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
))
613 static void assert_bind_count(const struct drm_i915_gem_object
*obj
)
616 * Combine the assertion that the object is bound and that we have
617 * pinned its pages. But we should never have bound the object
618 * more than we have pinned its pages. (For complete accuracy, we
619 * assume that no else is pinning the pages, but as a rough assertion
620 * that we will not run into problems later, this will do!)
622 GEM_BUG_ON(atomic_read(&obj
->mm
.pages_pin_count
) < atomic_read(&obj
->bind_count
));
626 * i915_vma_insert - finds a slot for the vma in its address space
628 * @size: requested size in bytes (can be larger than the VMA)
629 * @alignment: required alignment
630 * @flags: mask of PIN_* flags to use
632 * First we try to allocate some free space that meets the requirements for
633 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
634 * preferrably the oldest idle entry to make room for the new VMA.
637 * 0 on success, negative error code otherwise.
640 i915_vma_insert(struct i915_vma
*vma
, u64 size
, u64 alignment
, u64 flags
)
646 GEM_BUG_ON(i915_vma_is_bound(vma
, I915_VMA_GLOBAL_BIND
| I915_VMA_LOCAL_BIND
));
647 GEM_BUG_ON(drm_mm_node_allocated(&vma
->node
));
649 size
= max(size
, vma
->size
);
650 alignment
= max(alignment
, vma
->display_alignment
);
651 if (flags
& PIN_MAPPABLE
) {
652 size
= max_t(typeof(size
), size
, vma
->fence_size
);
653 alignment
= max_t(typeof(alignment
),
654 alignment
, vma
->fence_alignment
);
657 GEM_BUG_ON(!IS_ALIGNED(size
, I915_GTT_PAGE_SIZE
));
658 GEM_BUG_ON(!IS_ALIGNED(alignment
, I915_GTT_MIN_ALIGNMENT
));
659 GEM_BUG_ON(!is_power_of_2(alignment
));
661 start
= flags
& PIN_OFFSET_BIAS
? flags
& PIN_OFFSET_MASK
: 0;
662 GEM_BUG_ON(!IS_ALIGNED(start
, I915_GTT_PAGE_SIZE
));
664 end
= vma
->vm
->total
;
665 if (flags
& PIN_MAPPABLE
)
666 end
= min_t(u64
, end
, i915_vm_to_ggtt(vma
->vm
)->mappable_end
);
667 if (flags
& PIN_ZONE_4G
)
668 end
= min_t(u64
, end
, (1ULL << 32) - I915_GTT_PAGE_SIZE
);
669 GEM_BUG_ON(!IS_ALIGNED(end
, I915_GTT_PAGE_SIZE
));
671 /* If binding the object/GGTT view requires more space than the entire
672 * aperture has, reject it early before evicting everything in a vain
673 * attempt to find space.
676 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
677 size
, flags
& PIN_MAPPABLE
? "mappable" : "total",
683 if (vma
->obj
&& i915_vm_has_cache_coloring(vma
->vm
))
684 color
= vma
->obj
->cache_level
;
686 if (flags
& PIN_OFFSET_FIXED
) {
687 u64 offset
= flags
& PIN_OFFSET_MASK
;
688 if (!IS_ALIGNED(offset
, alignment
) ||
689 range_overflows(offset
, size
, end
))
692 ret
= i915_gem_gtt_reserve(vma
->vm
, &vma
->node
,
699 * We only support huge gtt pages through the 48b PPGTT,
700 * however we also don't want to force any alignment for
701 * objects which need to be tightly packed into the low 32bits.
703 * Note that we assume that GGTT are limited to 4GiB for the
704 * forseeable future. See also i915_ggtt_offset().
706 if (upper_32_bits(end
- 1) &&
707 vma
->page_sizes
.sg
> I915_GTT_PAGE_SIZE
) {
709 * We can't mix 64K and 4K PTEs in the same page-table
710 * (2M block), and so to avoid the ugliness and
711 * complexity of coloring we opt for just aligning 64K
715 rounddown_pow_of_two(vma
->page_sizes
.sg
|
716 I915_GTT_PAGE_SIZE_2M
);
719 * Check we don't expand for the limited Global GTT
720 * (mappable aperture is even more precious!). This
721 * also checks that we exclude the aliasing-ppgtt.
723 GEM_BUG_ON(i915_vma_is_ggtt(vma
));
725 alignment
= max(alignment
, page_alignment
);
727 if (vma
->page_sizes
.sg
& I915_GTT_PAGE_SIZE_64K
)
728 size
= round_up(size
, I915_GTT_PAGE_SIZE_2M
);
731 ret
= i915_gem_gtt_insert(vma
->vm
, &vma
->node
,
732 size
, alignment
, color
,
737 GEM_BUG_ON(vma
->node
.start
< start
);
738 GEM_BUG_ON(vma
->node
.start
+ vma
->node
.size
> end
);
740 GEM_BUG_ON(!drm_mm_node_allocated(&vma
->node
));
741 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma
, color
));
744 struct drm_i915_gem_object
*obj
= vma
->obj
;
746 atomic_inc(&obj
->bind_count
);
747 assert_bind_count(obj
);
749 list_add_tail(&vma
->vm_link
, &vma
->vm
->bound_list
);
755 i915_vma_detach(struct i915_vma
*vma
)
757 GEM_BUG_ON(!drm_mm_node_allocated(&vma
->node
));
758 GEM_BUG_ON(i915_vma_is_bound(vma
, I915_VMA_GLOBAL_BIND
| I915_VMA_LOCAL_BIND
));
761 * And finally now the object is completely decoupled from this
762 * vma, we can drop its hold on the backing storage and allow
763 * it to be reaped by the shrinker.
765 list_del(&vma
->vm_link
);
767 struct drm_i915_gem_object
*obj
= vma
->obj
;
769 assert_bind_count(obj
);
770 atomic_dec(&obj
->bind_count
);
774 static bool try_qad_pin(struct i915_vma
*vma
, unsigned int flags
)
779 bound
= atomic_read(&vma
->flags
);
781 if (unlikely(flags
& ~bound
))
784 if (unlikely(bound
& (I915_VMA_OVERFLOW
| I915_VMA_ERROR
)))
787 if (!(bound
& I915_VMA_PIN_MASK
))
790 GEM_BUG_ON(((bound
+ 1) & I915_VMA_PIN_MASK
) == 0);
791 } while (!atomic_try_cmpxchg(&vma
->flags
, &bound
, bound
+ 1));
797 * If pin_count==0, but we are bound, check under the lock to avoid
798 * racing with a concurrent i915_vma_unbind().
800 mutex_lock(&vma
->vm
->mutex
);
802 if (unlikely(bound
& (I915_VMA_OVERFLOW
| I915_VMA_ERROR
))) {
807 if (unlikely(flags
& ~bound
)) {
811 } while (!atomic_try_cmpxchg(&vma
->flags
, &bound
, bound
+ 1));
812 mutex_unlock(&vma
->vm
->mutex
);
817 static int vma_get_pages(struct i915_vma
*vma
)
821 if (atomic_add_unless(&vma
->pages_count
, 1, 0))
824 /* Allocations ahoy! */
825 if (mutex_lock_interruptible(&vma
->pages_mutex
))
828 if (!atomic_read(&vma
->pages_count
)) {
830 err
= i915_gem_object_pin_pages(vma
->obj
);
835 err
= vma
->ops
->set_pages(vma
);
838 i915_gem_object_unpin_pages(vma
->obj
);
842 atomic_inc(&vma
->pages_count
);
845 mutex_unlock(&vma
->pages_mutex
);
850 static void __vma_put_pages(struct i915_vma
*vma
, unsigned int count
)
852 /* We allocate under vma_get_pages, so beware the shrinker */
853 mutex_lock_nested(&vma
->pages_mutex
, SINGLE_DEPTH_NESTING
);
854 GEM_BUG_ON(atomic_read(&vma
->pages_count
) < count
);
855 if (atomic_sub_return(count
, &vma
->pages_count
) == 0) {
856 vma
->ops
->clear_pages(vma
);
857 GEM_BUG_ON(vma
->pages
);
859 i915_gem_object_unpin_pages(vma
->obj
);
861 mutex_unlock(&vma
->pages_mutex
);
864 static void vma_put_pages(struct i915_vma
*vma
)
866 if (atomic_add_unless(&vma
->pages_count
, -1, 1))
869 __vma_put_pages(vma
, 1);
872 static void vma_unbind_pages(struct i915_vma
*vma
)
876 lockdep_assert_held(&vma
->vm
->mutex
);
878 /* The upper portion of pages_count is the number of bindings */
879 count
= atomic_read(&vma
->pages_count
);
880 count
>>= I915_VMA_PAGES_BIAS
;
883 __vma_put_pages(vma
, count
| count
<< I915_VMA_PAGES_BIAS
);
886 int i915_vma_pin(struct i915_vma
*vma
, u64 size
, u64 alignment
, u64 flags
)
888 struct i915_vma_work
*work
= NULL
;
889 intel_wakeref_t wakeref
= 0;
893 BUILD_BUG_ON(PIN_GLOBAL
!= I915_VMA_GLOBAL_BIND
);
894 BUILD_BUG_ON(PIN_USER
!= I915_VMA_LOCAL_BIND
);
896 GEM_BUG_ON(flags
& PIN_UPDATE
);
897 GEM_BUG_ON(!(flags
& (PIN_USER
| PIN_GLOBAL
)));
899 /* First try and grab the pin without rebinding the vma */
900 if (try_qad_pin(vma
, flags
& I915_VMA_BIND_MASK
))
903 err
= vma_get_pages(vma
);
907 if (flags
& vma
->vm
->bind_async_flags
) {
908 work
= i915_vma_work();
915 if (flags
& PIN_GLOBAL
)
916 wakeref
= intel_runtime_pm_get(&vma
->vm
->i915
->runtime_pm
);
918 /* No more allocations allowed once we hold vm->mutex */
919 err
= mutex_lock_interruptible(&vma
->vm
->mutex
);
923 if (unlikely(i915_vma_is_closed(vma
))) {
928 bound
= atomic_read(&vma
->flags
);
929 if (unlikely(bound
& I915_VMA_ERROR
)) {
934 if (unlikely(!((bound
+ 1) & I915_VMA_PIN_MASK
))) {
935 err
= -EAGAIN
; /* pins are meant to be fairly temporary */
939 if (unlikely(!(flags
& ~bound
& I915_VMA_BIND_MASK
))) {
944 err
= i915_active_acquire(&vma
->active
);
948 if (!(bound
& I915_VMA_BIND_MASK
)) {
949 err
= i915_vma_insert(vma
, size
, alignment
, flags
);
953 if (i915_is_ggtt(vma
->vm
))
954 __i915_vma_set_map_and_fenceable(vma
);
957 GEM_BUG_ON(!vma
->pages
);
958 err
= i915_vma_bind(vma
,
959 vma
->obj
? vma
->obj
->cache_level
: 0,
964 /* There should only be at most 2 active bindings (user, global) */
965 GEM_BUG_ON(bound
+ I915_VMA_PAGES_ACTIVE
< bound
);
966 atomic_add(I915_VMA_PAGES_ACTIVE
, &vma
->pages_count
);
967 list_move_tail(&vma
->vm_link
, &vma
->vm
->bound_list
);
970 GEM_BUG_ON(!i915_vma_is_pinned(vma
));
971 GEM_BUG_ON(!i915_vma_is_bound(vma
, flags
));
972 GEM_BUG_ON(i915_vma_misplaced(vma
, size
, alignment
, flags
));
975 if (!i915_vma_is_bound(vma
, I915_VMA_BIND_MASK
)) {
976 i915_vma_detach(vma
);
977 drm_mm_remove_node(&vma
->node
);
980 i915_active_release(&vma
->active
);
982 mutex_unlock(&vma
->vm
->mutex
);
985 dma_fence_work_commit(&work
->base
);
987 intel_runtime_pm_put(&vma
->vm
->i915
->runtime_pm
, wakeref
);
993 static void flush_idle_contexts(struct intel_gt
*gt
)
995 struct intel_engine_cs
*engine
;
996 enum intel_engine_id id
;
998 for_each_engine(engine
, gt
, id
)
999 intel_engine_flush_barriers(engine
);
1001 intel_gt_wait_for_idle(gt
, MAX_SCHEDULE_TIMEOUT
);
1004 int i915_ggtt_pin(struct i915_vma
*vma
, u32 align
, unsigned int flags
)
1006 struct i915_address_space
*vm
= vma
->vm
;
1009 GEM_BUG_ON(!i915_vma_is_ggtt(vma
));
1012 err
= i915_vma_pin(vma
, 0, align
, flags
| PIN_GLOBAL
);
1013 if (err
!= -ENOSPC
) {
1015 err
= i915_vma_wait_for_bind(vma
);
1017 i915_vma_unpin(vma
);
1022 /* Unlike i915_vma_pin, we don't take no for an answer! */
1023 flush_idle_contexts(vm
->gt
);
1024 if (mutex_lock_interruptible(&vm
->mutex
) == 0) {
1025 i915_gem_evict_vm(vm
);
1026 mutex_unlock(&vm
->mutex
);
1031 void i915_vma_close(struct i915_vma
*vma
)
1033 struct intel_gt
*gt
= vma
->vm
->gt
;
1034 unsigned long flags
;
1036 GEM_BUG_ON(i915_vma_is_closed(vma
));
1039 * We defer actually closing, unbinding and destroying the VMA until
1040 * the next idle point, or if the object is freed in the meantime. By
1041 * postponing the unbind, we allow for it to be resurrected by the
1042 * client, avoiding the work required to rebind the VMA. This is
1043 * advantageous for DRI, where the client/server pass objects
1044 * between themselves, temporarily opening a local VMA to the
1045 * object, and then closing it again. The same object is then reused
1046 * on the next frame (or two, depending on the depth of the swap queue)
1047 * causing us to rebind the VMA once more. This ends up being a lot
1048 * of wasted work for the steady state.
1050 spin_lock_irqsave(>
->closed_lock
, flags
);
1051 list_add(&vma
->closed_link
, >
->closed_vma
);
1052 spin_unlock_irqrestore(>
->closed_lock
, flags
);
1055 static void __i915_vma_remove_closed(struct i915_vma
*vma
)
1057 struct intel_gt
*gt
= vma
->vm
->gt
;
1059 spin_lock_irq(>
->closed_lock
);
1060 list_del_init(&vma
->closed_link
);
1061 spin_unlock_irq(>
->closed_lock
);
1064 void i915_vma_reopen(struct i915_vma
*vma
)
1066 if (i915_vma_is_closed(vma
))
1067 __i915_vma_remove_closed(vma
);
1070 void i915_vma_release(struct kref
*ref
)
1072 struct i915_vma
*vma
= container_of(ref
, typeof(*vma
), ref
);
1074 if (drm_mm_node_allocated(&vma
->node
)) {
1075 mutex_lock(&vma
->vm
->mutex
);
1076 atomic_and(~I915_VMA_PIN_MASK
, &vma
->flags
);
1077 WARN_ON(__i915_vma_unbind(vma
));
1078 mutex_unlock(&vma
->vm
->mutex
);
1079 GEM_BUG_ON(drm_mm_node_allocated(&vma
->node
));
1081 GEM_BUG_ON(i915_vma_is_active(vma
));
1084 struct drm_i915_gem_object
*obj
= vma
->obj
;
1086 spin_lock(&obj
->vma
.lock
);
1087 list_del(&vma
->obj_link
);
1088 rb_erase(&vma
->obj_node
, &obj
->vma
.tree
);
1089 spin_unlock(&obj
->vma
.lock
);
1092 __i915_vma_remove_closed(vma
);
1093 i915_vm_put(vma
->vm
);
1095 i915_active_fini(&vma
->active
);
1099 void i915_vma_parked(struct intel_gt
*gt
)
1101 struct i915_vma
*vma
, *next
;
1104 spin_lock_irq(>
->closed_lock
);
1105 list_for_each_entry_safe(vma
, next
, >
->closed_vma
, closed_link
) {
1106 struct drm_i915_gem_object
*obj
= vma
->obj
;
1107 struct i915_address_space
*vm
= vma
->vm
;
1109 /* XXX All to avoid keeping a reference on i915_vma itself */
1111 if (!kref_get_unless_zero(&obj
->base
.refcount
))
1114 if (!i915_vm_tryopen(vm
)) {
1115 i915_gem_object_put(obj
);
1119 list_move(&vma
->closed_link
, &closed
);
1121 spin_unlock_irq(>
->closed_lock
);
1123 /* As the GT is held idle, no vma can be reopened as we destroy them */
1124 list_for_each_entry_safe(vma
, next
, &closed
, closed_link
) {
1125 struct drm_i915_gem_object
*obj
= vma
->obj
;
1126 struct i915_address_space
*vm
= vma
->vm
;
1128 INIT_LIST_HEAD(&vma
->closed_link
);
1129 __i915_vma_put(vma
);
1131 i915_gem_object_put(obj
);
1136 static void __i915_vma_iounmap(struct i915_vma
*vma
)
1138 GEM_BUG_ON(i915_vma_is_pinned(vma
));
1140 if (vma
->iomap
== NULL
)
1143 io_mapping_unmap(vma
->iomap
);
1147 void i915_vma_revoke_mmap(struct i915_vma
*vma
)
1149 struct drm_vma_offset_node
*node
;
1152 if (!i915_vma_has_userfault(vma
))
1155 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma
));
1156 GEM_BUG_ON(!vma
->obj
->userfault_count
);
1158 node
= &vma
->mmo
->vma_node
;
1159 vma_offset
= vma
->ggtt_view
.partial
.offset
<< PAGE_SHIFT
;
1160 unmap_mapping_range(vma
->vm
->i915
->drm
.anon_inode
->i_mapping
,
1161 drm_vma_node_offset_addr(node
) + vma_offset
,
1165 i915_vma_unset_userfault(vma
);
1166 if (!--vma
->obj
->userfault_count
)
1167 list_del(&vma
->obj
->userfault_link
);
1170 int __i915_vma_move_to_active(struct i915_vma
*vma
, struct i915_request
*rq
)
1174 GEM_BUG_ON(!i915_vma_is_pinned(vma
));
1176 /* Wait for the vma to be bound before we start! */
1177 err
= i915_request_await_active(rq
, &vma
->active
, 0);
1181 return i915_active_add_request(&vma
->active
, rq
);
1184 int i915_vma_move_to_active(struct i915_vma
*vma
,
1185 struct i915_request
*rq
,
1188 struct drm_i915_gem_object
*obj
= vma
->obj
;
1191 assert_object_held(obj
);
1193 err
= __i915_vma_move_to_active(vma
, rq
);
1197 if (flags
& EXEC_OBJECT_WRITE
) {
1198 struct intel_frontbuffer
*front
;
1200 front
= __intel_frontbuffer_get(obj
);
1201 if (unlikely(front
)) {
1202 if (intel_frontbuffer_invalidate(front
, ORIGIN_CS
))
1203 i915_active_add_request(&front
->write
, rq
);
1204 intel_frontbuffer_put(front
);
1207 dma_resv_add_excl_fence(vma
->resv
, &rq
->fence
);
1208 obj
->write_domain
= I915_GEM_DOMAIN_RENDER
;
1209 obj
->read_domains
= 0;
1211 err
= dma_resv_reserve_shared(vma
->resv
, 1);
1215 dma_resv_add_shared_fence(vma
->resv
, &rq
->fence
);
1216 obj
->write_domain
= 0;
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
)) {
1253 * Check that we have flushed all writes through the GGTT
1254 * before the unbind, other due to non-strict nature of those
1255 * indirect writes they may end up referencing the GGTT PTE
1258 * Note that we may be concurrently poking at the GGTT_WRITE
1259 * bit from set-domain, as we mark all GGTT vma associated
1260 * with an object. We know this is for another vma, as we
1261 * are currently unbinding this one -- so if this vma will be
1262 * reused, it will be refaulted and have its dirty bit set
1263 * before the next write.
1265 i915_vma_flush_writes(vma
);
1267 /* release the fence reg _after_ flushing */
1268 ret
= i915_vma_revoke_fence(vma
);
1272 /* Force a pagefault for domain tracking on next user access */
1273 i915_vma_revoke_mmap(vma
);
1275 __i915_vma_iounmap(vma
);
1276 clear_bit(I915_VMA_CAN_FENCE_BIT
, __i915_vma_flags(vma
));
1278 GEM_BUG_ON(vma
->fence
);
1279 GEM_BUG_ON(i915_vma_has_userfault(vma
));
1281 if (likely(atomic_read(&vma
->vm
->open
))) {
1282 trace_i915_vma_unbind(vma
);
1283 vma
->ops
->unbind_vma(vma
);
1285 atomic_and(~(I915_VMA_BIND_MASK
| I915_VMA_ERROR
| I915_VMA_GGTT_WRITE
),
1288 i915_vma_detach(vma
);
1289 vma_unbind_pages(vma
);
1291 drm_mm_remove_node(&vma
->node
); /* pairs with i915_vma_release() */
1295 int i915_vma_unbind(struct i915_vma
*vma
)
1297 struct i915_address_space
*vm
= vma
->vm
;
1298 intel_wakeref_t wakeref
= 0;
1301 if (!drm_mm_node_allocated(&vma
->node
))
1304 /* Optimistic wait before taking the mutex */
1305 err
= i915_vma_sync(vma
);
1309 if (i915_vma_is_pinned(vma
)) {
1310 vma_print_allocator(vma
, "is pinned");
1314 if (i915_vma_is_bound(vma
, I915_VMA_GLOBAL_BIND
))
1315 /* XXX not always required: nop_clear_range */
1316 wakeref
= intel_runtime_pm_get(&vm
->i915
->runtime_pm
);
1318 err
= mutex_lock_interruptible(&vm
->mutex
);
1322 err
= __i915_vma_unbind(vma
);
1323 mutex_unlock(&vm
->mutex
);
1327 intel_runtime_pm_put(&vm
->i915
->runtime_pm
, wakeref
);
1331 struct i915_vma
*i915_vma_make_unshrinkable(struct i915_vma
*vma
)
1333 i915_gem_object_make_unshrinkable(vma
->obj
);
1337 void i915_vma_make_shrinkable(struct i915_vma
*vma
)
1339 i915_gem_object_make_shrinkable(vma
->obj
);
1342 void i915_vma_make_purgeable(struct i915_vma
*vma
)
1344 i915_gem_object_make_purgeable(vma
->obj
);
1347 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1348 #include "selftests/i915_vma.c"
1351 static void i915_global_vma_shrink(void)
1353 kmem_cache_shrink(global
.slab_vmas
);
1356 static void i915_global_vma_exit(void)
1358 kmem_cache_destroy(global
.slab_vmas
);
1361 static struct i915_global_vma global
= { {
1362 .shrink
= i915_global_vma_shrink
,
1363 .exit
= i915_global_vma_exit
,
1366 int __init
i915_global_vma_init(void)
1368 global
.slab_vmas
= KMEM_CACHE(i915_vma
, SLAB_HWCACHE_ALIGN
);
1369 if (!global
.slab_vmas
)
1372 i915_global_register(&global
.base
);