void __i915_vma_evict(struct i915_vma *vma)
1288 GEM_BUG_ON(i915_vma_is_pinned(vma));
1290 if (i915_vma_is_map_and_fenceable(vma)) {
1291 /* Force a pagefault for domain tracking on next user access */
1292 i915_vma_revoke_mmap(vma);
1295 * Check that we have flushed all writes through the GGTT
1296 * before the unbind, other due to non-strict nature of those
1297 * indirect writes they may end up referencing the GGTT PTE
1300 * Note that we may be concurrently poking at the GGTT_WRITE
1301 * bit from set-domain, as we mark all GGTT vma associated
1302 * with an object. We know this is for another vma, as we
1303 * are currently unbinding this one -- so if this vma will be
1304 * reused, it will be refaulted and have its dirty bit set
1305 * before the next write.
1307 i915_vma_flush_writes(vma);
1309 /* release the fence reg _after_ flushing */
1310 i915_vma_revoke_fence(vma);
1312 __i915_vma_iounmap(vma);
1313 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1315 GEM_BUG_ON(vma->fence);
1316 GEM_BUG_ON(i915_vma_has_userfault(vma));
1318 if (likely(atomic_read(&vma->vm->open))) {
1319 trace_i915_vma_unbind(vma);
1320 vma->ops->unbind_vma(vma->vm, vma);
1322 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1325 i915_vma_detach(vma);
1326 vma_unbind_pages(vma);
1329 int __i915_vma_unbind(struct i915_vma *vma)
1333 lockdep_assert_held(&vma->vm->mutex);
1335 if (!drm_mm_node_allocated(&vma->node))
1338 if (i915_vma_is_pinned(vma)) {
1339 vma_print_allocator(vma, "is pinned");
1344 * After confirming that no one else is pinning this vma, wait for
1345 * any laggards who may have crept in during the wait (through
1346 * a residual pin skipping the vm->mutex) to complete.
1348 ret = i915_vma_sync(vma);
1352 GEM_BUG_ON(i915_vma_is_active(vma));
1353 __i915_vma_evict(vma);
1355 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1359 int i915_vma_unbind(struct i915_vma *vma)
1361 struct i915_address_space *vm = vma->vm;
1362 intel_wakeref_t wakeref = 0;
1365 /* Optimistic wait before taking the mutex */
1366 err = i915_vma_sync(vma);
1370 if (!drm_mm_node_allocated(&vma->node))
1373 if (i915_vma_is_pinned(vma)) {
1374 vma_print_allocator(vma, "is pinned");
1378 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1379 /* XXX not always required: nop_clear_range */
1380 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1382 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1386 err = __i915_vma_unbind(vma);
1387 mutex_unlock(&vm->mutex);
1391 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1395 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1397 i915_gem_object_make_unshrinkable(vma->obj);
1401 void i915_vma_make_shrinkable(struct i915_vma *vma)
1403 i915_gem_object_make_shrinkable(vma->obj);
1406 void i915_vma_make_purgeable(struct i915_vma *vma)
1408 i915_gem_object_make_purgeable(vma->obj);
1411 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1412 #include "selftests/i915_vma.c"
1415 void i915_vma_module_exit(void)
1417 kmem_cache_destroy(slab_vmas);
1420 int __init i915_vma_module_init(void)
1422 slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);