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 enum i915_cache_level cache_level;
301 static int __vma_bind(struct dma_fence_work *work)
303 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
304 struct i915_vma *vma = vw->vma;
307 err = vma->ops->bind_vma(vma, vw->cache_level, vw->flags);
309 atomic_or(I915_VMA_ERROR, &vma->flags);
314 static void __vma_release(struct dma_fence_work *work)
316 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
319 __i915_gem_object_unpin_pages(vw->pinned);
322 static const struct dma_fence_work_ops bind_ops = {
325 .release = __vma_release,
328 struct i915_vma_work *i915_vma_work(void)
330 struct i915_vma_work *vw;
332 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
336 dma_fence_work_init(&vw->base, &bind_ops);
337 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
343 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
345 * @cache_level: mapping cache level
346 * @flags: flags like global or local mapping
347 * @work: preallocated worker for allocating and binding the PTE
349 * DMA addresses are taken from the scatter-gather table of this object (or of
350 * this VMA in case of non-default GGTT views) and PTE entries set up.
351 * Note that DMA addresses are also the only part of the SG table we care about.
353 int i915_vma_bind(struct i915_vma *vma,
354 enum i915_cache_level cache_level,
356 struct i915_vma_work *work)
362 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
363 GEM_BUG_ON(vma->size > vma->node.size);
365 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
370 if (GEM_DEBUG_WARN_ON(!flags))
374 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
376 vma_flags = atomic_read(&vma->flags);
377 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
378 if (flags & PIN_UPDATE)
379 bind_flags |= vma_flags;
381 bind_flags &= ~vma_flags;
385 GEM_BUG_ON(!vma->pages);
387 trace_i915_vma_bind(vma, bind_flags);
388 if (work && (bind_flags & ~vma_flags) & vma->vm->bind_async_flags) {
390 work->cache_level = cache_level;
391 work->flags = bind_flags | I915_VMA_ALLOC;
394 * Note we only want to chain up to the migration fence on
395 * the pages (not the object itself). As we don't track that,
396 * yet, we have to use the exclusive fence instead.
398 * Also note that we do not want to track the async vma as
399 * part of the obj->resv->excl_fence as it only affects
400 * execution and not content or object's backing store lifetime.
402 GEM_BUG_ON(i915_active_has_exclusive(&vma->active));
403 i915_active_set_exclusive(&vma->active, &work->base.dma);
404 work->base.dma.error = 0; /* enable the queue_work() */
407 __i915_gem_object_pin_pages(vma->obj);
408 work->pinned = vma->obj;
411 GEM_BUG_ON((bind_flags & ~vma_flags) & vma->vm->bind_async_flags);
412 ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
417 atomic_or(bind_flags, &vma->flags);
421 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
426 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
431 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
432 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
434 ptr = READ_ONCE(vma->iomap);
436 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
444 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
445 io_mapping_unmap(ptr);
452 err = i915_vma_pin_fence(vma);
456 i915_vma_set_ggtt_write(vma);
458 /* NB Access through the GTT requires the device to be awake. */
462 __i915_vma_unpin(vma);
464 return IO_ERR_PTR(err);
467 void i915_vma_flush_writes(struct i915_vma *vma)
469 if (i915_vma_unset_ggtt_write(vma))
470 intel_gt_flush_ggtt_writes(vma->vm->gt);
473 void i915_vma_unpin_iomap(struct i915_vma *vma)
475 GEM_BUG_ON(vma->iomap == NULL);
477 i915_vma_flush_writes(vma);
479 i915_vma_unpin_fence(vma);
483 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
485 struct i915_vma *vma;
486 struct drm_i915_gem_object *obj;
488 vma = fetch_and_zero(p_vma);
498 if (flags & I915_VMA_RELEASE_MAP)
499 i915_gem_object_unpin_map(obj);
501 i915_gem_object_put(obj);
504 bool i915_vma_misplaced(const struct i915_vma *vma,
505 u64 size, u64 alignment, u64 flags)
507 if (!drm_mm_node_allocated(&vma->node))
510 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
513 if (vma->node.size < size)
516 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
517 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
520 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
523 if (flags & PIN_OFFSET_BIAS &&
524 vma->node.start < (flags & PIN_OFFSET_MASK))
527 if (flags & PIN_OFFSET_FIXED &&
528 vma->node.start != (flags & PIN_OFFSET_MASK))
534 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
536 bool mappable, fenceable;
538 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
539 GEM_BUG_ON(!vma->fence_size);
541 fenceable = (vma->node.size >= vma->fence_size &&
542 IS_ALIGNED(vma->node.start, vma->fence_alignment));
544 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
546 if (mappable && fenceable)
547 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
549 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
552 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
554 struct drm_mm_node *node = &vma->node;
555 struct drm_mm_node *other;
558 * On some machines we have to be careful when putting differing types
559 * of snoopable memory together to avoid the prefetcher crossing memory
560 * domains and dying. During vm initialisation, we decide whether or not
561 * these constraints apply and set the drm_mm.color_adjust
564 if (!i915_vm_has_cache_coloring(vma->vm))
567 /* Only valid to be called on an already inserted vma */
568 GEM_BUG_ON(!drm_mm_node_allocated(node));
569 GEM_BUG_ON(list_empty(&node->node_list));
571 other = list_prev_entry(node, node_list);
572 if (i915_node_color_differs(other, color) &&
573 !drm_mm_hole_follows(other))
576 other = list_next_entry(node, node_list);
577 if (i915_node_color_differs(other, color) &&
578 !drm_mm_hole_follows(node))
584 static void assert_bind_count(const struct drm_i915_gem_object *obj)
587 * Combine the assertion that the object is bound and that we have
588 * pinned its pages. But we should never have bound the object
589 * more than we have pinned its pages. (For complete accuracy, we
590 * assume that no else is pinning the pages, but as a rough assertion
591 * that we will not run into problems later, this will do!)
593 GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count));
597 * i915_vma_insert - finds a slot for the vma in its address space
599 * @size: requested size in bytes (can be larger than the VMA)
600 * @alignment: required alignment
601 * @flags: mask of PIN_* flags to use
603 * First we try to allocate some free space that meets the requirements for
604 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
605 * preferrably the oldest idle entry to make room for the new VMA.
608 * 0 on success, negative error code otherwise.
611 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
617 GEM_BUG_ON(i915_vma_is_closed(vma));
618 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
619 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
621 size = max(size, vma->size);
622 alignment = max(alignment, vma->display_alignment);
623 if (flags & PIN_MAPPABLE) {
624 size = max_t(typeof(size), size, vma->fence_size);
625 alignment = max_t(typeof(alignment),
626 alignment, vma->fence_alignment);
629 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
630 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
631 GEM_BUG_ON(!is_power_of_2(alignment));
633 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
634 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
636 end = vma->vm->total;
637 if (flags & PIN_MAPPABLE)
638 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
639 if (flags & PIN_ZONE_4G)
640 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
641 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
643 /* If binding the object/GGTT view requires more space than the entire
644 * aperture has, reject it early before evicting everything in a vain
645 * attempt to find space.
648 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
649 size, flags & PIN_MAPPABLE ? "mappable" : "total",
655 if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
656 color = vma->obj->cache_level;
658 if (flags & PIN_OFFSET_FIXED) {
659 u64 offset = flags & PIN_OFFSET_MASK;
660 if (!IS_ALIGNED(offset, alignment) ||
661 range_overflows(offset, size, end))
664 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
671 * We only support huge gtt pages through the 48b PPGTT,
672 * however we also don't want to force any alignment for
673 * objects which need to be tightly packed into the low 32bits.
675 * Note that we assume that GGTT are limited to 4GiB for the
676 * forseeable future. See also i915_ggtt_offset().
678 if (upper_32_bits(end - 1) &&
679 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
681 * We can't mix 64K and 4K PTEs in the same page-table
682 * (2M block), and so to avoid the ugliness and
683 * complexity of coloring we opt for just aligning 64K
687 rounddown_pow_of_two(vma->page_sizes.sg |
688 I915_GTT_PAGE_SIZE_2M);
691 * Check we don't expand for the limited Global GTT
692 * (mappable aperture is even more precious!). This
693 * also checks that we exclude the aliasing-ppgtt.
695 GEM_BUG_ON(i915_vma_is_ggtt(vma));
697 alignment = max(alignment, page_alignment);
699 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
700 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
703 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
704 size, alignment, color,
709 GEM_BUG_ON(vma->node.start < start);
710 GEM_BUG_ON(vma->node.start + vma->node.size > end);
712 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
713 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
716 struct drm_i915_gem_object *obj = vma->obj;
718 atomic_inc(&obj->bind_count);
719 assert_bind_count(obj);
721 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
727 i915_vma_detach(struct i915_vma *vma)
729 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
730 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
733 * And finally now the object is completely decoupled from this
734 * vma, we can drop its hold on the backing storage and allow
735 * it to be reaped by the shrinker.
737 list_del(&vma->vm_link);
739 struct drm_i915_gem_object *obj = vma->obj;
741 assert_bind_count(obj);
742 atomic_dec(&obj->bind_count);
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)
793 if (atomic_add_unless(&vma->pages_count, 1, 0))
796 /* Allocations ahoy! */
797 if (mutex_lock_interruptible(&vma->pages_mutex))
800 if (!atomic_read(&vma->pages_count)) {
802 err = i915_gem_object_pin_pages(vma->obj);
807 err = vma->ops->set_pages(vma);
810 i915_gem_object_unpin_pages(vma->obj);
814 atomic_inc(&vma->pages_count);
817 mutex_unlock(&vma->pages_mutex);
822 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
824 /* We allocate under vma_get_pages, so beware the shrinker */
825 mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
826 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
827 if (atomic_sub_return(count, &vma->pages_count) == 0) {
828 vma->ops->clear_pages(vma);
829 GEM_BUG_ON(vma->pages);
831 i915_gem_object_unpin_pages(vma->obj);
833 mutex_unlock(&vma->pages_mutex);
836 static void vma_put_pages(struct i915_vma *vma)
838 if (atomic_add_unless(&vma->pages_count, -1, 1))
841 __vma_put_pages(vma, 1);
844 static void vma_unbind_pages(struct i915_vma *vma)
848 lockdep_assert_held(&vma->vm->mutex);
850 /* The upper portion of pages_count is the number of bindings */
851 count = atomic_read(&vma->pages_count);
852 count >>= I915_VMA_PAGES_BIAS;
855 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
858 int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
860 struct i915_vma_work *work = NULL;
861 intel_wakeref_t wakeref = 0;
865 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
866 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
868 GEM_BUG_ON(flags & PIN_UPDATE);
869 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
871 /* First try and grab the pin without rebinding the vma */
872 if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
875 err = vma_get_pages(vma);
879 if (flags & vma->vm->bind_async_flags) {
880 work = i915_vma_work();
887 if (flags & PIN_GLOBAL)
888 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
890 /* No more allocations allowed once we hold vm->mutex */
891 err = mutex_lock_interruptible(&vma->vm->mutex);
895 bound = atomic_read(&vma->flags);
896 if (unlikely(bound & I915_VMA_ERROR)) {
901 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
902 err = -EAGAIN; /* pins are meant to be fairly temporary */
906 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
911 err = i915_active_acquire(&vma->active);
915 if (!(bound & I915_VMA_BIND_MASK)) {
916 err = i915_vma_insert(vma, size, alignment, flags);
920 if (i915_is_ggtt(vma->vm))
921 __i915_vma_set_map_and_fenceable(vma);
924 GEM_BUG_ON(!vma->pages);
925 err = i915_vma_bind(vma,
926 vma->obj ? vma->obj->cache_level : 0,
931 /* There should only be at most 2 active bindings (user, global) */
932 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
933 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
934 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
937 GEM_BUG_ON(!i915_vma_is_pinned(vma));
938 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
939 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
942 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
943 i915_vma_detach(vma);
944 drm_mm_remove_node(&vma->node);
947 i915_active_release(&vma->active);
949 mutex_unlock(&vma->vm->mutex);
952 dma_fence_work_commit(&work->base);
954 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
960 static void flush_idle_contexts(struct intel_gt *gt)
962 struct intel_engine_cs *engine;
963 enum intel_engine_id id;
965 for_each_engine(engine, gt, id)
966 intel_engine_flush_barriers(engine);
968 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
971 int i915_ggtt_pin(struct i915_vma *vma, u32 align, unsigned int flags)
973 struct i915_address_space *vm = vma->vm;
976 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
979 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
983 /* Unlike i915_vma_pin, we don't take no for an answer! */
984 flush_idle_contexts(vm->gt);
985 if (mutex_lock_interruptible(&vm->mutex) == 0) {
986 i915_gem_evict_vm(vm);
987 mutex_unlock(&vm->mutex);
992 void i915_vma_close(struct i915_vma *vma)
994 struct intel_gt *gt = vma->vm->gt;
997 GEM_BUG_ON(i915_vma_is_closed(vma));
1000 * We defer actually closing, unbinding and destroying the VMA until
1001 * the next idle point, or if the object is freed in the meantime. By
1002 * postponing the unbind, we allow for it to be resurrected by the
1003 * client, avoiding the work required to rebind the VMA. This is
1004 * advantageous for DRI, where the client/server pass objects
1005 * between themselves, temporarily opening a local VMA to the
1006 * object, and then closing it again. The same object is then reused
1007 * on the next frame (or two, depending on the depth of the swap queue)
1008 * causing us to rebind the VMA once more. This ends up being a lot
1009 * of wasted work for the steady state.
1011 spin_lock_irqsave(>->closed_lock, flags);
1012 list_add(&vma->closed_link, >->closed_vma);
1013 spin_unlock_irqrestore(>->closed_lock, flags);
1016 static void __i915_vma_remove_closed(struct i915_vma *vma)
1018 struct intel_gt *gt = vma->vm->gt;
1020 spin_lock_irq(>->closed_lock);
1021 list_del_init(&vma->closed_link);
1022 spin_unlock_irq(>->closed_lock);
1025 void i915_vma_reopen(struct i915_vma *vma)
1027 if (i915_vma_is_closed(vma))
1028 __i915_vma_remove_closed(vma);
1031 void i915_vma_release(struct kref *ref)
1033 struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1035 if (drm_mm_node_allocated(&vma->node)) {
1036 mutex_lock(&vma->vm->mutex);
1037 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1038 WARN_ON(__i915_vma_unbind(vma));
1039 mutex_unlock(&vma->vm->mutex);
1040 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1042 GEM_BUG_ON(i915_vma_is_active(vma));
1045 struct drm_i915_gem_object *obj = vma->obj;
1047 spin_lock(&obj->vma.lock);
1048 list_del(&vma->obj_link);
1049 rb_erase(&vma->obj_node, &obj->vma.tree);
1050 spin_unlock(&obj->vma.lock);
1053 __i915_vma_remove_closed(vma);
1054 i915_vm_put(vma->vm);
1056 i915_active_fini(&vma->active);
1060 void i915_vma_parked(struct intel_gt *gt)
1062 struct i915_vma *vma, *next;
1064 spin_lock_irq(>->closed_lock);
1065 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1066 struct drm_i915_gem_object *obj = vma->obj;
1067 struct i915_address_space *vm = vma->vm;
1069 /* XXX All to avoid keeping a reference on i915_vma itself */
1071 if (!kref_get_unless_zero(&obj->base.refcount))
1074 if (i915_vm_tryopen(vm)) {
1075 list_del_init(&vma->closed_link);
1077 i915_gem_object_put(obj);
1081 spin_unlock_irq(>->closed_lock);
1084 __i915_vma_put(vma);
1085 i915_gem_object_put(obj);
1090 /* Restart after dropping lock */
1091 spin_lock_irq(>->closed_lock);
1092 next = list_first_entry(>->closed_vma,
1093 typeof(*next), closed_link);
1095 spin_unlock_irq(>->closed_lock);
1098 static void __i915_vma_iounmap(struct i915_vma *vma)
1100 GEM_BUG_ON(i915_vma_is_pinned(vma));
1102 if (vma->iomap == NULL)
1105 io_mapping_unmap(vma->iomap);
1109 void i915_vma_revoke_mmap(struct i915_vma *vma)
1111 struct drm_vma_offset_node *node;
1114 if (!i915_vma_has_userfault(vma))
1117 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1118 GEM_BUG_ON(!vma->obj->userfault_count);
1120 node = &vma->mmo->vma_node;
1121 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1122 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1123 drm_vma_node_offset_addr(node) + vma_offset,
1127 i915_vma_unset_userfault(vma);
1128 if (!--vma->obj->userfault_count)
1129 list_del(&vma->obj->userfault_link);
1132 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1136 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1138 /* Wait for the vma to be bound before we start! */
1139 err = i915_request_await_active(rq, &vma->active);
1143 return i915_active_add_request(&vma->active, rq);
1146 int i915_vma_move_to_active(struct i915_vma *vma,
1147 struct i915_request *rq,
1150 struct drm_i915_gem_object *obj = vma->obj;
1153 assert_object_held(obj);
1155 err = __i915_vma_move_to_active(vma, rq);
1159 if (flags & EXEC_OBJECT_WRITE) {
1160 struct intel_frontbuffer *front;
1162 front = __intel_frontbuffer_get(obj);
1163 if (unlikely(front)) {
1164 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1165 i915_active_add_request(&front->write, rq);
1166 intel_frontbuffer_put(front);
1169 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1170 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1171 obj->read_domains = 0;
1173 err = dma_resv_reserve_shared(vma->resv, 1);
1177 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1178 obj->write_domain = 0;
1180 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1181 obj->mm.dirty = true;
1183 GEM_BUG_ON(!i915_vma_is_active(vma));
1187 int __i915_vma_unbind(struct i915_vma *vma)
1191 lockdep_assert_held(&vma->vm->mutex);
1194 * First wait upon any activity as retiring the request may
1195 * have side-effects such as unpinning or even unbinding this vma.
1197 * XXX Actually waiting under the vm->mutex is a hinderance and
1198 * should be pipelined wherever possible. In cases where that is
1199 * unavoidable, we should lift the wait to before the mutex.
1201 ret = i915_vma_sync(vma);
1205 if (i915_vma_is_pinned(vma)) {
1206 vma_print_allocator(vma, "is pinned");
1211 * After confirming that no one else is pinning this vma, wait for
1212 * any laggards who may have crept in during the wait (through
1213 * a residual pin skipping the vm->mutex) to complete.
1215 ret = i915_vma_sync(vma);
1219 if (!drm_mm_node_allocated(&vma->node))
1222 GEM_BUG_ON(i915_vma_is_pinned(vma));
1223 GEM_BUG_ON(i915_vma_is_active(vma));
1225 if (i915_vma_is_map_and_fenceable(vma)) {
1227 * Check that we have flushed all writes through the GGTT
1228 * before the unbind, other due to non-strict nature of those
1229 * indirect writes they may end up referencing the GGTT PTE
1232 i915_vma_flush_writes(vma);
1233 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1235 /* release the fence reg _after_ flushing */
1236 ret = i915_vma_revoke_fence(vma);
1240 /* Force a pagefault for domain tracking on next user access */
1241 i915_vma_revoke_mmap(vma);
1243 __i915_vma_iounmap(vma);
1244 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1246 GEM_BUG_ON(vma->fence);
1247 GEM_BUG_ON(i915_vma_has_userfault(vma));
1249 if (likely(atomic_read(&vma->vm->open))) {
1250 trace_i915_vma_unbind(vma);
1251 vma->ops->unbind_vma(vma);
1253 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR), &vma->flags);
1255 i915_vma_detach(vma);
1256 vma_unbind_pages(vma);
1258 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1262 int i915_vma_unbind(struct i915_vma *vma)
1264 struct i915_address_space *vm = vma->vm;
1265 intel_wakeref_t wakeref = 0;
1268 if (!drm_mm_node_allocated(&vma->node))
1271 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1272 /* XXX not always required: nop_clear_range */
1273 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1275 err = mutex_lock_interruptible(&vm->mutex);
1279 err = __i915_vma_unbind(vma);
1280 mutex_unlock(&vm->mutex);
1283 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1288 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1290 i915_gem_object_make_unshrinkable(vma->obj);
1294 void i915_vma_make_shrinkable(struct i915_vma *vma)
1296 i915_gem_object_make_shrinkable(vma->obj);
1299 void i915_vma_make_purgeable(struct i915_vma *vma)
1301 i915_gem_object_make_purgeable(vma->obj);
1304 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1305 #include "selftests/i915_vma.c"
1308 static void i915_global_vma_shrink(void)
1310 kmem_cache_shrink(global.slab_vmas);
1313 static void i915_global_vma_exit(void)
1315 kmem_cache_destroy(global.slab_vmas);
1318 static struct i915_global_vma global = { {
1319 .shrink = i915_global_vma_shrink,
1320 .exit = i915_global_vma_exit,
1323 int __init i915_global_vma_init(void)
1325 global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1326 if (!global.slab_vmas)
1329 i915_global_register(&global.base);