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 static struct i915_vma *i915_vma_alloc(void)
45 return kmem_cache_zalloc(slab_vmas, GFP_KERNEL);
48 static 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->size = obj->base.size;
117 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
119 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0);
121 /* Declare ourselves safe for use inside shrinkers */
122 if (IS_ENABLED(CONFIG_LOCKDEP)) {
123 fs_reclaim_acquire(GFP_KERNEL);
124 might_lock(&vma->active.mutex);
125 fs_reclaim_release(GFP_KERNEL);
128 INIT_LIST_HEAD(&vma->closed_link);
130 if (view && view->type != I915_GGTT_VIEW_NORMAL) {
131 vma->ggtt_view = *view;
132 if (view->type == I915_GGTT_VIEW_PARTIAL) {
133 GEM_BUG_ON(range_overflows_t(u64,
134 view->partial.offset,
136 obj->base.size >> PAGE_SHIFT));
137 vma->size = view->partial.size;
138 vma->size <<= PAGE_SHIFT;
139 GEM_BUG_ON(vma->size > obj->base.size);
140 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
141 vma->size = intel_rotation_info_size(&view->rotated);
142 vma->size <<= PAGE_SHIFT;
143 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
144 vma->size = intel_remapped_info_size(&view->remapped);
145 vma->size <<= PAGE_SHIFT;
149 if (unlikely(vma->size > vm->total))
152 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
154 spin_lock(&obj->vma.lock);
156 if (i915_is_ggtt(vm)) {
157 if (unlikely(overflows_type(vma->size, u32)))
160 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
161 i915_gem_object_get_tiling(obj),
162 i915_gem_object_get_stride(obj));
163 if (unlikely(vma->fence_size < vma->size || /* overflow */
164 vma->fence_size > vm->total))
167 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
169 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
170 i915_gem_object_get_tiling(obj),
171 i915_gem_object_get_stride(obj));
172 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
174 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
178 p = &obj->vma.tree.rb_node;
183 pos = rb_entry(rb, struct i915_vma, obj_node);
186 * If the view already exists in the tree, another thread
187 * already created a matching vma, so return the older instance
188 * and dispose of ours.
190 cmp = i915_vma_compare(pos, vm, view);
198 rb_link_node(&vma->obj_node, rb, p);
199 rb_insert_color(&vma->obj_node, &obj->vma.tree);
201 if (i915_vma_is_ggtt(vma))
203 * We put the GGTT vma at the start of the vma-list, followed
204 * by the ppGGTT vma. This allows us to break early when
205 * iterating over only the GGTT vma for an object, see
206 * for_each_ggtt_vma()
208 list_add(&vma->obj_link, &obj->vma.list);
210 list_add_tail(&vma->obj_link, &obj->vma.list);
212 spin_unlock(&obj->vma.lock);
217 spin_unlock(&obj->vma.lock);
224 static struct i915_vma *
225 i915_vma_lookup(struct drm_i915_gem_object *obj,
226 struct i915_address_space *vm,
227 const struct i915_ggtt_view *view)
231 rb = obj->vma.tree.rb_node;
233 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
236 cmp = i915_vma_compare(vma, vm, view);
250 * i915_vma_instance - return the singleton instance of the VMA
251 * @obj: parent &struct drm_i915_gem_object to be mapped
252 * @vm: address space in which the mapping is located
253 * @view: additional mapping requirements
255 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
256 * the same @view characteristics. If a match is not found, one is created.
257 * Once created, the VMA is kept until either the object is freed, or the
258 * address space is closed.
260 * Returns the vma, or an error pointer.
263 i915_vma_instance(struct drm_i915_gem_object *obj,
264 struct i915_address_space *vm,
265 const struct i915_ggtt_view *view)
267 struct i915_vma *vma;
269 GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm));
270 GEM_BUG_ON(!atomic_read(&vm->open));
272 spin_lock(&obj->vma.lock);
273 vma = i915_vma_lookup(obj, vm, view);
274 spin_unlock(&obj->vma.lock);
276 /* vma_create() will resolve the race if another creates the vma */
278 vma = vma_create(obj, vm, view);
280 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
284 struct i915_vma_work {
285 struct dma_fence_work base;
286 struct i915_address_space *vm;
287 struct i915_vm_pt_stash stash;
288 struct i915_vma *vma;
289 struct drm_i915_gem_object *pinned;
290 struct i915_sw_dma_fence_cb cb;
291 enum i915_cache_level cache_level;
295 static void __vma_bind(struct dma_fence_work *work)
297 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
298 struct i915_vma *vma = vw->vma;
300 vma->ops->bind_vma(vw->vm, &vw->stash,
301 vma, vw->cache_level, vw->flags);
304 static void __vma_release(struct dma_fence_work *work)
306 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
309 __i915_gem_object_unpin_pages(vw->pinned);
310 i915_gem_object_put(vw->pinned);
313 i915_vm_free_pt_stash(vw->vm, &vw->stash);
317 static const struct dma_fence_work_ops bind_ops = {
320 .release = __vma_release,
323 struct i915_vma_work *i915_vma_work(void)
325 struct i915_vma_work *vw;
327 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
331 dma_fence_work_init(&vw->base, &bind_ops);
332 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
337 int i915_vma_wait_for_bind(struct i915_vma *vma)
341 if (rcu_access_pointer(vma->active.excl.fence)) {
342 struct dma_fence *fence;
345 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
348 err = dma_fence_wait(fence, true);
349 dma_fence_put(fence);
357 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
359 * @cache_level: mapping cache level
360 * @flags: flags like global or local mapping
361 * @work: preallocated worker for allocating and binding the PTE
363 * DMA addresses are taken from the scatter-gather table of this object (or of
364 * this VMA in case of non-default GGTT views) and PTE entries set up.
365 * Note that DMA addresses are also the only part of the SG table we care about.
367 int i915_vma_bind(struct i915_vma *vma,
368 enum i915_cache_level cache_level,
370 struct i915_vma_work *work)
375 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
376 GEM_BUG_ON(vma->size > vma->node.size);
378 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
383 if (GEM_DEBUG_WARN_ON(!flags))
387 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
389 vma_flags = atomic_read(&vma->flags);
390 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
392 bind_flags &= ~vma_flags;
396 GEM_BUG_ON(!vma->pages);
398 trace_i915_vma_bind(vma, bind_flags);
399 if (work && bind_flags & vma->vm->bind_async_flags) {
400 struct dma_fence *prev;
403 work->cache_level = cache_level;
404 work->flags = bind_flags;
407 * Note we only want to chain up to the migration fence on
408 * the pages (not the object itself). As we don't track that,
409 * yet, we have to use the exclusive fence instead.
411 * Also note that we do not want to track the async vma as
412 * part of the obj->resv->excl_fence as it only affects
413 * execution and not content or object's backing store lifetime.
415 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
417 __i915_sw_fence_await_dma_fence(&work->base.chain,
423 work->base.dma.error = 0; /* enable the queue_work() */
425 __i915_gem_object_pin_pages(vma->obj);
426 work->pinned = i915_gem_object_get(vma->obj);
428 vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
431 atomic_or(bind_flags, &vma->flags);
435 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
440 if (!i915_gem_object_is_lmem(vma->obj)) {
441 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
447 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
448 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
450 ptr = READ_ONCE(vma->iomap);
453 * TODO: consider just using i915_gem_object_pin_map() for lmem
454 * instead, which already supports mapping non-contiguous chunks
455 * of pages, that way we can also drop the
456 * I915_BO_ALLOC_CONTIGUOUS when allocating the object.
458 if (i915_gem_object_is_lmem(vma->obj))
459 ptr = i915_gem_object_lmem_io_map(vma->obj, 0,
460 vma->obj->base.size);
462 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
470 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
471 io_mapping_unmap(ptr);
478 err = i915_vma_pin_fence(vma);
482 i915_vma_set_ggtt_write(vma);
484 /* NB Access through the GTT requires the device to be awake. */
488 __i915_vma_unpin(vma);
490 return IO_ERR_PTR(err);
493 void i915_vma_flush_writes(struct i915_vma *vma)
495 if (i915_vma_unset_ggtt_write(vma))
496 intel_gt_flush_ggtt_writes(vma->vm->gt);
499 void i915_vma_unpin_iomap(struct i915_vma *vma)
501 GEM_BUG_ON(vma->iomap == NULL);
503 i915_vma_flush_writes(vma);
505 i915_vma_unpin_fence(vma);
509 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
511 struct i915_vma *vma;
512 struct drm_i915_gem_object *obj;
514 vma = fetch_and_zero(p_vma);
523 if (flags & I915_VMA_RELEASE_MAP)
524 i915_gem_object_unpin_map(obj);
526 i915_gem_object_put(obj);
529 bool i915_vma_misplaced(const struct i915_vma *vma,
530 u64 size, u64 alignment, u64 flags)
532 if (!drm_mm_node_allocated(&vma->node))
535 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
538 if (vma->node.size < size)
541 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
542 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
545 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
548 if (flags & PIN_OFFSET_BIAS &&
549 vma->node.start < (flags & PIN_OFFSET_MASK))
552 if (flags & PIN_OFFSET_FIXED &&
553 vma->node.start != (flags & PIN_OFFSET_MASK))
559 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
561 bool mappable, fenceable;
563 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
564 GEM_BUG_ON(!vma->fence_size);
566 fenceable = (vma->node.size >= vma->fence_size &&
567 IS_ALIGNED(vma->node.start, vma->fence_alignment));
569 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
571 if (mappable && fenceable)
572 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
574 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
577 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
579 struct drm_mm_node *node = &vma->node;
580 struct drm_mm_node *other;
583 * On some machines we have to be careful when putting differing types
584 * of snoopable memory together to avoid the prefetcher crossing memory
585 * domains and dying. During vm initialisation, we decide whether or not
586 * these constraints apply and set the drm_mm.color_adjust
589 if (!i915_vm_has_cache_coloring(vma->vm))
592 /* Only valid to be called on an already inserted vma */
593 GEM_BUG_ON(!drm_mm_node_allocated(node));
594 GEM_BUG_ON(list_empty(&node->node_list));
596 other = list_prev_entry(node, node_list);
597 if (i915_node_color_differs(other, color) &&
598 !drm_mm_hole_follows(other))
601 other = list_next_entry(node, node_list);
602 if (i915_node_color_differs(other, color) &&
603 !drm_mm_hole_follows(node))
610 * i915_vma_insert - finds a slot for the vma in its address space
612 * @size: requested size in bytes (can be larger than the VMA)
613 * @alignment: required alignment
614 * @flags: mask of PIN_* flags to use
616 * First we try to allocate some free space that meets the requirements for
617 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
618 * preferrably the oldest idle entry to make room for the new VMA.
621 * 0 on success, negative error code otherwise.
624 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
630 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
631 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
633 size = max(size, vma->size);
634 alignment = max(alignment, vma->display_alignment);
635 if (flags & PIN_MAPPABLE) {
636 size = max_t(typeof(size), size, vma->fence_size);
637 alignment = max_t(typeof(alignment),
638 alignment, vma->fence_alignment);
641 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
642 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
643 GEM_BUG_ON(!is_power_of_2(alignment));
645 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
646 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
648 end = vma->vm->total;
649 if (flags & PIN_MAPPABLE)
650 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
651 if (flags & PIN_ZONE_4G)
652 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
653 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
655 /* If binding the object/GGTT view requires more space than the entire
656 * aperture has, reject it early before evicting everything in a vain
657 * attempt to find space.
660 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
661 size, flags & PIN_MAPPABLE ? "mappable" : "total",
667 if (i915_vm_has_cache_coloring(vma->vm))
668 color = vma->obj->cache_level;
670 if (flags & PIN_OFFSET_FIXED) {
671 u64 offset = flags & PIN_OFFSET_MASK;
672 if (!IS_ALIGNED(offset, alignment) ||
673 range_overflows(offset, size, end))
676 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
683 * We only support huge gtt pages through the 48b PPGTT,
684 * however we also don't want to force any alignment for
685 * objects which need to be tightly packed into the low 32bits.
687 * Note that we assume that GGTT are limited to 4GiB for the
688 * forseeable future. See also i915_ggtt_offset().
690 if (upper_32_bits(end - 1) &&
691 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
693 * We can't mix 64K and 4K PTEs in the same page-table
694 * (2M block), and so to avoid the ugliness and
695 * complexity of coloring we opt for just aligning 64K
699 rounddown_pow_of_two(vma->page_sizes.sg |
700 I915_GTT_PAGE_SIZE_2M);
703 * Check we don't expand for the limited Global GTT
704 * (mappable aperture is even more precious!). This
705 * also checks that we exclude the aliasing-ppgtt.
707 GEM_BUG_ON(i915_vma_is_ggtt(vma));
709 alignment = max(alignment, page_alignment);
711 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
712 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
715 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
716 size, alignment, color,
721 GEM_BUG_ON(vma->node.start < start);
722 GEM_BUG_ON(vma->node.start + vma->node.size > end);
724 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
725 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
727 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
733 i915_vma_detach(struct i915_vma *vma)
735 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
736 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
739 * And finally now the object is completely decoupled from this
740 * vma, we can drop its hold on the backing storage and allow
741 * it to be reaped by the shrinker.
743 list_del(&vma->vm_link);
746 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
751 bound = atomic_read(&vma->flags);
753 if (unlikely(flags & ~bound))
756 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
759 if (!(bound & I915_VMA_PIN_MASK))
762 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
763 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
769 * If pin_count==0, but we are bound, check under the lock to avoid
770 * racing with a concurrent i915_vma_unbind().
772 mutex_lock(&vma->vm->mutex);
774 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
779 if (unlikely(flags & ~bound)) {
783 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
784 mutex_unlock(&vma->vm->mutex);
789 static int vma_get_pages(struct i915_vma *vma)
792 bool pinned_pages = true;
794 if (atomic_add_unless(&vma->pages_count, 1, 0))
797 err = i915_gem_object_pin_pages(vma->obj);
801 /* Allocations ahoy! */
802 if (mutex_lock_interruptible(&vma->pages_mutex)) {
807 if (!atomic_read(&vma->pages_count)) {
808 err = vma->ops->set_pages(vma);
811 pinned_pages = false;
813 atomic_inc(&vma->pages_count);
816 mutex_unlock(&vma->pages_mutex);
819 __i915_gem_object_unpin_pages(vma->obj);
824 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
826 /* We allocate under vma_get_pages, so beware the shrinker */
827 mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
828 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
829 if (atomic_sub_return(count, &vma->pages_count) == 0) {
830 vma->ops->clear_pages(vma);
831 GEM_BUG_ON(vma->pages);
833 i915_gem_object_unpin_pages(vma->obj);
835 mutex_unlock(&vma->pages_mutex);
838 static void vma_put_pages(struct i915_vma *vma)
840 if (atomic_add_unless(&vma->pages_count, -1, 1))
843 __vma_put_pages(vma, 1);
846 static void vma_unbind_pages(struct i915_vma *vma)
850 lockdep_assert_held(&vma->vm->mutex);
852 /* The upper portion of pages_count is the number of bindings */
853 count = atomic_read(&vma->pages_count);
854 count >>= I915_VMA_PAGES_BIAS;
857 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
860 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
861 u64 size, u64 alignment, u64 flags)
863 struct i915_vma_work *work = NULL;
864 intel_wakeref_t wakeref = 0;
868 #ifdef CONFIG_PROVE_LOCKING
869 if (debug_locks && !WARN_ON(!ww))
870 assert_vma_held(vma);
873 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
874 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
876 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
878 /* First try and grab the pin without rebinding the vma */
879 if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
882 err = vma_get_pages(vma);
886 if (flags & PIN_GLOBAL)
887 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
889 if (flags & vma->vm->bind_async_flags) {
891 err = i915_vm_lock_objects(vma->vm, ww);
895 work = i915_vma_work();
901 work->vm = i915_vm_get(vma->vm);
903 /* Allocate enough page directories to used PTE */
904 if (vma->vm->allocate_va_range) {
905 err = i915_vm_alloc_pt_stash(vma->vm,
911 err = i915_vm_map_pt_stash(vma->vm, &work->stash);
918 * Differentiate between user/kernel vma inside the aliasing-ppgtt.
920 * We conflate the Global GTT with the user's vma when using the
921 * aliasing-ppgtt, but it is still vitally important to try and
922 * keep the use cases distinct. For example, userptr objects are
923 * not allowed inside the Global GTT as that will cause lock
924 * inversions when we have to evict them the mmu_notifier callbacks -
925 * but they are allowed to be part of the user ppGTT which can never
926 * be mapped. As such we try to give the distinct users of the same
927 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
928 * and i915_ppgtt separate].
930 * NB this may cause us to mask real lock inversions -- while the
931 * code is safe today, lockdep may not be able to spot future
934 err = mutex_lock_interruptible_nested(&vma->vm->mutex,
935 !(flags & PIN_GLOBAL));
939 /* No more allocations allowed now we hold vm->mutex */
941 if (unlikely(i915_vma_is_closed(vma))) {
946 bound = atomic_read(&vma->flags);
947 if (unlikely(bound & I915_VMA_ERROR)) {
952 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
953 err = -EAGAIN; /* pins are meant to be fairly temporary */
957 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
962 err = i915_active_acquire(&vma->active);
966 if (!(bound & I915_VMA_BIND_MASK)) {
967 err = i915_vma_insert(vma, size, alignment, flags);
971 if (i915_is_ggtt(vma->vm))
972 __i915_vma_set_map_and_fenceable(vma);
975 GEM_BUG_ON(!vma->pages);
976 err = i915_vma_bind(vma,
977 vma->obj->cache_level,
982 /* There should only be at most 2 active bindings (user, global) */
983 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
984 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
985 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
988 GEM_BUG_ON(!i915_vma_is_pinned(vma));
989 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
990 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
993 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
994 i915_vma_detach(vma);
995 drm_mm_remove_node(&vma->node);
998 i915_active_release(&vma->active);
1000 mutex_unlock(&vma->vm->mutex);
1003 dma_fence_work_commit_imm(&work->base);
1006 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
1011 static void flush_idle_contexts(struct intel_gt *gt)
1013 struct intel_engine_cs *engine;
1014 enum intel_engine_id id;
1016 for_each_engine(engine, gt, id)
1017 intel_engine_flush_barriers(engine);
1019 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1022 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1023 u32 align, unsigned int flags)
1025 struct i915_address_space *vm = vma->vm;
1028 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1030 #ifdef CONFIG_LOCKDEP
1031 WARN_ON(!ww && dma_resv_held(vma->obj->base.resv));
1036 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);
1038 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1039 if (err != -ENOSPC) {
1041 err = i915_vma_wait_for_bind(vma);
1043 i915_vma_unpin(vma);
1048 /* Unlike i915_vma_pin, we don't take no for an answer! */
1049 flush_idle_contexts(vm->gt);
1050 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1051 i915_gem_evict_vm(vm);
1052 mutex_unlock(&vm->mutex);
1057 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1060 * We defer actually closing, unbinding and destroying the VMA until
1061 * the next idle point, or if the object is freed in the meantime. By
1062 * postponing the unbind, we allow for it to be resurrected by the
1063 * client, avoiding the work required to rebind the VMA. This is
1064 * advantageous for DRI, where the client/server pass objects
1065 * between themselves, temporarily opening a local VMA to the
1066 * object, and then closing it again. The same object is then reused
1067 * on the next frame (or two, depending on the depth of the swap queue)
1068 * causing us to rebind the VMA once more. This ends up being a lot
1069 * of wasted work for the steady state.
1071 GEM_BUG_ON(i915_vma_is_closed(vma));
1072 list_add(&vma->closed_link, >->closed_vma);
1075 void i915_vma_close(struct i915_vma *vma)
1077 struct intel_gt *gt = vma->vm->gt;
1078 unsigned long flags;
1080 if (i915_vma_is_ggtt(vma))
1083 GEM_BUG_ON(!atomic_read(&vma->open_count));
1084 if (atomic_dec_and_lock_irqsave(&vma->open_count,
1087 __vma_close(vma, gt);
1088 spin_unlock_irqrestore(>->closed_lock, flags);
1092 static void __i915_vma_remove_closed(struct i915_vma *vma)
1094 struct intel_gt *gt = vma->vm->gt;
1096 spin_lock_irq(>->closed_lock);
1097 list_del_init(&vma->closed_link);
1098 spin_unlock_irq(>->closed_lock);
1101 void i915_vma_reopen(struct i915_vma *vma)
1103 if (i915_vma_is_closed(vma))
1104 __i915_vma_remove_closed(vma);
1107 void i915_vma_release(struct kref *ref)
1109 struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1110 struct drm_i915_gem_object *obj = vma->obj;
1112 if (drm_mm_node_allocated(&vma->node)) {
1113 mutex_lock(&vma->vm->mutex);
1114 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1115 WARN_ON(__i915_vma_unbind(vma));
1116 mutex_unlock(&vma->vm->mutex);
1117 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1119 GEM_BUG_ON(i915_vma_is_active(vma));
1121 spin_lock(&obj->vma.lock);
1122 list_del(&vma->obj_link);
1123 if (!RB_EMPTY_NODE(&vma->obj_node))
1124 rb_erase(&vma->obj_node, &obj->vma.tree);
1125 spin_unlock(&obj->vma.lock);
1127 __i915_vma_remove_closed(vma);
1128 i915_vm_put(vma->vm);
1130 i915_active_fini(&vma->active);
1134 void i915_vma_parked(struct intel_gt *gt)
1136 struct i915_vma *vma, *next;
1139 spin_lock_irq(>->closed_lock);
1140 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1141 struct drm_i915_gem_object *obj = vma->obj;
1142 struct i915_address_space *vm = vma->vm;
1144 /* XXX All to avoid keeping a reference on i915_vma itself */
1146 if (!kref_get_unless_zero(&obj->base.refcount))
1149 if (!i915_vm_tryopen(vm)) {
1150 i915_gem_object_put(obj);
1154 list_move(&vma->closed_link, &closed);
1156 spin_unlock_irq(>->closed_lock);
1158 /* As the GT is held idle, no vma can be reopened as we destroy them */
1159 list_for_each_entry_safe(vma, next, &closed, closed_link) {
1160 struct drm_i915_gem_object *obj = vma->obj;
1161 struct i915_address_space *vm = vma->vm;
1163 INIT_LIST_HEAD(&vma->closed_link);
1164 __i915_vma_put(vma);
1166 i915_gem_object_put(obj);
1171 static void __i915_vma_iounmap(struct i915_vma *vma)
1173 GEM_BUG_ON(i915_vma_is_pinned(vma));
1175 if (vma->iomap == NULL)
1178 io_mapping_unmap(vma->iomap);
1182 void i915_vma_revoke_mmap(struct i915_vma *vma)
1184 struct drm_vma_offset_node *node;
1187 if (!i915_vma_has_userfault(vma))
1190 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1191 GEM_BUG_ON(!vma->obj->userfault_count);
1193 node = &vma->mmo->vma_node;
1194 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1195 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1196 drm_vma_node_offset_addr(node) + vma_offset,
1200 i915_vma_unset_userfault(vma);
1201 if (!--vma->obj->userfault_count)
1202 list_del(&vma->obj->userfault_link);
1206 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1208 return __i915_request_await_exclusive(rq, &vma->active);
1211 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1215 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1217 /* Wait for the vma to be bound before we start! */
1218 err = __i915_request_await_bind(rq, vma);
1222 return i915_active_add_request(&vma->active, rq);
1225 int _i915_vma_move_to_active(struct i915_vma *vma,
1226 struct i915_request *rq,
1227 struct dma_fence *fence,
1230 struct drm_i915_gem_object *obj = vma->obj;
1233 assert_object_held(obj);
1235 err = __i915_vma_move_to_active(vma, rq);
1239 if (flags & EXEC_OBJECT_WRITE) {
1240 struct intel_frontbuffer *front;
1242 front = __intel_frontbuffer_get(obj);
1243 if (unlikely(front)) {
1244 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1245 i915_active_add_request(&front->write, rq);
1246 intel_frontbuffer_put(front);
1250 dma_resv_add_excl_fence(vma->obj->base.resv, fence);
1251 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1252 obj->read_domains = 0;
1255 if (!(flags & __EXEC_OBJECT_NO_RESERVE)) {
1256 err = dma_resv_reserve_shared(vma->obj->base.resv, 1);
1262 dma_resv_add_shared_fence(vma->obj->base.resv, fence);
1263 obj->write_domain = 0;
1267 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1268 i915_active_add_request(&vma->fence->active, rq);
1270 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1271 obj->mm.dirty = true;
1273 GEM_BUG_ON(!i915_vma_is_active(vma));
1277 void __i915_vma_evict(struct i915_vma *vma)
1279 GEM_BUG_ON(i915_vma_is_pinned(vma));
1281 if (i915_vma_is_map_and_fenceable(vma)) {
1282 /* Force a pagefault for domain tracking on next user access */
1283 i915_vma_revoke_mmap(vma);
1286 * Check that we have flushed all writes through the GGTT
1287 * before the unbind, other due to non-strict nature of those
1288 * indirect writes they may end up referencing the GGTT PTE
1291 * Note that we may be concurrently poking at the GGTT_WRITE
1292 * bit from set-domain, as we mark all GGTT vma associated
1293 * with an object. We know this is for another vma, as we
1294 * are currently unbinding this one -- so if this vma will be
1295 * reused, it will be refaulted and have its dirty bit set
1296 * before the next write.
1298 i915_vma_flush_writes(vma);
1300 /* release the fence reg _after_ flushing */
1301 i915_vma_revoke_fence(vma);
1303 __i915_vma_iounmap(vma);
1304 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1306 GEM_BUG_ON(vma->fence);
1307 GEM_BUG_ON(i915_vma_has_userfault(vma));
1309 if (likely(atomic_read(&vma->vm->open))) {
1310 trace_i915_vma_unbind(vma);
1311 vma->ops->unbind_vma(vma->vm, vma);
1313 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1316 i915_vma_detach(vma);
1317 vma_unbind_pages(vma);
1320 int __i915_vma_unbind(struct i915_vma *vma)
1324 lockdep_assert_held(&vma->vm->mutex);
1326 if (!drm_mm_node_allocated(&vma->node))
1329 if (i915_vma_is_pinned(vma)) {
1330 vma_print_allocator(vma, "is pinned");
1335 * After confirming that no one else is pinning this vma, wait for
1336 * any laggards who may have crept in during the wait (through
1337 * a residual pin skipping the vm->mutex) to complete.
1339 ret = i915_vma_sync(vma);
1343 GEM_BUG_ON(i915_vma_is_active(vma));
1344 __i915_vma_evict(vma);
1346 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1350 int i915_vma_unbind(struct i915_vma *vma)
1352 struct i915_address_space *vm = vma->vm;
1353 intel_wakeref_t wakeref = 0;
1356 /* Optimistic wait before taking the mutex */
1357 err = i915_vma_sync(vma);
1361 if (!drm_mm_node_allocated(&vma->node))
1364 if (i915_vma_is_pinned(vma)) {
1365 vma_print_allocator(vma, "is pinned");
1369 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1370 /* XXX not always required: nop_clear_range */
1371 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1373 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1377 err = __i915_vma_unbind(vma);
1378 mutex_unlock(&vm->mutex);
1382 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1386 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1388 i915_gem_object_make_unshrinkable(vma->obj);
1392 void i915_vma_make_shrinkable(struct i915_vma *vma)
1394 i915_gem_object_make_shrinkable(vma->obj);
1397 void i915_vma_make_purgeable(struct i915_vma *vma)
1399 i915_gem_object_make_purgeable(vma->obj);
1402 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1403 #include "selftests/i915_vma.c"
1406 void i915_vma_module_exit(void)
1408 kmem_cache_destroy(slab_vmas);
1411 int __init i915_vma_module_init(void)
1413 slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);