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 spin_lock(&obj->vma.lock);
163 if (i915_is_ggtt(vm)) {
164 if (unlikely(overflows_type(vma->size, u32)))
167 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
168 i915_gem_object_get_tiling(obj),
169 i915_gem_object_get_stride(obj));
170 if (unlikely(vma->fence_size < vma->size || /* overflow */
171 vma->fence_size > vm->total))
174 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
176 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
177 i915_gem_object_get_tiling(obj),
178 i915_gem_object_get_stride(obj));
179 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
181 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
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);
229 spin_unlock(&obj->vma.lock);
232 return ERR_PTR(-E2BIG);
235 static struct i915_vma *
236 vma_lookup(struct drm_i915_gem_object *obj,
237 struct i915_address_space *vm,
238 const struct i915_ggtt_view *view)
242 rb = obj->vma.tree.rb_node;
244 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
247 cmp = i915_vma_compare(vma, vm, view);
261 * i915_vma_instance - return the singleton instance of the VMA
262 * @obj: parent &struct drm_i915_gem_object to be mapped
263 * @vm: address space in which the mapping is located
264 * @view: additional mapping requirements
266 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
267 * the same @view characteristics. If a match is not found, one is created.
268 * Once created, the VMA is kept until either the object is freed, or the
269 * address space is closed.
271 * Returns the vma, or an error pointer.
274 i915_vma_instance(struct drm_i915_gem_object *obj,
275 struct i915_address_space *vm,
276 const struct i915_ggtt_view *view)
278 struct i915_vma *vma;
280 GEM_BUG_ON(view && !i915_is_ggtt(vm));
281 GEM_BUG_ON(!atomic_read(&vm->open));
283 spin_lock(&obj->vma.lock);
284 vma = vma_lookup(obj, vm, view);
285 spin_unlock(&obj->vma.lock);
287 /* vma_create() will resolve the race if another creates the vma */
289 vma = vma_create(obj, vm, view);
291 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
295 struct i915_vma_work {
296 struct dma_fence_work base;
297 struct i915_vma *vma;
298 struct drm_i915_gem_object *pinned;
299 struct i915_sw_dma_fence_cb cb;
300 enum i915_cache_level cache_level;
304 static int __vma_bind(struct dma_fence_work *work)
306 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
307 struct i915_vma *vma = vw->vma;
310 err = vma->ops->bind_vma(vma, vw->cache_level, vw->flags);
312 atomic_or(I915_VMA_ERROR, &vma->flags);
317 static void __vma_release(struct dma_fence_work *work)
319 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
322 __i915_gem_object_unpin_pages(vw->pinned);
325 static const struct dma_fence_work_ops bind_ops = {
328 .release = __vma_release,
331 struct i915_vma_work *i915_vma_work(void)
333 struct i915_vma_work *vw;
335 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
339 dma_fence_work_init(&vw->base, &bind_ops);
340 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
345 int i915_vma_wait_for_bind(struct i915_vma *vma)
349 if (rcu_access_pointer(vma->active.excl.fence)) {
350 struct dma_fence *fence;
353 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
356 err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
357 dma_fence_put(fence);
365 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
367 * @cache_level: mapping cache level
368 * @flags: flags like global or local mapping
369 * @work: preallocated worker for allocating and binding the PTE
371 * DMA addresses are taken from the scatter-gather table of this object (or of
372 * this VMA in case of non-default GGTT views) and PTE entries set up.
373 * Note that DMA addresses are also the only part of the SG table we care about.
375 int i915_vma_bind(struct i915_vma *vma,
376 enum i915_cache_level cache_level,
378 struct i915_vma_work *work)
384 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
385 GEM_BUG_ON(vma->size > vma->node.size);
387 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
392 if (GEM_DEBUG_WARN_ON(!flags))
396 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
398 vma_flags = atomic_read(&vma->flags);
399 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
400 if (flags & PIN_UPDATE)
401 bind_flags |= vma_flags;
403 bind_flags &= ~vma_flags;
407 GEM_BUG_ON(!vma->pages);
409 trace_i915_vma_bind(vma, bind_flags);
410 if (work && (bind_flags & ~vma_flags) & vma->vm->bind_async_flags) {
411 struct dma_fence *prev;
414 work->cache_level = cache_level;
415 work->flags = bind_flags | I915_VMA_ALLOC;
418 * Note we only want to chain up to the migration fence on
419 * the pages (not the object itself). As we don't track that,
420 * yet, we have to use the exclusive fence instead.
422 * Also note that we do not want to track the async vma as
423 * part of the obj->resv->excl_fence as it only affects
424 * execution and not content or object's backing store lifetime.
426 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
428 __i915_sw_fence_await_dma_fence(&work->base.chain,
434 work->base.dma.error = 0; /* enable the queue_work() */
437 __i915_gem_object_pin_pages(vma->obj);
438 work->pinned = vma->obj;
441 ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
446 atomic_or(bind_flags, &vma->flags);
450 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
455 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
460 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
461 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
463 ptr = READ_ONCE(vma->iomap);
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);
527 if (flags & I915_VMA_RELEASE_MAP)
528 i915_gem_object_unpin_map(obj);
530 i915_gem_object_put(obj);
533 bool i915_vma_misplaced(const struct i915_vma *vma,
534 u64 size, u64 alignment, u64 flags)
536 if (!drm_mm_node_allocated(&vma->node))
539 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
542 if (vma->node.size < size)
545 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
546 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
549 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
552 if (flags & PIN_OFFSET_BIAS &&
553 vma->node.start < (flags & PIN_OFFSET_MASK))
556 if (flags & PIN_OFFSET_FIXED &&
557 vma->node.start != (flags & PIN_OFFSET_MASK))
563 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
565 bool mappable, fenceable;
567 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
568 GEM_BUG_ON(!vma->fence_size);
570 fenceable = (vma->node.size >= vma->fence_size &&
571 IS_ALIGNED(vma->node.start, vma->fence_alignment));
573 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
575 if (mappable && fenceable)
576 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
578 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
581 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
583 struct drm_mm_node *node = &vma->node;
584 struct drm_mm_node *other;
587 * On some machines we have to be careful when putting differing types
588 * of snoopable memory together to avoid the prefetcher crossing memory
589 * domains and dying. During vm initialisation, we decide whether or not
590 * these constraints apply and set the drm_mm.color_adjust
593 if (!i915_vm_has_cache_coloring(vma->vm))
596 /* Only valid to be called on an already inserted vma */
597 GEM_BUG_ON(!drm_mm_node_allocated(node));
598 GEM_BUG_ON(list_empty(&node->node_list));
600 other = list_prev_entry(node, node_list);
601 if (i915_node_color_differs(other, color) &&
602 !drm_mm_hole_follows(other))
605 other = list_next_entry(node, node_list);
606 if (i915_node_color_differs(other, color) &&
607 !drm_mm_hole_follows(node))
614 * i915_vma_insert - finds a slot for the vma in its address space
616 * @size: requested size in bytes (can be larger than the VMA)
617 * @alignment: required alignment
618 * @flags: mask of PIN_* flags to use
620 * First we try to allocate some free space that meets the requirements for
621 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
622 * preferrably the oldest idle entry to make room for the new VMA.
625 * 0 on success, negative error code otherwise.
628 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
634 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
635 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
637 size = max(size, vma->size);
638 alignment = max(alignment, vma->display_alignment);
639 if (flags & PIN_MAPPABLE) {
640 size = max_t(typeof(size), size, vma->fence_size);
641 alignment = max_t(typeof(alignment),
642 alignment, vma->fence_alignment);
645 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
646 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
647 GEM_BUG_ON(!is_power_of_2(alignment));
649 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
650 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
652 end = vma->vm->total;
653 if (flags & PIN_MAPPABLE)
654 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
655 if (flags & PIN_ZONE_4G)
656 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
657 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
659 /* If binding the object/GGTT view requires more space than the entire
660 * aperture has, reject it early before evicting everything in a vain
661 * attempt to find space.
664 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
665 size, flags & PIN_MAPPABLE ? "mappable" : "total",
671 if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
672 color = vma->obj->cache_level;
674 if (flags & PIN_OFFSET_FIXED) {
675 u64 offset = flags & PIN_OFFSET_MASK;
676 if (!IS_ALIGNED(offset, alignment) ||
677 range_overflows(offset, size, end))
680 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
687 * We only support huge gtt pages through the 48b PPGTT,
688 * however we also don't want to force any alignment for
689 * objects which need to be tightly packed into the low 32bits.
691 * Note that we assume that GGTT are limited to 4GiB for the
692 * forseeable future. See also i915_ggtt_offset().
694 if (upper_32_bits(end - 1) &&
695 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
697 * We can't mix 64K and 4K PTEs in the same page-table
698 * (2M block), and so to avoid the ugliness and
699 * complexity of coloring we opt for just aligning 64K
703 rounddown_pow_of_two(vma->page_sizes.sg |
704 I915_GTT_PAGE_SIZE_2M);
707 * Check we don't expand for the limited Global GTT
708 * (mappable aperture is even more precious!). This
709 * also checks that we exclude the aliasing-ppgtt.
711 GEM_BUG_ON(i915_vma_is_ggtt(vma));
713 alignment = max(alignment, page_alignment);
715 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
716 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
719 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
720 size, alignment, color,
725 GEM_BUG_ON(vma->node.start < start);
726 GEM_BUG_ON(vma->node.start + vma->node.size > end);
728 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
729 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
731 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
737 i915_vma_detach(struct i915_vma *vma)
739 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
740 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
743 * And finally now the object is completely decoupled from this
744 * vma, we can drop its hold on the backing storage and allow
745 * it to be reaped by the shrinker.
747 list_del(&vma->vm_link);
750 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
755 bound = atomic_read(&vma->flags);
757 if (unlikely(flags & ~bound))
760 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
763 if (!(bound & I915_VMA_PIN_MASK))
766 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
767 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
773 * If pin_count==0, but we are bound, check under the lock to avoid
774 * racing with a concurrent i915_vma_unbind().
776 mutex_lock(&vma->vm->mutex);
778 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
783 if (unlikely(flags & ~bound)) {
787 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
788 mutex_unlock(&vma->vm->mutex);
793 static int vma_get_pages(struct i915_vma *vma)
797 if (atomic_add_unless(&vma->pages_count, 1, 0))
800 /* Allocations ahoy! */
801 if (mutex_lock_interruptible(&vma->pages_mutex))
804 if (!atomic_read(&vma->pages_count)) {
806 err = i915_gem_object_pin_pages(vma->obj);
811 err = vma->ops->set_pages(vma);
814 i915_gem_object_unpin_pages(vma->obj);
818 atomic_inc(&vma->pages_count);
821 mutex_unlock(&vma->pages_mutex);
826 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
828 /* We allocate under vma_get_pages, so beware the shrinker */
829 mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
830 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
831 if (atomic_sub_return(count, &vma->pages_count) == 0) {
832 vma->ops->clear_pages(vma);
833 GEM_BUG_ON(vma->pages);
835 i915_gem_object_unpin_pages(vma->obj);
837 mutex_unlock(&vma->pages_mutex);
840 static void vma_put_pages(struct i915_vma *vma)
842 if (atomic_add_unless(&vma->pages_count, -1, 1))
845 __vma_put_pages(vma, 1);
848 static void vma_unbind_pages(struct i915_vma *vma)
852 lockdep_assert_held(&vma->vm->mutex);
854 /* The upper portion of pages_count is the number of bindings */
855 count = atomic_read(&vma->pages_count);
856 count >>= I915_VMA_PAGES_BIAS;
859 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
862 int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
864 struct i915_vma_work *work = NULL;
865 intel_wakeref_t wakeref = 0;
869 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
870 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
872 GEM_BUG_ON(flags & PIN_UPDATE);
873 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
875 /* First try and grab the pin without rebinding the vma */
876 if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
879 err = vma_get_pages(vma);
883 if (flags & vma->vm->bind_async_flags) {
884 work = i915_vma_work();
891 if (flags & PIN_GLOBAL)
892 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
895 * Differentiate between user/kernel vma inside the aliasing-ppgtt.
897 * We conflate the Global GTT with the user's vma when using the
898 * aliasing-ppgtt, but it is still vitally important to try and
899 * keep the use cases distinct. For example, userptr objects are
900 * not allowed inside the Global GTT as that will cause lock
901 * inversions when we have to evict them the mmu_notifier callbacks -
902 * but they are allowed to be part of the user ppGTT which can never
903 * be mapped. As such we try to give the distinct users of the same
904 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
905 * and i915_ppgtt separate].
907 * NB this may cause us to mask real lock inversions -- while the
908 * code is safe today, lockdep may not be able to spot future
911 err = mutex_lock_interruptible_nested(&vma->vm->mutex,
912 !(flags & PIN_GLOBAL));
916 /* No more allocations allowed now we hold vm->mutex */
918 if (unlikely(i915_vma_is_closed(vma))) {
923 bound = atomic_read(&vma->flags);
924 if (unlikely(bound & I915_VMA_ERROR)) {
929 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
930 err = -EAGAIN; /* pins are meant to be fairly temporary */
934 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
939 err = i915_active_acquire(&vma->active);
943 if (!(bound & I915_VMA_BIND_MASK)) {
944 err = i915_vma_insert(vma, size, alignment, flags);
948 if (i915_is_ggtt(vma->vm))
949 __i915_vma_set_map_and_fenceable(vma);
952 GEM_BUG_ON(!vma->pages);
953 err = i915_vma_bind(vma,
954 vma->obj ? vma->obj->cache_level : 0,
959 /* There should only be at most 2 active bindings (user, global) */
960 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
961 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
962 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
965 GEM_BUG_ON(!i915_vma_is_pinned(vma));
966 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
967 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
970 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
971 i915_vma_detach(vma);
972 drm_mm_remove_node(&vma->node);
975 i915_active_release(&vma->active);
977 mutex_unlock(&vma->vm->mutex);
980 dma_fence_work_commit_imm(&work->base);
982 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
988 static void flush_idle_contexts(struct intel_gt *gt)
990 struct intel_engine_cs *engine;
991 enum intel_engine_id id;
993 for_each_engine(engine, gt, id)
994 intel_engine_flush_barriers(engine);
996 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
999 int i915_ggtt_pin(struct i915_vma *vma, u32 align, unsigned int flags)
1001 struct i915_address_space *vm = vma->vm;
1004 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1007 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1008 if (err != -ENOSPC) {
1010 err = i915_vma_wait_for_bind(vma);
1012 i915_vma_unpin(vma);
1017 /* Unlike i915_vma_pin, we don't take no for an answer! */
1018 flush_idle_contexts(vm->gt);
1019 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1020 i915_gem_evict_vm(vm);
1021 mutex_unlock(&vm->mutex);
1026 void i915_vma_close(struct i915_vma *vma)
1028 struct intel_gt *gt = vma->vm->gt;
1029 unsigned long flags;
1031 GEM_BUG_ON(i915_vma_is_closed(vma));
1034 * We defer actually closing, unbinding and destroying the VMA until
1035 * the next idle point, or if the object is freed in the meantime. By
1036 * postponing the unbind, we allow for it to be resurrected by the
1037 * client, avoiding the work required to rebind the VMA. This is
1038 * advantageous for DRI, where the client/server pass objects
1039 * between themselves, temporarily opening a local VMA to the
1040 * object, and then closing it again. The same object is then reused
1041 * on the next frame (or two, depending on the depth of the swap queue)
1042 * causing us to rebind the VMA once more. This ends up being a lot
1043 * of wasted work for the steady state.
1045 spin_lock_irqsave(>->closed_lock, flags);
1046 list_add(&vma->closed_link, >->closed_vma);
1047 spin_unlock_irqrestore(>->closed_lock, flags);
1050 static void __i915_vma_remove_closed(struct i915_vma *vma)
1052 struct intel_gt *gt = vma->vm->gt;
1054 spin_lock_irq(>->closed_lock);
1055 list_del_init(&vma->closed_link);
1056 spin_unlock_irq(>->closed_lock);
1059 void i915_vma_reopen(struct i915_vma *vma)
1061 if (i915_vma_is_closed(vma))
1062 __i915_vma_remove_closed(vma);
1065 void i915_vma_release(struct kref *ref)
1067 struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1069 if (drm_mm_node_allocated(&vma->node)) {
1070 mutex_lock(&vma->vm->mutex);
1071 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1072 WARN_ON(__i915_vma_unbind(vma));
1073 mutex_unlock(&vma->vm->mutex);
1074 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1076 GEM_BUG_ON(i915_vma_is_active(vma));
1079 struct drm_i915_gem_object *obj = vma->obj;
1081 spin_lock(&obj->vma.lock);
1082 list_del(&vma->obj_link);
1083 rb_erase(&vma->obj_node, &obj->vma.tree);
1084 spin_unlock(&obj->vma.lock);
1087 __i915_vma_remove_closed(vma);
1088 i915_vm_put(vma->vm);
1090 i915_active_fini(&vma->active);
1094 void i915_vma_parked(struct intel_gt *gt)
1096 struct i915_vma *vma, *next;
1099 spin_lock_irq(>->closed_lock);
1100 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1101 struct drm_i915_gem_object *obj = vma->obj;
1102 struct i915_address_space *vm = vma->vm;
1104 /* XXX All to avoid keeping a reference on i915_vma itself */
1106 if (!kref_get_unless_zero(&obj->base.refcount))
1109 if (!i915_vm_tryopen(vm)) {
1110 i915_gem_object_put(obj);
1114 list_move(&vma->closed_link, &closed);
1116 spin_unlock_irq(>->closed_lock);
1118 /* As the GT is held idle, no vma can be reopened as we destroy them */
1119 list_for_each_entry_safe(vma, next, &closed, closed_link) {
1120 struct drm_i915_gem_object *obj = vma->obj;
1121 struct i915_address_space *vm = vma->vm;
1123 INIT_LIST_HEAD(&vma->closed_link);
1124 __i915_vma_put(vma);
1126 i915_gem_object_put(obj);
1131 static void __i915_vma_iounmap(struct i915_vma *vma)
1133 GEM_BUG_ON(i915_vma_is_pinned(vma));
1135 if (vma->iomap == NULL)
1138 io_mapping_unmap(vma->iomap);
1142 void i915_vma_revoke_mmap(struct i915_vma *vma)
1144 struct drm_vma_offset_node *node;
1147 if (!i915_vma_has_userfault(vma))
1150 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1151 GEM_BUG_ON(!vma->obj->userfault_count);
1153 node = &vma->mmo->vma_node;
1154 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1155 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1156 drm_vma_node_offset_addr(node) + vma_offset,
1160 i915_vma_unset_userfault(vma);
1161 if (!--vma->obj->userfault_count)
1162 list_del(&vma->obj->userfault_link);
1165 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1169 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1171 /* Wait for the vma to be bound before we start! */
1172 err = i915_request_await_active(rq, &vma->active,
1173 I915_ACTIVE_AWAIT_EXCL);
1177 return i915_active_add_request(&vma->active, rq);
1180 int i915_vma_move_to_active(struct i915_vma *vma,
1181 struct i915_request *rq,
1184 struct drm_i915_gem_object *obj = vma->obj;
1187 assert_object_held(obj);
1189 err = __i915_vma_move_to_active(vma, rq);
1193 if (flags & EXEC_OBJECT_WRITE) {
1194 struct intel_frontbuffer *front;
1196 front = __intel_frontbuffer_get(obj);
1197 if (unlikely(front)) {
1198 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1199 i915_active_add_request(&front->write, rq);
1200 intel_frontbuffer_put(front);
1203 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1204 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1205 obj->read_domains = 0;
1207 err = dma_resv_reserve_shared(vma->resv, 1);
1211 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1212 obj->write_domain = 0;
1215 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1216 i915_active_add_request(&vma->fence->active, rq);
1218 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1219 obj->mm.dirty = true;
1221 GEM_BUG_ON(!i915_vma_is_active(vma));
1225 int __i915_vma_unbind(struct i915_vma *vma)
1229 lockdep_assert_held(&vma->vm->mutex);
1231 if (i915_vma_is_pinned(vma)) {
1232 vma_print_allocator(vma, "is pinned");
1237 * After confirming that no one else is pinning this vma, wait for
1238 * any laggards who may have crept in during the wait (through
1239 * a residual pin skipping the vm->mutex) to complete.
1241 ret = i915_vma_sync(vma);
1245 if (!drm_mm_node_allocated(&vma->node))
1248 GEM_BUG_ON(i915_vma_is_pinned(vma));
1249 GEM_BUG_ON(i915_vma_is_active(vma));
1251 if (i915_vma_is_map_and_fenceable(vma)) {
1252 /* Force a pagefault for domain tracking on next user access */
1253 i915_vma_revoke_mmap(vma);
1256 * Check that we have flushed all writes through the GGTT
1257 * before the unbind, other due to non-strict nature of those
1258 * indirect writes they may end up referencing the GGTT PTE
1261 * Note that we may be concurrently poking at the GGTT_WRITE
1262 * bit from set-domain, as we mark all GGTT vma associated
1263 * with an object. We know this is for another vma, as we
1264 * are currently unbinding this one -- so if this vma will be
1265 * reused, it will be refaulted and have its dirty bit set
1266 * before the next write.
1268 i915_vma_flush_writes(vma);
1270 /* release the fence reg _after_ flushing */
1271 i915_vma_revoke_fence(vma);
1273 __i915_vma_iounmap(vma);
1274 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1276 GEM_BUG_ON(vma->fence);
1277 GEM_BUG_ON(i915_vma_has_userfault(vma));
1279 if (likely(atomic_read(&vma->vm->open))) {
1280 trace_i915_vma_unbind(vma);
1281 vma->ops->unbind_vma(vma);
1283 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1286 i915_vma_detach(vma);
1287 vma_unbind_pages(vma);
1289 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1293 int i915_vma_unbind(struct i915_vma *vma)
1295 struct i915_address_space *vm = vma->vm;
1296 intel_wakeref_t wakeref = 0;
1299 if (!drm_mm_node_allocated(&vma->node))
1302 /* Optimistic wait before taking the mutex */
1303 err = i915_vma_sync(vma);
1307 if (i915_vma_is_pinned(vma)) {
1308 vma_print_allocator(vma, "is pinned");
1312 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1313 /* XXX not always required: nop_clear_range */
1314 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1316 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1320 err = __i915_vma_unbind(vma);
1321 mutex_unlock(&vm->mutex);
1325 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1329 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1331 i915_gem_object_make_unshrinkable(vma->obj);
1335 void i915_vma_make_shrinkable(struct i915_vma *vma)
1337 i915_gem_object_make_shrinkable(vma->obj);
1340 void i915_vma_make_purgeable(struct i915_vma *vma)
1342 i915_gem_object_make_purgeable(vma->obj);
1345 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1346 #include "selftests/i915_vma.c"
1349 static void i915_global_vma_shrink(void)
1351 kmem_cache_shrink(global.slab_vmas);
1354 static void i915_global_vma_exit(void)
1356 kmem_cache_destroy(global.slab_vmas);
1359 static struct i915_global_vma global = { {
1360 .shrink = i915_global_vma_shrink,
1361 .exit = i915_global_vma_exit,
1364 int __init i915_global_vma_init(void)
1366 global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1367 if (!global.slab_vmas)
1370 i915_global_register(&global.base);