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 "gt/intel_engine.h"
31 #include "gt/intel_engine_heartbeat.h"
32 #include "gt/intel_gt.h"
33 #include "gt/intel_gt_requests.h"
36 #include "i915_globals.h"
37 #include "i915_sw_fence_work.h"
38 #include "i915_trace.h"
41 static struct i915_global_vma {
42 struct i915_global base;
43 struct kmem_cache *slab_vmas;
46 struct i915_vma *i915_vma_alloc(void)
48 return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
51 void i915_vma_free(struct i915_vma *vma)
53 return kmem_cache_free(global.slab_vmas, vma);
56 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
58 #include <linux/stackdepot.h>
60 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
62 unsigned long *entries;
63 unsigned int nr_entries;
66 if (!vma->node.stack) {
67 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
68 vma->node.start, vma->node.size, reason);
72 nr_entries = stack_depot_fetch(vma->node.stack, &entries);
73 stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
74 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
75 vma->node.start, vma->node.size, reason, buf);
80 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
86 static inline struct i915_vma *active_to_vma(struct i915_active *ref)
88 return container_of(ref, typeof(struct i915_vma), active);
91 static int __i915_vma_active(struct i915_active *ref)
93 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
97 static void __i915_vma_retire(struct i915_active *ref)
99 i915_vma_put(active_to_vma(ref));
102 static struct i915_vma *
103 vma_create(struct drm_i915_gem_object *obj,
104 struct i915_address_space *vm,
105 const struct i915_ggtt_view *view)
107 struct i915_vma *vma;
108 struct rb_node *rb, **p;
110 /* The aliasing_ppgtt should never be used directly! */
111 GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
113 vma = i915_vma_alloc();
115 return ERR_PTR(-ENOMEM);
117 kref_init(&vma->ref);
118 mutex_init(&vma->pages_mutex);
119 vma->vm = i915_vm_get(vm);
120 vma->ops = &vm->vma_ops;
122 vma->resv = obj->base.resv;
123 vma->size = obj->base.size;
124 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
126 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire);
128 /* Declare ourselves safe for use inside shrinkers */
129 if (IS_ENABLED(CONFIG_LOCKDEP)) {
130 fs_reclaim_acquire(GFP_KERNEL);
131 might_lock(&vma->active.mutex);
132 fs_reclaim_release(GFP_KERNEL);
135 INIT_LIST_HEAD(&vma->closed_link);
137 if (view && view->type != I915_GGTT_VIEW_NORMAL) {
138 vma->ggtt_view = *view;
139 if (view->type == I915_GGTT_VIEW_PARTIAL) {
140 GEM_BUG_ON(range_overflows_t(u64,
141 view->partial.offset,
143 obj->base.size >> PAGE_SHIFT));
144 vma->size = view->partial.size;
145 vma->size <<= PAGE_SHIFT;
146 GEM_BUG_ON(vma->size > obj->base.size);
147 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
148 vma->size = intel_rotation_info_size(&view->rotated);
149 vma->size <<= PAGE_SHIFT;
150 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
151 vma->size = intel_remapped_info_size(&view->remapped);
152 vma->size <<= PAGE_SHIFT;
156 if (unlikely(vma->size > vm->total))
159 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
161 if (i915_is_ggtt(vm)) {
162 if (unlikely(overflows_type(vma->size, u32)))
165 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
166 i915_gem_object_get_tiling(obj),
167 i915_gem_object_get_stride(obj));
168 if (unlikely(vma->fence_size < vma->size || /* overflow */
169 vma->fence_size > vm->total))
172 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
174 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
175 i915_gem_object_get_tiling(obj),
176 i915_gem_object_get_stride(obj));
177 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
179 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
182 spin_lock(&obj->vma.lock);
185 p = &obj->vma.tree.rb_node;
187 struct i915_vma *pos;
191 pos = rb_entry(rb, struct i915_vma, obj_node);
194 * If the view already exists in the tree, another thread
195 * already created a matching vma, so return the older instance
196 * and dispose of ours.
198 cmp = i915_vma_compare(pos, vm, view);
200 spin_unlock(&obj->vma.lock);
210 rb_link_node(&vma->obj_node, rb, p);
211 rb_insert_color(&vma->obj_node, &obj->vma.tree);
213 if (i915_vma_is_ggtt(vma))
215 * We put the GGTT vma at the start of the vma-list, followed
216 * by the ppGGTT vma. This allows us to break early when
217 * iterating over only the GGTT vma for an object, see
218 * for_each_ggtt_vma()
220 list_add(&vma->obj_link, &obj->vma.list);
222 list_add_tail(&vma->obj_link, &obj->vma.list);
224 spin_unlock(&obj->vma.lock);
230 return ERR_PTR(-E2BIG);
233 static struct i915_vma *
234 vma_lookup(struct drm_i915_gem_object *obj,
235 struct i915_address_space *vm,
236 const struct i915_ggtt_view *view)
240 rb = obj->vma.tree.rb_node;
242 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
245 cmp = i915_vma_compare(vma, vm, view);
259 * i915_vma_instance - return the singleton instance of the VMA
260 * @obj: parent &struct drm_i915_gem_object to be mapped
261 * @vm: address space in which the mapping is located
262 * @view: additional mapping requirements
264 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
265 * the same @view characteristics. If a match is not found, one is created.
266 * Once created, the VMA is kept until either the object is freed, or the
267 * address space is closed.
269 * Returns the vma, or an error pointer.
272 i915_vma_instance(struct drm_i915_gem_object *obj,
273 struct i915_address_space *vm,
274 const struct i915_ggtt_view *view)
276 struct i915_vma *vma;
278 GEM_BUG_ON(view && !i915_is_ggtt(vm));
279 GEM_BUG_ON(!atomic_read(&vm->open));
281 spin_lock(&obj->vma.lock);
282 vma = vma_lookup(obj, vm, view);
283 spin_unlock(&obj->vma.lock);
285 /* vma_create() will resolve the race if another creates the vma */
287 vma = vma_create(obj, vm, view);
289 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
293 struct i915_vma_work {
294 struct dma_fence_work base;
295 struct i915_vma *vma;
296 struct drm_i915_gem_object *pinned;
297 struct i915_sw_dma_fence_cb cb;
298 enum i915_cache_level cache_level;
302 static int __vma_bind(struct dma_fence_work *work)
304 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
305 struct i915_vma *vma = vw->vma;
308 err = vma->ops->bind_vma(vma, vw->cache_level, vw->flags);
310 atomic_or(I915_VMA_ERROR, &vma->flags);
315 static void __vma_release(struct dma_fence_work *work)
317 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
320 __i915_gem_object_unpin_pages(vw->pinned);
323 static const struct dma_fence_work_ops bind_ops = {
326 .release = __vma_release,
329 struct i915_vma_work *i915_vma_work(void)
331 struct i915_vma_work *vw;
333 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
337 dma_fence_work_init(&vw->base, &bind_ops);
338 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
343 int i915_vma_wait_for_bind(struct i915_vma *vma)
347 if (rcu_access_pointer(vma->active.excl.fence)) {
348 struct dma_fence *fence;
351 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
354 err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
355 dma_fence_put(fence);
363 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
365 * @cache_level: mapping cache level
366 * @flags: flags like global or local mapping
367 * @work: preallocated worker for allocating and binding the PTE
369 * DMA addresses are taken from the scatter-gather table of this object (or of
370 * this VMA in case of non-default GGTT views) and PTE entries set up.
371 * Note that DMA addresses are also the only part of the SG table we care about.
373 int i915_vma_bind(struct i915_vma *vma,
374 enum i915_cache_level cache_level,
376 struct i915_vma_work *work)
382 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
383 GEM_BUG_ON(vma->size > vma->node.size);
385 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
390 if (GEM_DEBUG_WARN_ON(!flags))
394 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
396 vma_flags = atomic_read(&vma->flags);
397 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
398 if (flags & PIN_UPDATE)
399 bind_flags |= vma_flags;
401 bind_flags &= ~vma_flags;
405 GEM_BUG_ON(!vma->pages);
407 trace_i915_vma_bind(vma, bind_flags);
408 if (work && (bind_flags & ~vma_flags) & vma->vm->bind_async_flags) {
409 struct dma_fence *prev;
412 work->cache_level = cache_level;
413 work->flags = bind_flags | I915_VMA_ALLOC;
416 * Note we only want to chain up to the migration fence on
417 * the pages (not the object itself). As we don't track that,
418 * yet, we have to use the exclusive fence instead.
420 * Also note that we do not want to track the async vma as
421 * part of the obj->resv->excl_fence as it only affects
422 * execution and not content or object's backing store lifetime.
424 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
426 __i915_sw_fence_await_dma_fence(&work->base.chain,
432 work->base.dma.error = 0; /* enable the queue_work() */
435 __i915_gem_object_pin_pages(vma->obj);
436 work->pinned = vma->obj;
439 ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
444 atomic_or(bind_flags, &vma->flags);
448 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
453 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
458 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
459 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
461 ptr = READ_ONCE(vma->iomap);
463 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
471 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
472 io_mapping_unmap(ptr);
479 err = i915_vma_pin_fence(vma);
483 i915_vma_set_ggtt_write(vma);
485 /* NB Access through the GTT requires the device to be awake. */
489 __i915_vma_unpin(vma);
491 return IO_ERR_PTR(err);
494 void i915_vma_flush_writes(struct i915_vma *vma)
496 if (i915_vma_unset_ggtt_write(vma))
497 intel_gt_flush_ggtt_writes(vma->vm->gt);
500 void i915_vma_unpin_iomap(struct i915_vma *vma)
502 GEM_BUG_ON(vma->iomap == NULL);
504 i915_vma_flush_writes(vma);
506 i915_vma_unpin_fence(vma);
510 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
512 struct i915_vma *vma;
513 struct drm_i915_gem_object *obj;
515 vma = fetch_and_zero(p_vma);
525 if (flags & I915_VMA_RELEASE_MAP)
526 i915_gem_object_unpin_map(obj);
528 i915_gem_object_put(obj);
531 bool i915_vma_misplaced(const struct i915_vma *vma,
532 u64 size, u64 alignment, u64 flags)
534 if (!drm_mm_node_allocated(&vma->node))
537 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
540 if (vma->node.size < size)
543 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
544 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
547 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
550 if (flags & PIN_OFFSET_BIAS &&
551 vma->node.start < (flags & PIN_OFFSET_MASK))
554 if (flags & PIN_OFFSET_FIXED &&
555 vma->node.start != (flags & PIN_OFFSET_MASK))
561 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
563 bool mappable, fenceable;
565 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
566 GEM_BUG_ON(!vma->fence_size);
568 fenceable = (vma->node.size >= vma->fence_size &&
569 IS_ALIGNED(vma->node.start, vma->fence_alignment));
571 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
573 if (mappable && fenceable)
574 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
576 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
579 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
581 struct drm_mm_node *node = &vma->node;
582 struct drm_mm_node *other;
585 * On some machines we have to be careful when putting differing types
586 * of snoopable memory together to avoid the prefetcher crossing memory
587 * domains and dying. During vm initialisation, we decide whether or not
588 * these constraints apply and set the drm_mm.color_adjust
591 if (!i915_vm_has_cache_coloring(vma->vm))
594 /* Only valid to be called on an already inserted vma */
595 GEM_BUG_ON(!drm_mm_node_allocated(node));
596 GEM_BUG_ON(list_empty(&node->node_list));
598 other = list_prev_entry(node, node_list);
599 if (i915_node_color_differs(other, color) &&
600 !drm_mm_hole_follows(other))
603 other = list_next_entry(node, node_list);
604 if (i915_node_color_differs(other, color) &&
605 !drm_mm_hole_follows(node))
611 static void assert_bind_count(const struct drm_i915_gem_object *obj)
614 * Combine the assertion that the object is bound and that we have
615 * pinned its pages. But we should never have bound the object
616 * more than we have pinned its pages. (For complete accuracy, we
617 * assume that no else is pinning the pages, but as a rough assertion
618 * that we will not run into problems later, this will do!)
620 GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count));
624 * i915_vma_insert - finds a slot for the vma in its address space
626 * @size: requested size in bytes (can be larger than the VMA)
627 * @alignment: required alignment
628 * @flags: mask of PIN_* flags to use
630 * First we try to allocate some free space that meets the requirements for
631 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
632 * preferrably the oldest idle entry to make room for the new VMA.
635 * 0 on success, negative error code otherwise.
638 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
644 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
645 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
647 size = max(size, vma->size);
648 alignment = max(alignment, vma->display_alignment);
649 if (flags & PIN_MAPPABLE) {
650 size = max_t(typeof(size), size, vma->fence_size);
651 alignment = max_t(typeof(alignment),
652 alignment, vma->fence_alignment);
655 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
656 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
657 GEM_BUG_ON(!is_power_of_2(alignment));
659 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
660 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
662 end = vma->vm->total;
663 if (flags & PIN_MAPPABLE)
664 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
665 if (flags & PIN_ZONE_4G)
666 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
667 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
669 /* If binding the object/GGTT view requires more space than the entire
670 * aperture has, reject it early before evicting everything in a vain
671 * attempt to find space.
674 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
675 size, flags & PIN_MAPPABLE ? "mappable" : "total",
681 if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
682 color = vma->obj->cache_level;
684 if (flags & PIN_OFFSET_FIXED) {
685 u64 offset = flags & PIN_OFFSET_MASK;
686 if (!IS_ALIGNED(offset, alignment) ||
687 range_overflows(offset, size, end))
690 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
697 * We only support huge gtt pages through the 48b PPGTT,
698 * however we also don't want to force any alignment for
699 * objects which need to be tightly packed into the low 32bits.
701 * Note that we assume that GGTT are limited to 4GiB for the
702 * forseeable future. See also i915_ggtt_offset().
704 if (upper_32_bits(end - 1) &&
705 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
707 * We can't mix 64K and 4K PTEs in the same page-table
708 * (2M block), and so to avoid the ugliness and
709 * complexity of coloring we opt for just aligning 64K
713 rounddown_pow_of_two(vma->page_sizes.sg |
714 I915_GTT_PAGE_SIZE_2M);
717 * Check we don't expand for the limited Global GTT
718 * (mappable aperture is even more precious!). This
719 * also checks that we exclude the aliasing-ppgtt.
721 GEM_BUG_ON(i915_vma_is_ggtt(vma));
723 alignment = max(alignment, page_alignment);
725 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
726 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
729 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
730 size, alignment, color,
735 GEM_BUG_ON(vma->node.start < start);
736 GEM_BUG_ON(vma->node.start + vma->node.size > end);
738 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
739 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
742 struct drm_i915_gem_object *obj = vma->obj;
744 atomic_inc(&obj->bind_count);
745 assert_bind_count(obj);
747 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
753 i915_vma_detach(struct i915_vma *vma)
755 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
756 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
759 * And finally now the object is completely decoupled from this
760 * vma, we can drop its hold on the backing storage and allow
761 * it to be reaped by the shrinker.
763 list_del(&vma->vm_link);
765 struct drm_i915_gem_object *obj = vma->obj;
767 assert_bind_count(obj);
768 atomic_dec(&obj->bind_count);
772 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
777 bound = atomic_read(&vma->flags);
779 if (unlikely(flags & ~bound))
782 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
785 if (!(bound & I915_VMA_PIN_MASK))
788 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
789 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
795 * If pin_count==0, but we are bound, check under the lock to avoid
796 * racing with a concurrent i915_vma_unbind().
798 mutex_lock(&vma->vm->mutex);
800 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
805 if (unlikely(flags & ~bound)) {
809 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
810 mutex_unlock(&vma->vm->mutex);
815 static int vma_get_pages(struct i915_vma *vma)
819 if (atomic_add_unless(&vma->pages_count, 1, 0))
822 /* Allocations ahoy! */
823 if (mutex_lock_interruptible(&vma->pages_mutex))
826 if (!atomic_read(&vma->pages_count)) {
828 err = i915_gem_object_pin_pages(vma->obj);
833 err = vma->ops->set_pages(vma);
836 i915_gem_object_unpin_pages(vma->obj);
840 atomic_inc(&vma->pages_count);
843 mutex_unlock(&vma->pages_mutex);
848 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
850 /* We allocate under vma_get_pages, so beware the shrinker */
851 mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
852 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
853 if (atomic_sub_return(count, &vma->pages_count) == 0) {
854 vma->ops->clear_pages(vma);
855 GEM_BUG_ON(vma->pages);
857 i915_gem_object_unpin_pages(vma->obj);
859 mutex_unlock(&vma->pages_mutex);
862 static void vma_put_pages(struct i915_vma *vma)
864 if (atomic_add_unless(&vma->pages_count, -1, 1))
867 __vma_put_pages(vma, 1);
870 static void vma_unbind_pages(struct i915_vma *vma)
874 lockdep_assert_held(&vma->vm->mutex);
876 /* The upper portion of pages_count is the number of bindings */
877 count = atomic_read(&vma->pages_count);
878 count >>= I915_VMA_PAGES_BIAS;
881 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
884 int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
886 struct i915_vma_work *work = NULL;
887 intel_wakeref_t wakeref = 0;
891 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
892 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
894 GEM_BUG_ON(flags & PIN_UPDATE);
895 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
897 /* First try and grab the pin without rebinding the vma */
898 if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
901 err = vma_get_pages(vma);
905 if (flags & vma->vm->bind_async_flags) {
906 work = i915_vma_work();
913 if (flags & PIN_GLOBAL)
914 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
916 /* No more allocations allowed once we hold vm->mutex */
917 err = mutex_lock_interruptible(&vma->vm->mutex);
921 if (unlikely(i915_vma_is_closed(vma))) {
926 bound = atomic_read(&vma->flags);
927 if (unlikely(bound & I915_VMA_ERROR)) {
932 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
933 err = -EAGAIN; /* pins are meant to be fairly temporary */
937 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
942 err = i915_active_acquire(&vma->active);
946 if (!(bound & I915_VMA_BIND_MASK)) {
947 err = i915_vma_insert(vma, size, alignment, flags);
951 if (i915_is_ggtt(vma->vm))
952 __i915_vma_set_map_and_fenceable(vma);
955 GEM_BUG_ON(!vma->pages);
956 err = i915_vma_bind(vma,
957 vma->obj ? vma->obj->cache_level : 0,
962 /* There should only be at most 2 active bindings (user, global) */
963 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
964 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
965 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
968 GEM_BUG_ON(!i915_vma_is_pinned(vma));
969 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
970 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
973 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
974 i915_vma_detach(vma);
975 drm_mm_remove_node(&vma->node);
978 i915_active_release(&vma->active);
980 mutex_unlock(&vma->vm->mutex);
983 dma_fence_work_commit(&work->base);
985 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
991 static void flush_idle_contexts(struct intel_gt *gt)
993 struct intel_engine_cs *engine;
994 enum intel_engine_id id;
996 for_each_engine(engine, gt, id)
997 intel_engine_flush_barriers(engine);
999 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1002 int i915_ggtt_pin(struct i915_vma *vma, u32 align, unsigned int flags)
1004 struct i915_address_space *vm = vma->vm;
1007 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1010 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1011 if (err != -ENOSPC) {
1013 err = i915_vma_wait_for_bind(vma);
1015 i915_vma_unpin(vma);
1020 /* Unlike i915_vma_pin, we don't take no for an answer! */
1021 flush_idle_contexts(vm->gt);
1022 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1023 i915_gem_evict_vm(vm);
1024 mutex_unlock(&vm->mutex);
1029 void i915_vma_close(struct i915_vma *vma)
1031 struct intel_gt *gt = vma->vm->gt;
1032 unsigned long flags;
1034 GEM_BUG_ON(i915_vma_is_closed(vma));
1037 * We defer actually closing, unbinding and destroying the VMA until
1038 * the next idle point, or if the object is freed in the meantime. By
1039 * postponing the unbind, we allow for it to be resurrected by the
1040 * client, avoiding the work required to rebind the VMA. This is
1041 * advantageous for DRI, where the client/server pass objects
1042 * between themselves, temporarily opening a local VMA to the
1043 * object, and then closing it again. The same object is then reused
1044 * on the next frame (or two, depending on the depth of the swap queue)
1045 * causing us to rebind the VMA once more. This ends up being a lot
1046 * of wasted work for the steady state.
1048 spin_lock_irqsave(>->closed_lock, flags);
1049 list_add(&vma->closed_link, >->closed_vma);
1050 spin_unlock_irqrestore(>->closed_lock, flags);
1053 static void __i915_vma_remove_closed(struct i915_vma *vma)
1055 struct intel_gt *gt = vma->vm->gt;
1057 spin_lock_irq(>->closed_lock);
1058 list_del_init(&vma->closed_link);
1059 spin_unlock_irq(>->closed_lock);
1062 void i915_vma_reopen(struct i915_vma *vma)
1064 if (i915_vma_is_closed(vma))
1065 __i915_vma_remove_closed(vma);
1068 void i915_vma_release(struct kref *ref)
1070 struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1072 if (drm_mm_node_allocated(&vma->node)) {
1073 mutex_lock(&vma->vm->mutex);
1074 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1075 WARN_ON(__i915_vma_unbind(vma));
1076 mutex_unlock(&vma->vm->mutex);
1077 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1079 GEM_BUG_ON(i915_vma_is_active(vma));
1082 struct drm_i915_gem_object *obj = vma->obj;
1084 spin_lock(&obj->vma.lock);
1085 list_del(&vma->obj_link);
1086 rb_erase(&vma->obj_node, &obj->vma.tree);
1087 spin_unlock(&obj->vma.lock);
1090 __i915_vma_remove_closed(vma);
1091 i915_vm_put(vma->vm);
1093 i915_active_fini(&vma->active);
1097 void i915_vma_parked(struct intel_gt *gt)
1099 struct i915_vma *vma, *next;
1102 spin_lock_irq(>->closed_lock);
1103 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1104 struct drm_i915_gem_object *obj = vma->obj;
1105 struct i915_address_space *vm = vma->vm;
1107 /* XXX All to avoid keeping a reference on i915_vma itself */
1109 if (!kref_get_unless_zero(&obj->base.refcount))
1112 if (!i915_vm_tryopen(vm)) {
1113 i915_gem_object_put(obj);
1117 list_move(&vma->closed_link, &closed);
1119 spin_unlock_irq(>->closed_lock);
1121 /* As the GT is held idle, no vma can be reopened as we destroy them */
1122 list_for_each_entry_safe(vma, next, &closed, closed_link) {
1123 struct drm_i915_gem_object *obj = vma->obj;
1124 struct i915_address_space *vm = vma->vm;
1126 INIT_LIST_HEAD(&vma->closed_link);
1127 __i915_vma_put(vma);
1129 i915_gem_object_put(obj);
1134 static void __i915_vma_iounmap(struct i915_vma *vma)
1136 GEM_BUG_ON(i915_vma_is_pinned(vma));
1138 if (vma->iomap == NULL)
1141 io_mapping_unmap(vma->iomap);
1145 void i915_vma_revoke_mmap(struct i915_vma *vma)
1147 struct drm_vma_offset_node *node;
1150 if (!i915_vma_has_userfault(vma))
1153 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1154 GEM_BUG_ON(!vma->obj->userfault_count);
1156 node = &vma->mmo->vma_node;
1157 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1158 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1159 drm_vma_node_offset_addr(node) + vma_offset,
1163 i915_vma_unset_userfault(vma);
1164 if (!--vma->obj->userfault_count)
1165 list_del(&vma->obj->userfault_link);
1168 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1172 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1174 /* Wait for the vma to be bound before we start! */
1175 err = i915_request_await_active(rq, &vma->active, 0);
1179 return i915_active_add_request(&vma->active, rq);
1182 int i915_vma_move_to_active(struct i915_vma *vma,
1183 struct i915_request *rq,
1186 struct drm_i915_gem_object *obj = vma->obj;
1189 assert_object_held(obj);
1191 err = __i915_vma_move_to_active(vma, rq);
1195 if (flags & EXEC_OBJECT_WRITE) {
1196 struct intel_frontbuffer *front;
1198 front = __intel_frontbuffer_get(obj);
1199 if (unlikely(front)) {
1200 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1201 i915_active_add_request(&front->write, rq);
1202 intel_frontbuffer_put(front);
1205 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1206 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1207 obj->read_domains = 0;
1209 err = dma_resv_reserve_shared(vma->resv, 1);
1213 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1214 obj->write_domain = 0;
1216 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1217 obj->mm.dirty = true;
1219 GEM_BUG_ON(!i915_vma_is_active(vma));
1223 int __i915_vma_unbind(struct i915_vma *vma)
1227 lockdep_assert_held(&vma->vm->mutex);
1230 * First wait upon any activity as retiring the request may
1231 * have side-effects such as unpinning or even unbinding this vma.
1233 * XXX Actually waiting under the vm->mutex is a hinderance and
1234 * should be pipelined wherever possible. In cases where that is
1235 * unavoidable, we should lift the wait to before the mutex.
1237 ret = i915_vma_sync(vma);
1241 if (i915_vma_is_pinned(vma)) {
1242 vma_print_allocator(vma, "is pinned");
1247 * After confirming that no one else is pinning this vma, wait for
1248 * any laggards who may have crept in during the wait (through
1249 * a residual pin skipping the vm->mutex) to complete.
1251 ret = i915_vma_sync(vma);
1255 if (!drm_mm_node_allocated(&vma->node))
1258 GEM_BUG_ON(i915_vma_is_pinned(vma));
1259 GEM_BUG_ON(i915_vma_is_active(vma));
1261 if (i915_vma_is_map_and_fenceable(vma)) {
1263 * Check that we have flushed all writes through the GGTT
1264 * before the unbind, other due to non-strict nature of those
1265 * indirect writes they may end up referencing the GGTT PTE
1268 * Note that we may be concurrently poking at the GGTT_WRITE
1269 * bit from set-domain, as we mark all GGTT vma associated
1270 * with an object. We know this is for another vma, as we
1271 * are currently unbinding this one -- so if this vma will be
1272 * reused, it will be refaulted and have its dirty bit set
1273 * before the next write.
1275 i915_vma_flush_writes(vma);
1277 /* release the fence reg _after_ flushing */
1278 ret = i915_vma_revoke_fence(vma);
1282 /* Force a pagefault for domain tracking on next user access */
1283 i915_vma_revoke_mmap(vma);
1285 __i915_vma_iounmap(vma);
1286 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1288 GEM_BUG_ON(vma->fence);
1289 GEM_BUG_ON(i915_vma_has_userfault(vma));
1291 if (likely(atomic_read(&vma->vm->open))) {
1292 trace_i915_vma_unbind(vma);
1293 vma->ops->unbind_vma(vma);
1295 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1298 i915_vma_detach(vma);
1299 vma_unbind_pages(vma);
1301 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1305 int i915_vma_unbind(struct i915_vma *vma)
1307 struct i915_address_space *vm = vma->vm;
1308 intel_wakeref_t wakeref = 0;
1311 if (!drm_mm_node_allocated(&vma->node))
1314 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1315 /* XXX not always required: nop_clear_range */
1316 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1318 /* Optimistic wait before taking the mutex */
1319 err = i915_vma_sync(vma);
1323 err = mutex_lock_interruptible(&vm->mutex);
1327 err = __i915_vma_unbind(vma);
1328 mutex_unlock(&vm->mutex);
1332 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1336 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1338 i915_gem_object_make_unshrinkable(vma->obj);
1342 void i915_vma_make_shrinkable(struct i915_vma *vma)
1344 i915_gem_object_make_shrinkable(vma->obj);
1347 void i915_vma_make_purgeable(struct i915_vma *vma)
1349 i915_gem_object_make_purgeable(vma->obj);
1352 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1353 #include "selftests/i915_vma.c"
1356 static void i915_global_vma_shrink(void)
1358 kmem_cache_shrink(global.slab_vmas);
1361 static void i915_global_vma_exit(void)
1363 kmem_cache_destroy(global.slab_vmas);
1366 static struct i915_global_vma global = { {
1367 .shrink = i915_global_vma_shrink,
1368 .exit = i915_global_vma_exit,
1371 int __init i915_global_vma_init(void)
1373 global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1374 if (!global.slab_vmas)
1377 i915_global_register(&global.base);