2 * Copyright © 2008,2010 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
24 * Eric Anholt <eric@anholt.net>
25 * Chris Wilson <chris@chris-wilson.co.uk>
30 #include <drm/i915_drm.h>
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/dma_remapping.h>
36 #define __EXEC_OBJECT_HAS_PIN (1<<31)
37 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
38 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
39 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
41 #define BATCH_OFFSET_BIAS (256*1024)
44 struct list_head vmas
;
47 struct i915_vma
*lut
[0];
48 struct hlist_head buckets
[0];
52 static struct eb_vmas
*
53 eb_create(struct drm_i915_gem_execbuffer2
*args
)
55 struct eb_vmas
*eb
= NULL
;
57 if (args
->flags
& I915_EXEC_HANDLE_LUT
) {
58 unsigned size
= args
->buffer_count
;
59 size
*= sizeof(struct i915_vma
*);
60 size
+= sizeof(struct eb_vmas
);
61 eb
= kmalloc(size
, GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
65 unsigned size
= args
->buffer_count
;
66 unsigned count
= PAGE_SIZE
/ sizeof(struct hlist_head
) / 2;
67 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE
/ sizeof(struct hlist_head
));
68 while (count
> 2*size
)
70 eb
= kzalloc(count
*sizeof(struct hlist_head
) +
71 sizeof(struct eb_vmas
),
78 eb
->and = -args
->buffer_count
;
80 INIT_LIST_HEAD(&eb
->vmas
);
85 eb_reset(struct eb_vmas
*eb
)
88 memset(eb
->buckets
, 0, (eb
->and+1)*sizeof(struct hlist_head
));
92 eb_lookup_vmas(struct eb_vmas
*eb
,
93 struct drm_i915_gem_exec_object2
*exec
,
94 const struct drm_i915_gem_execbuffer2
*args
,
95 struct i915_address_space
*vm
,
96 struct drm_file
*file
)
98 struct drm_i915_private
*dev_priv
= vm
->dev
->dev_private
;
99 struct drm_i915_gem_object
*obj
;
100 struct list_head objects
;
103 INIT_LIST_HEAD(&objects
);
104 spin_lock(&file
->table_lock
);
105 /* Grab a reference to the object and release the lock so we can lookup
106 * or create the VMA without using GFP_ATOMIC */
107 for (i
= 0; i
< args
->buffer_count
; i
++) {
108 obj
= to_intel_bo(idr_find(&file
->object_idr
, exec
[i
].handle
));
110 spin_unlock(&file
->table_lock
);
111 DRM_DEBUG("Invalid object handle %d at index %d\n",
117 if (!list_empty(&obj
->obj_exec_link
)) {
118 spin_unlock(&file
->table_lock
);
119 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
120 obj
, exec
[i
].handle
, i
);
125 drm_gem_object_reference(&obj
->base
);
126 list_add_tail(&obj
->obj_exec_link
, &objects
);
128 spin_unlock(&file
->table_lock
);
131 while (!list_empty(&objects
)) {
132 struct i915_vma
*vma
;
133 struct i915_address_space
*bind_vm
= vm
;
135 if (exec
[i
].flags
& EXEC_OBJECT_NEEDS_GTT
&&
136 USES_FULL_PPGTT(vm
->dev
)) {
141 /* If we have secure dispatch, or the userspace assures us that
142 * they know what they're doing, use the GGTT VM.
144 if (((args
->flags
& I915_EXEC_SECURE
) &&
145 (i
== (args
->buffer_count
- 1))))
146 bind_vm
= &dev_priv
->gtt
.base
;
148 obj
= list_first_entry(&objects
,
149 struct drm_i915_gem_object
,
153 * NOTE: We can leak any vmas created here when something fails
154 * later on. But that's no issue since vma_unbind can deal with
155 * vmas which are not actually bound. And since only
156 * lookup_or_create exists as an interface to get at the vma
157 * from the (obj, vm) we don't run the risk of creating
158 * duplicated vmas for the same vm.
160 vma
= i915_gem_obj_lookup_or_create_vma(obj
, bind_vm
);
162 DRM_DEBUG("Failed to lookup VMA\n");
167 /* Transfer ownership from the objects list to the vmas list. */
168 list_add_tail(&vma
->exec_list
, &eb
->vmas
);
169 list_del_init(&obj
->obj_exec_link
);
171 vma
->exec_entry
= &exec
[i
];
175 uint32_t handle
= args
->flags
& I915_EXEC_HANDLE_LUT
? i
: exec
[i
].handle
;
176 vma
->exec_handle
= handle
;
177 hlist_add_head(&vma
->exec_node
,
178 &eb
->buckets
[handle
& eb
->and]);
187 while (!list_empty(&objects
)) {
188 obj
= list_first_entry(&objects
,
189 struct drm_i915_gem_object
,
191 list_del_init(&obj
->obj_exec_link
);
192 drm_gem_object_unreference(&obj
->base
);
195 * Objects already transfered to the vmas list will be unreferenced by
202 static struct i915_vma
*eb_get_vma(struct eb_vmas
*eb
, unsigned long handle
)
205 if (handle
>= -eb
->and)
207 return eb
->lut
[handle
];
209 struct hlist_head
*head
;
210 struct hlist_node
*node
;
212 head
= &eb
->buckets
[handle
& eb
->and];
213 hlist_for_each(node
, head
) {
214 struct i915_vma
*vma
;
216 vma
= hlist_entry(node
, struct i915_vma
, exec_node
);
217 if (vma
->exec_handle
== handle
)
225 i915_gem_execbuffer_unreserve_vma(struct i915_vma
*vma
)
227 struct drm_i915_gem_exec_object2
*entry
;
228 struct drm_i915_gem_object
*obj
= vma
->obj
;
230 if (!drm_mm_node_allocated(&vma
->node
))
233 entry
= vma
->exec_entry
;
235 if (entry
->flags
& __EXEC_OBJECT_HAS_FENCE
)
236 i915_gem_object_unpin_fence(obj
);
238 if (entry
->flags
& __EXEC_OBJECT_HAS_PIN
)
241 entry
->flags
&= ~(__EXEC_OBJECT_HAS_FENCE
| __EXEC_OBJECT_HAS_PIN
);
244 static void eb_destroy(struct eb_vmas
*eb
)
246 while (!list_empty(&eb
->vmas
)) {
247 struct i915_vma
*vma
;
249 vma
= list_first_entry(&eb
->vmas
,
252 list_del_init(&vma
->exec_list
);
253 i915_gem_execbuffer_unreserve_vma(vma
);
254 drm_gem_object_unreference(&vma
->obj
->base
);
259 static inline int use_cpu_reloc(struct drm_i915_gem_object
*obj
)
261 return (HAS_LLC(obj
->base
.dev
) ||
262 obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
||
263 !obj
->map_and_fenceable
||
264 obj
->cache_level
!= I915_CACHE_NONE
);
268 relocate_entry_cpu(struct drm_i915_gem_object
*obj
,
269 struct drm_i915_gem_relocation_entry
*reloc
,
270 uint64_t target_offset
)
272 struct drm_device
*dev
= obj
->base
.dev
;
273 uint32_t page_offset
= offset_in_page(reloc
->offset
);
274 uint64_t delta
= reloc
->delta
+ target_offset
;
278 ret
= i915_gem_object_set_to_cpu_domain(obj
, true);
282 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
283 reloc
->offset
>> PAGE_SHIFT
));
284 *(uint32_t *)(vaddr
+ page_offset
) = lower_32_bits(delta
);
286 if (INTEL_INFO(dev
)->gen
>= 8) {
287 page_offset
= offset_in_page(page_offset
+ sizeof(uint32_t));
289 if (page_offset
== 0) {
290 kunmap_atomic(vaddr
);
291 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
292 (reloc
->offset
+ sizeof(uint32_t)) >> PAGE_SHIFT
));
295 *(uint32_t *)(vaddr
+ page_offset
) = upper_32_bits(delta
);
298 kunmap_atomic(vaddr
);
304 relocate_entry_gtt(struct drm_i915_gem_object
*obj
,
305 struct drm_i915_gem_relocation_entry
*reloc
,
306 uint64_t target_offset
)
308 struct drm_device
*dev
= obj
->base
.dev
;
309 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
310 uint64_t delta
= reloc
->delta
+ target_offset
;
311 uint32_t __iomem
*reloc_entry
;
312 void __iomem
*reloc_page
;
315 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
319 ret
= i915_gem_object_put_fence(obj
);
323 /* Map the page containing the relocation we're going to perform. */
324 reloc
->offset
+= i915_gem_obj_ggtt_offset(obj
);
325 reloc_page
= io_mapping_map_atomic_wc(dev_priv
->gtt
.mappable
,
326 reloc
->offset
& PAGE_MASK
);
327 reloc_entry
= (uint32_t __iomem
*)
328 (reloc_page
+ offset_in_page(reloc
->offset
));
329 iowrite32(lower_32_bits(delta
), reloc_entry
);
331 if (INTEL_INFO(dev
)->gen
>= 8) {
334 if (offset_in_page(reloc
->offset
+ sizeof(uint32_t)) == 0) {
335 io_mapping_unmap_atomic(reloc_page
);
336 reloc_page
= io_mapping_map_atomic_wc(
337 dev_priv
->gtt
.mappable
,
338 reloc
->offset
+ sizeof(uint32_t));
339 reloc_entry
= reloc_page
;
342 iowrite32(upper_32_bits(delta
), reloc_entry
);
345 io_mapping_unmap_atomic(reloc_page
);
351 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object
*obj
,
353 struct drm_i915_gem_relocation_entry
*reloc
)
355 struct drm_device
*dev
= obj
->base
.dev
;
356 struct drm_gem_object
*target_obj
;
357 struct drm_i915_gem_object
*target_i915_obj
;
358 struct i915_vma
*target_vma
;
359 uint64_t target_offset
;
362 /* we've already hold a reference to all valid objects */
363 target_vma
= eb_get_vma(eb
, reloc
->target_handle
);
364 if (unlikely(target_vma
== NULL
))
366 target_i915_obj
= target_vma
->obj
;
367 target_obj
= &target_vma
->obj
->base
;
369 target_offset
= target_vma
->node
.start
;
371 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
372 * pipe_control writes because the gpu doesn't properly redirect them
373 * through the ppgtt for non_secure batchbuffers. */
374 if (unlikely(IS_GEN6(dev
) &&
375 reloc
->write_domain
== I915_GEM_DOMAIN_INSTRUCTION
&&
376 !target_i915_obj
->has_global_gtt_mapping
)) {
377 struct i915_vma
*vma
=
378 list_first_entry(&target_i915_obj
->vma_list
,
379 typeof(*vma
), vma_link
);
380 vma
->bind_vma(vma
, target_i915_obj
->cache_level
, GLOBAL_BIND
);
383 /* Validate that the target is in a valid r/w GPU domain */
384 if (unlikely(reloc
->write_domain
& (reloc
->write_domain
- 1))) {
385 DRM_DEBUG("reloc with multiple write domains: "
386 "obj %p target %d offset %d "
387 "read %08x write %08x",
388 obj
, reloc
->target_handle
,
391 reloc
->write_domain
);
394 if (unlikely((reloc
->write_domain
| reloc
->read_domains
)
395 & ~I915_GEM_GPU_DOMAINS
)) {
396 DRM_DEBUG("reloc with read/write non-GPU domains: "
397 "obj %p target %d offset %d "
398 "read %08x write %08x",
399 obj
, reloc
->target_handle
,
402 reloc
->write_domain
);
406 target_obj
->pending_read_domains
|= reloc
->read_domains
;
407 target_obj
->pending_write_domain
|= reloc
->write_domain
;
409 /* If the relocation already has the right value in it, no
410 * more work needs to be done.
412 if (target_offset
== reloc
->presumed_offset
)
415 /* Check that the relocation address is valid... */
416 if (unlikely(reloc
->offset
>
417 obj
->base
.size
- (INTEL_INFO(dev
)->gen
>= 8 ? 8 : 4))) {
418 DRM_DEBUG("Relocation beyond object bounds: "
419 "obj %p target %d offset %d size %d.\n",
420 obj
, reloc
->target_handle
,
422 (int) obj
->base
.size
);
425 if (unlikely(reloc
->offset
& 3)) {
426 DRM_DEBUG("Relocation not 4-byte aligned: "
427 "obj %p target %d offset %d.\n",
428 obj
, reloc
->target_handle
,
429 (int) reloc
->offset
);
433 /* We can't wait for rendering with pagefaults disabled */
434 if (obj
->active
&& in_atomic())
437 if (use_cpu_reloc(obj
))
438 ret
= relocate_entry_cpu(obj
, reloc
, target_offset
);
440 ret
= relocate_entry_gtt(obj
, reloc
, target_offset
);
445 /* and update the user's relocation entry */
446 reloc
->presumed_offset
= target_offset
;
452 i915_gem_execbuffer_relocate_vma(struct i915_vma
*vma
,
455 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
456 struct drm_i915_gem_relocation_entry stack_reloc
[N_RELOC(512)];
457 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
458 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
461 user_relocs
= to_user_ptr(entry
->relocs_ptr
);
463 remain
= entry
->relocation_count
;
465 struct drm_i915_gem_relocation_entry
*r
= stack_reloc
;
467 if (count
> ARRAY_SIZE(stack_reloc
))
468 count
= ARRAY_SIZE(stack_reloc
);
471 if (__copy_from_user_inatomic(r
, user_relocs
, count
*sizeof(r
[0])))
475 u64 offset
= r
->presumed_offset
;
477 ret
= i915_gem_execbuffer_relocate_entry(vma
->obj
, eb
, r
);
481 if (r
->presumed_offset
!= offset
&&
482 __copy_to_user_inatomic(&user_relocs
->presumed_offset
,
484 sizeof(r
->presumed_offset
))) {
498 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma
*vma
,
500 struct drm_i915_gem_relocation_entry
*relocs
)
502 const struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
505 for (i
= 0; i
< entry
->relocation_count
; i
++) {
506 ret
= i915_gem_execbuffer_relocate_entry(vma
->obj
, eb
, &relocs
[i
]);
515 i915_gem_execbuffer_relocate(struct eb_vmas
*eb
)
517 struct i915_vma
*vma
;
520 /* This is the fast path and we cannot handle a pagefault whilst
521 * holding the struct mutex lest the user pass in the relocations
522 * contained within a mmaped bo. For in such a case we, the page
523 * fault handler would call i915_gem_fault() and we would try to
524 * acquire the struct mutex again. Obviously this is bad and so
525 * lockdep complains vehemently.
528 list_for_each_entry(vma
, &eb
->vmas
, exec_list
) {
529 ret
= i915_gem_execbuffer_relocate_vma(vma
, eb
);
539 i915_gem_execbuffer_reserve_vma(struct i915_vma
*vma
,
540 struct intel_engine_cs
*ring
,
543 struct drm_i915_gem_object
*obj
= vma
->obj
;
544 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
545 bool has_fenced_gpu_access
= INTEL_INFO(ring
->dev
)->gen
< 4;
550 if (entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
)
551 flags
|= PIN_MAPPABLE
;
552 if (entry
->flags
& EXEC_OBJECT_NEEDS_GTT
)
554 if (entry
->flags
& __EXEC_OBJECT_NEEDS_BIAS
)
555 flags
|= BATCH_OFFSET_BIAS
| PIN_OFFSET_BIAS
;
557 ret
= i915_gem_object_pin(obj
, vma
->vm
, entry
->alignment
, flags
);
561 entry
->flags
|= __EXEC_OBJECT_HAS_PIN
;
563 if (has_fenced_gpu_access
) {
564 if (entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
) {
565 ret
= i915_gem_object_get_fence(obj
);
569 if (i915_gem_object_pin_fence(obj
))
570 entry
->flags
|= __EXEC_OBJECT_HAS_FENCE
;
572 obj
->pending_fenced_gpu_access
= true;
576 if (entry
->offset
!= vma
->node
.start
) {
577 entry
->offset
= vma
->node
.start
;
581 if (entry
->flags
& EXEC_OBJECT_WRITE
) {
582 obj
->base
.pending_read_domains
= I915_GEM_DOMAIN_RENDER
;
583 obj
->base
.pending_write_domain
= I915_GEM_DOMAIN_RENDER
;
590 need_reloc_mappable(struct i915_vma
*vma
)
592 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
594 if (entry
->relocation_count
== 0)
597 if (!i915_is_ggtt(vma
->vm
))
600 /* See also use_cpu_reloc() */
601 if (HAS_LLC(vma
->obj
->base
.dev
))
604 if (vma
->obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
)
611 eb_vma_misplaced(struct i915_vma
*vma
)
613 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
614 struct drm_i915_gem_object
*obj
= vma
->obj
;
616 WARN_ON(entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
&&
617 !i915_is_ggtt(vma
->vm
));
619 if (entry
->alignment
&&
620 vma
->node
.start
& (entry
->alignment
- 1))
623 if (entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
&& !obj
->map_and_fenceable
)
626 if (entry
->flags
& __EXEC_OBJECT_NEEDS_BIAS
&&
627 vma
->node
.start
< BATCH_OFFSET_BIAS
)
634 i915_gem_execbuffer_reserve(struct intel_engine_cs
*ring
,
635 struct list_head
*vmas
,
638 struct drm_i915_gem_object
*obj
;
639 struct i915_vma
*vma
;
640 struct i915_address_space
*vm
;
641 struct list_head ordered_vmas
;
642 bool has_fenced_gpu_access
= INTEL_INFO(ring
->dev
)->gen
< 4;
645 if (list_empty(vmas
))
648 i915_gem_retire_requests_ring(ring
);
650 vm
= list_first_entry(vmas
, struct i915_vma
, exec_list
)->vm
;
652 INIT_LIST_HEAD(&ordered_vmas
);
653 while (!list_empty(vmas
)) {
654 struct drm_i915_gem_exec_object2
*entry
;
655 bool need_fence
, need_mappable
;
657 vma
= list_first_entry(vmas
, struct i915_vma
, exec_list
);
659 entry
= vma
->exec_entry
;
662 has_fenced_gpu_access
&&
663 entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
&&
664 obj
->tiling_mode
!= I915_TILING_NONE
;
665 need_mappable
= need_fence
|| need_reloc_mappable(vma
);
668 entry
->flags
|= __EXEC_OBJECT_NEEDS_MAP
;
669 list_move(&vma
->exec_list
, &ordered_vmas
);
671 list_move_tail(&vma
->exec_list
, &ordered_vmas
);
673 obj
->base
.pending_read_domains
= I915_GEM_GPU_DOMAINS
& ~I915_GEM_DOMAIN_COMMAND
;
674 obj
->base
.pending_write_domain
= 0;
675 obj
->pending_fenced_gpu_access
= false;
677 list_splice(&ordered_vmas
, vmas
);
679 /* Attempt to pin all of the buffers into the GTT.
680 * This is done in 3 phases:
682 * 1a. Unbind all objects that do not match the GTT constraints for
683 * the execbuffer (fenceable, mappable, alignment etc).
684 * 1b. Increment pin count for already bound objects.
685 * 2. Bind new objects.
686 * 3. Decrement pin count.
688 * This avoid unnecessary unbinding of later objects in order to make
689 * room for the earlier objects *unless* we need to defragment.
695 /* Unbind any ill-fitting objects or pin. */
696 list_for_each_entry(vma
, vmas
, exec_list
) {
697 if (!drm_mm_node_allocated(&vma
->node
))
700 if (eb_vma_misplaced(vma
))
701 ret
= i915_vma_unbind(vma
);
703 ret
= i915_gem_execbuffer_reserve_vma(vma
, ring
, need_relocs
);
708 /* Bind fresh objects */
709 list_for_each_entry(vma
, vmas
, exec_list
) {
710 if (drm_mm_node_allocated(&vma
->node
))
713 ret
= i915_gem_execbuffer_reserve_vma(vma
, ring
, need_relocs
);
719 if (ret
!= -ENOSPC
|| retry
++)
722 /* Decrement pin count for bound objects */
723 list_for_each_entry(vma
, vmas
, exec_list
)
724 i915_gem_execbuffer_unreserve_vma(vma
);
726 ret
= i915_gem_evict_vm(vm
, true);
733 i915_gem_execbuffer_relocate_slow(struct drm_device
*dev
,
734 struct drm_i915_gem_execbuffer2
*args
,
735 struct drm_file
*file
,
736 struct intel_engine_cs
*ring
,
738 struct drm_i915_gem_exec_object2
*exec
)
740 struct drm_i915_gem_relocation_entry
*reloc
;
741 struct i915_address_space
*vm
;
742 struct i915_vma
*vma
;
746 unsigned count
= args
->buffer_count
;
748 if (WARN_ON(list_empty(&eb
->vmas
)))
751 vm
= list_first_entry(&eb
->vmas
, struct i915_vma
, exec_list
)->vm
;
753 /* We may process another execbuffer during the unlock... */
754 while (!list_empty(&eb
->vmas
)) {
755 vma
= list_first_entry(&eb
->vmas
, struct i915_vma
, exec_list
);
756 list_del_init(&vma
->exec_list
);
757 i915_gem_execbuffer_unreserve_vma(vma
);
758 drm_gem_object_unreference(&vma
->obj
->base
);
761 mutex_unlock(&dev
->struct_mutex
);
764 for (i
= 0; i
< count
; i
++)
765 total
+= exec
[i
].relocation_count
;
767 reloc_offset
= drm_malloc_ab(count
, sizeof(*reloc_offset
));
768 reloc
= drm_malloc_ab(total
, sizeof(*reloc
));
769 if (reloc
== NULL
|| reloc_offset
== NULL
) {
770 drm_free_large(reloc
);
771 drm_free_large(reloc_offset
);
772 mutex_lock(&dev
->struct_mutex
);
777 for (i
= 0; i
< count
; i
++) {
778 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
779 u64 invalid_offset
= (u64
)-1;
782 user_relocs
= to_user_ptr(exec
[i
].relocs_ptr
);
784 if (copy_from_user(reloc
+total
, user_relocs
,
785 exec
[i
].relocation_count
* sizeof(*reloc
))) {
787 mutex_lock(&dev
->struct_mutex
);
791 /* As we do not update the known relocation offsets after
792 * relocating (due to the complexities in lock handling),
793 * we need to mark them as invalid now so that we force the
794 * relocation processing next time. Just in case the target
795 * object is evicted and then rebound into its old
796 * presumed_offset before the next execbuffer - if that
797 * happened we would make the mistake of assuming that the
798 * relocations were valid.
800 for (j
= 0; j
< exec
[i
].relocation_count
; j
++) {
801 if (__copy_to_user(&user_relocs
[j
].presumed_offset
,
803 sizeof(invalid_offset
))) {
805 mutex_lock(&dev
->struct_mutex
);
810 reloc_offset
[i
] = total
;
811 total
+= exec
[i
].relocation_count
;
814 ret
= i915_mutex_lock_interruptible(dev
);
816 mutex_lock(&dev
->struct_mutex
);
820 /* reacquire the objects */
822 ret
= eb_lookup_vmas(eb
, exec
, args
, vm
, file
);
826 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
827 ret
= i915_gem_execbuffer_reserve(ring
, &eb
->vmas
, &need_relocs
);
831 list_for_each_entry(vma
, &eb
->vmas
, exec_list
) {
832 int offset
= vma
->exec_entry
- exec
;
833 ret
= i915_gem_execbuffer_relocate_vma_slow(vma
, eb
,
834 reloc
+ reloc_offset
[offset
]);
839 /* Leave the user relocations as are, this is the painfully slow path,
840 * and we want to avoid the complication of dropping the lock whilst
841 * having buffers reserved in the aperture and so causing spurious
842 * ENOSPC for random operations.
846 drm_free_large(reloc
);
847 drm_free_large(reloc_offset
);
852 i915_gem_execbuffer_move_to_gpu(struct intel_engine_cs
*ring
,
853 struct list_head
*vmas
)
855 struct i915_vma
*vma
;
856 uint32_t flush_domains
= 0;
857 bool flush_chipset
= false;
860 list_for_each_entry(vma
, vmas
, exec_list
) {
861 struct drm_i915_gem_object
*obj
= vma
->obj
;
862 ret
= i915_gem_object_sync(obj
, ring
);
866 if (obj
->base
.write_domain
& I915_GEM_DOMAIN_CPU
)
867 flush_chipset
|= i915_gem_clflush_object(obj
, false);
869 flush_domains
|= obj
->base
.write_domain
;
873 i915_gem_chipset_flush(ring
->dev
);
875 if (flush_domains
& I915_GEM_DOMAIN_GTT
)
878 /* Unconditionally invalidate gpu caches and ensure that we do flush
879 * any residual writes from the previous batch.
881 return intel_ring_invalidate_all_caches(ring
);
885 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2
*exec
)
887 if (exec
->flags
& __I915_EXEC_UNKNOWN_FLAGS
)
890 return ((exec
->batch_start_offset
| exec
->batch_len
) & 0x7) == 0;
894 validate_exec_list(struct drm_i915_gem_exec_object2
*exec
,
898 unsigned relocs_total
= 0;
899 unsigned relocs_max
= UINT_MAX
/ sizeof(struct drm_i915_gem_relocation_entry
);
901 for (i
= 0; i
< count
; i
++) {
902 char __user
*ptr
= to_user_ptr(exec
[i
].relocs_ptr
);
903 int length
; /* limited by fault_in_pages_readable() */
905 if (exec
[i
].flags
& __EXEC_OBJECT_UNKNOWN_FLAGS
)
908 /* First check for malicious input causing overflow in
909 * the worst case where we need to allocate the entire
910 * relocation tree as a single array.
912 if (exec
[i
].relocation_count
> relocs_max
- relocs_total
)
914 relocs_total
+= exec
[i
].relocation_count
;
916 length
= exec
[i
].relocation_count
*
917 sizeof(struct drm_i915_gem_relocation_entry
);
919 * We must check that the entire relocation array is safe
920 * to read, but since we may need to update the presumed
921 * offsets during execution, check for full write access.
923 if (!access_ok(VERIFY_WRITE
, ptr
, length
))
926 if (likely(!i915
.prefault_disable
)) {
927 if (fault_in_multipages_readable(ptr
, length
))
935 static struct intel_context
*
936 i915_gem_validate_context(struct drm_device
*dev
, struct drm_file
*file
,
937 struct intel_engine_cs
*ring
, const u32 ctx_id
)
939 struct intel_context
*ctx
= NULL
;
940 struct i915_ctx_hang_stats
*hs
;
942 if (ring
->id
!= RCS
&& ctx_id
!= DEFAULT_CONTEXT_HANDLE
)
943 return ERR_PTR(-EINVAL
);
945 ctx
= i915_gem_context_get(file
->driver_priv
, ctx_id
);
949 hs
= &ctx
->hang_stats
;
951 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id
);
952 return ERR_PTR(-EIO
);
959 i915_gem_execbuffer_move_to_active(struct list_head
*vmas
,
960 struct intel_engine_cs
*ring
)
962 struct i915_vma
*vma
;
964 list_for_each_entry(vma
, vmas
, exec_list
) {
965 struct drm_i915_gem_object
*obj
= vma
->obj
;
966 u32 old_read
= obj
->base
.read_domains
;
967 u32 old_write
= obj
->base
.write_domain
;
969 obj
->base
.write_domain
= obj
->base
.pending_write_domain
;
970 if (obj
->base
.write_domain
== 0)
971 obj
->base
.pending_read_domains
|= obj
->base
.read_domains
;
972 obj
->base
.read_domains
= obj
->base
.pending_read_domains
;
973 obj
->fenced_gpu_access
= obj
->pending_fenced_gpu_access
;
975 i915_vma_move_to_active(vma
, ring
);
976 if (obj
->base
.write_domain
) {
978 obj
->last_write_seqno
= intel_ring_get_seqno(ring
);
980 intel_fb_obj_invalidate(obj
, ring
);
982 /* update for the implicit flush after a batch */
983 obj
->base
.write_domain
&= ~I915_GEM_GPU_DOMAINS
;
986 trace_i915_gem_object_change_domain(obj
, old_read
, old_write
);
991 i915_gem_execbuffer_retire_commands(struct drm_device
*dev
,
992 struct drm_file
*file
,
993 struct intel_engine_cs
*ring
,
994 struct drm_i915_gem_object
*obj
)
996 /* Unconditionally force add_request to emit a full flush. */
997 ring
->gpu_caches_dirty
= true;
999 /* Add a breadcrumb for the completion of the batch buffer */
1000 (void)__i915_add_request(ring
, file
, obj
, NULL
);
1004 i915_reset_gen7_sol_offsets(struct drm_device
*dev
,
1005 struct intel_engine_cs
*ring
)
1007 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1010 if (!IS_GEN7(dev
) || ring
!= &dev_priv
->ring
[RCS
]) {
1011 DRM_DEBUG("sol reset is gen7/rcs only\n");
1015 ret
= intel_ring_begin(ring
, 4 * 3);
1019 for (i
= 0; i
< 4; i
++) {
1020 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
1021 intel_ring_emit(ring
, GEN7_SO_WRITE_OFFSET(i
));
1022 intel_ring_emit(ring
, 0);
1025 intel_ring_advance(ring
);
1031 legacy_ringbuffer_submission(struct drm_device
*dev
, struct drm_file
*file
,
1032 struct intel_engine_cs
*ring
,
1033 struct intel_context
*ctx
,
1034 struct drm_i915_gem_execbuffer2
*args
,
1035 struct list_head
*vmas
,
1036 struct drm_i915_gem_object
*batch_obj
,
1037 u64 exec_start
, u32 flags
)
1039 struct drm_clip_rect
*cliprects
= NULL
;
1040 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1046 if (args
->num_cliprects
!= 0) {
1047 if (ring
!= &dev_priv
->ring
[RCS
]) {
1048 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1052 if (INTEL_INFO(dev
)->gen
>= 5) {
1053 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1057 if (args
->num_cliprects
> UINT_MAX
/ sizeof(*cliprects
)) {
1058 DRM_DEBUG("execbuf with %u cliprects\n",
1059 args
->num_cliprects
);
1063 cliprects
= kcalloc(args
->num_cliprects
,
1066 if (cliprects
== NULL
) {
1071 if (copy_from_user(cliprects
,
1072 to_user_ptr(args
->cliprects_ptr
),
1073 sizeof(*cliprects
)*args
->num_cliprects
)) {
1078 if (args
->DR4
== 0xffffffff) {
1079 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1083 if (args
->DR1
|| args
->DR4
|| args
->cliprects_ptr
) {
1084 DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
1089 ret
= i915_gem_execbuffer_move_to_gpu(ring
, vmas
);
1093 ret
= i915_switch_context(ring
, ctx
);
1097 instp_mode
= args
->flags
& I915_EXEC_CONSTANTS_MASK
;
1098 instp_mask
= I915_EXEC_CONSTANTS_MASK
;
1099 switch (instp_mode
) {
1100 case I915_EXEC_CONSTANTS_REL_GENERAL
:
1101 case I915_EXEC_CONSTANTS_ABSOLUTE
:
1102 case I915_EXEC_CONSTANTS_REL_SURFACE
:
1103 if (instp_mode
!= 0 && ring
!= &dev_priv
->ring
[RCS
]) {
1104 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1109 if (instp_mode
!= dev_priv
->relative_constants_mode
) {
1110 if (INTEL_INFO(dev
)->gen
< 4) {
1111 DRM_DEBUG("no rel constants on pre-gen4\n");
1116 if (INTEL_INFO(dev
)->gen
> 5 &&
1117 instp_mode
== I915_EXEC_CONSTANTS_REL_SURFACE
) {
1118 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1123 /* The HW changed the meaning on this bit on gen6 */
1124 if (INTEL_INFO(dev
)->gen
>= 6)
1125 instp_mask
&= ~I915_EXEC_CONSTANTS_REL_SURFACE
;
1129 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode
);
1134 if (ring
== &dev_priv
->ring
[RCS
] &&
1135 instp_mode
!= dev_priv
->relative_constants_mode
) {
1136 ret
= intel_ring_begin(ring
, 4);
1140 intel_ring_emit(ring
, MI_NOOP
);
1141 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
1142 intel_ring_emit(ring
, INSTPM
);
1143 intel_ring_emit(ring
, instp_mask
<< 16 | instp_mode
);
1144 intel_ring_advance(ring
);
1146 dev_priv
->relative_constants_mode
= instp_mode
;
1149 if (args
->flags
& I915_EXEC_GEN7_SOL_RESET
) {
1150 ret
= i915_reset_gen7_sol_offsets(dev
, ring
);
1155 exec_len
= args
->batch_len
;
1157 for (i
= 0; i
< args
->num_cliprects
; i
++) {
1158 ret
= i915_emit_box(dev
, &cliprects
[i
],
1159 args
->DR1
, args
->DR4
);
1163 ret
= ring
->dispatch_execbuffer(ring
,
1164 exec_start
, exec_len
,
1170 ret
= ring
->dispatch_execbuffer(ring
,
1171 exec_start
, exec_len
,
1177 trace_i915_gem_ring_dispatch(ring
, intel_ring_get_seqno(ring
), flags
);
1179 i915_gem_execbuffer_move_to_active(vmas
, ring
);
1180 i915_gem_execbuffer_retire_commands(dev
, file
, ring
, batch_obj
);
1188 * Find one BSD ring to dispatch the corresponding BSD command.
1189 * The Ring ID is returned.
1191 static int gen8_dispatch_bsd_ring(struct drm_device
*dev
,
1192 struct drm_file
*file
)
1194 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1195 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
1197 /* Check whether the file_priv is using one ring */
1198 if (file_priv
->bsd_ring
)
1199 return file_priv
->bsd_ring
->id
;
1201 /* If no, use the ping-pong mechanism to select one ring */
1204 mutex_lock(&dev
->struct_mutex
);
1205 if (dev_priv
->mm
.bsd_ring_dispatch_index
== 0) {
1207 dev_priv
->mm
.bsd_ring_dispatch_index
= 1;
1210 dev_priv
->mm
.bsd_ring_dispatch_index
= 0;
1212 file_priv
->bsd_ring
= &dev_priv
->ring
[ring_id
];
1213 mutex_unlock(&dev
->struct_mutex
);
1218 static struct drm_i915_gem_object
*
1219 eb_get_batch(struct eb_vmas
*eb
)
1221 struct i915_vma
*vma
= list_entry(eb
->vmas
.prev
, typeof(*vma
), exec_list
);
1224 * SNA is doing fancy tricks with compressing batch buffers, which leads
1225 * to negative relocation deltas. Usually that works out ok since the
1226 * relocate address is still positive, except when the batch is placed
1227 * very low in the GTT. Ensure this doesn't happen.
1229 * Note that actual hangs have only been observed on gen7, but for
1230 * paranoia do it everywhere.
1232 vma
->exec_entry
->flags
|= __EXEC_OBJECT_NEEDS_BIAS
;
1238 i915_gem_do_execbuffer(struct drm_device
*dev
, void *data
,
1239 struct drm_file
*file
,
1240 struct drm_i915_gem_execbuffer2
*args
,
1241 struct drm_i915_gem_exec_object2
*exec
)
1243 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1245 struct drm_i915_gem_object
*batch_obj
;
1246 struct intel_engine_cs
*ring
;
1247 struct intel_context
*ctx
;
1248 struct i915_address_space
*vm
;
1249 const u32 ctx_id
= i915_execbuffer2_get_context_id(*args
);
1250 u64 exec_start
= args
->batch_start_offset
;
1255 if (!i915_gem_check_execbuffer(args
))
1258 ret
= validate_exec_list(exec
, args
->buffer_count
);
1263 if (args
->flags
& I915_EXEC_SECURE
) {
1264 if (!file
->is_master
|| !capable(CAP_SYS_ADMIN
))
1267 flags
|= I915_DISPATCH_SECURE
;
1269 if (args
->flags
& I915_EXEC_IS_PINNED
)
1270 flags
|= I915_DISPATCH_PINNED
;
1272 if ((args
->flags
& I915_EXEC_RING_MASK
) > LAST_USER_RING
) {
1273 DRM_DEBUG("execbuf with unknown ring: %d\n",
1274 (int)(args
->flags
& I915_EXEC_RING_MASK
));
1278 if ((args
->flags
& I915_EXEC_RING_MASK
) == I915_EXEC_DEFAULT
)
1279 ring
= &dev_priv
->ring
[RCS
];
1280 else if ((args
->flags
& I915_EXEC_RING_MASK
) == I915_EXEC_BSD
) {
1281 if (HAS_BSD2(dev
)) {
1283 ring_id
= gen8_dispatch_bsd_ring(dev
, file
);
1284 ring
= &dev_priv
->ring
[ring_id
];
1286 ring
= &dev_priv
->ring
[VCS
];
1288 ring
= &dev_priv
->ring
[(args
->flags
& I915_EXEC_RING_MASK
) - 1];
1290 if (!intel_ring_initialized(ring
)) {
1291 DRM_DEBUG("execbuf with invalid ring: %d\n",
1292 (int)(args
->flags
& I915_EXEC_RING_MASK
));
1296 if (args
->buffer_count
< 1) {
1297 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1301 intel_runtime_pm_get(dev_priv
);
1303 ret
= i915_mutex_lock_interruptible(dev
);
1307 if (dev_priv
->ums
.mm_suspended
) {
1308 mutex_unlock(&dev
->struct_mutex
);
1313 ctx
= i915_gem_validate_context(dev
, file
, ring
, ctx_id
);
1315 mutex_unlock(&dev
->struct_mutex
);
1320 i915_gem_context_reference(ctx
);
1323 if (!USES_FULL_PPGTT(dev
))
1324 vm
= &dev_priv
->gtt
.base
;
1326 eb
= eb_create(args
);
1328 i915_gem_context_unreference(ctx
);
1329 mutex_unlock(&dev
->struct_mutex
);
1334 /* Look up object handles */
1335 ret
= eb_lookup_vmas(eb
, exec
, args
, vm
, file
);
1339 /* take note of the batch buffer before we might reorder the lists */
1340 batch_obj
= eb_get_batch(eb
);
1342 /* Move the objects en-masse into the GTT, evicting if necessary. */
1343 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
1344 ret
= i915_gem_execbuffer_reserve(ring
, &eb
->vmas
, &need_relocs
);
1348 /* The objects are in their final locations, apply the relocations. */
1350 ret
= i915_gem_execbuffer_relocate(eb
);
1352 if (ret
== -EFAULT
) {
1353 ret
= i915_gem_execbuffer_relocate_slow(dev
, args
, file
, ring
,
1355 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
1361 /* Set the pending read domains for the batch buffer to COMMAND */
1362 if (batch_obj
->base
.pending_write_domain
) {
1363 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1367 batch_obj
->base
.pending_read_domains
|= I915_GEM_DOMAIN_COMMAND
;
1369 if (i915_needs_cmd_parser(ring
)) {
1370 ret
= i915_parse_cmds(ring
,
1372 args
->batch_start_offset
,
1378 * XXX: Actually do this when enabling batch copy...
1380 * Set the DISPATCH_SECURE bit to remove the NON_SECURE bit
1381 * from MI_BATCH_BUFFER_START commands issued in the
1382 * dispatch_execbuffer implementations. We specifically don't
1383 * want that set when the command parser is enabled.
1387 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1388 * batch" bit. Hence we need to pin secure batches into the global gtt.
1389 * hsw should have this fixed, but bdw mucks it up again. */
1390 if (flags
& I915_DISPATCH_SECURE
&&
1391 !batch_obj
->has_global_gtt_mapping
) {
1392 /* When we have multiple VMs, we'll need to make sure that we
1393 * allocate space first */
1394 struct i915_vma
*vma
= i915_gem_obj_to_ggtt(batch_obj
);
1396 vma
->bind_vma(vma
, batch_obj
->cache_level
, GLOBAL_BIND
);
1399 if (flags
& I915_DISPATCH_SECURE
)
1400 exec_start
+= i915_gem_obj_ggtt_offset(batch_obj
);
1402 exec_start
+= i915_gem_obj_offset(batch_obj
, vm
);
1404 ret
= legacy_ringbuffer_submission(dev
, file
, ring
, ctx
,
1405 args
, &eb
->vmas
, batch_obj
, exec_start
, flags
);
1410 /* the request owns the ref now */
1411 i915_gem_context_unreference(ctx
);
1414 mutex_unlock(&dev
->struct_mutex
);
1417 /* intel_gpu_busy should also get a ref, so it will free when the device
1418 * is really idle. */
1419 intel_runtime_pm_put(dev_priv
);
1424 * Legacy execbuffer just creates an exec2 list from the original exec object
1425 * list array and passes it to the real function.
1428 i915_gem_execbuffer(struct drm_device
*dev
, void *data
,
1429 struct drm_file
*file
)
1431 struct drm_i915_gem_execbuffer
*args
= data
;
1432 struct drm_i915_gem_execbuffer2 exec2
;
1433 struct drm_i915_gem_exec_object
*exec_list
= NULL
;
1434 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1437 if (args
->buffer_count
< 1) {
1438 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1442 /* Copy in the exec list from userland */
1443 exec_list
= drm_malloc_ab(sizeof(*exec_list
), args
->buffer_count
);
1444 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
), args
->buffer_count
);
1445 if (exec_list
== NULL
|| exec2_list
== NULL
) {
1446 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1447 args
->buffer_count
);
1448 drm_free_large(exec_list
);
1449 drm_free_large(exec2_list
);
1452 ret
= copy_from_user(exec_list
,
1453 to_user_ptr(args
->buffers_ptr
),
1454 sizeof(*exec_list
) * args
->buffer_count
);
1456 DRM_DEBUG("copy %d exec entries failed %d\n",
1457 args
->buffer_count
, ret
);
1458 drm_free_large(exec_list
);
1459 drm_free_large(exec2_list
);
1463 for (i
= 0; i
< args
->buffer_count
; i
++) {
1464 exec2_list
[i
].handle
= exec_list
[i
].handle
;
1465 exec2_list
[i
].relocation_count
= exec_list
[i
].relocation_count
;
1466 exec2_list
[i
].relocs_ptr
= exec_list
[i
].relocs_ptr
;
1467 exec2_list
[i
].alignment
= exec_list
[i
].alignment
;
1468 exec2_list
[i
].offset
= exec_list
[i
].offset
;
1469 if (INTEL_INFO(dev
)->gen
< 4)
1470 exec2_list
[i
].flags
= EXEC_OBJECT_NEEDS_FENCE
;
1472 exec2_list
[i
].flags
= 0;
1475 exec2
.buffers_ptr
= args
->buffers_ptr
;
1476 exec2
.buffer_count
= args
->buffer_count
;
1477 exec2
.batch_start_offset
= args
->batch_start_offset
;
1478 exec2
.batch_len
= args
->batch_len
;
1479 exec2
.DR1
= args
->DR1
;
1480 exec2
.DR4
= args
->DR4
;
1481 exec2
.num_cliprects
= args
->num_cliprects
;
1482 exec2
.cliprects_ptr
= args
->cliprects_ptr
;
1483 exec2
.flags
= I915_EXEC_RENDER
;
1484 i915_execbuffer2_set_context_id(exec2
, 0);
1486 ret
= i915_gem_do_execbuffer(dev
, data
, file
, &exec2
, exec2_list
);
1488 struct drm_i915_gem_exec_object __user
*user_exec_list
=
1489 to_user_ptr(args
->buffers_ptr
);
1491 /* Copy the new buffer offsets back to the user's exec list. */
1492 for (i
= 0; i
< args
->buffer_count
; i
++) {
1493 ret
= __copy_to_user(&user_exec_list
[i
].offset
,
1494 &exec2_list
[i
].offset
,
1495 sizeof(user_exec_list
[i
].offset
));
1498 DRM_DEBUG("failed to copy %d exec entries "
1499 "back to user (%d)\n",
1500 args
->buffer_count
, ret
);
1506 drm_free_large(exec_list
);
1507 drm_free_large(exec2_list
);
1512 i915_gem_execbuffer2(struct drm_device
*dev
, void *data
,
1513 struct drm_file
*file
)
1515 struct drm_i915_gem_execbuffer2
*args
= data
;
1516 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1519 if (args
->buffer_count
< 1 ||
1520 args
->buffer_count
> UINT_MAX
/ sizeof(*exec2_list
)) {
1521 DRM_DEBUG("execbuf2 with %d buffers\n", args
->buffer_count
);
1525 if (args
->rsvd2
!= 0) {
1526 DRM_DEBUG("dirty rvsd2 field\n");
1530 exec2_list
= kmalloc(sizeof(*exec2_list
)*args
->buffer_count
,
1531 GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
1532 if (exec2_list
== NULL
)
1533 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
),
1534 args
->buffer_count
);
1535 if (exec2_list
== NULL
) {
1536 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1537 args
->buffer_count
);
1540 ret
= copy_from_user(exec2_list
,
1541 to_user_ptr(args
->buffers_ptr
),
1542 sizeof(*exec2_list
) * args
->buffer_count
);
1544 DRM_DEBUG("copy %d exec entries failed %d\n",
1545 args
->buffer_count
, ret
);
1546 drm_free_large(exec2_list
);
1550 ret
= i915_gem_do_execbuffer(dev
, data
, file
, args
, exec2_list
);
1552 /* Copy the new buffer offsets back to the user's exec list. */
1553 struct drm_i915_gem_exec_object2 __user
*user_exec_list
=
1554 to_user_ptr(args
->buffers_ptr
);
1557 for (i
= 0; i
< args
->buffer_count
; i
++) {
1558 ret
= __copy_to_user(&user_exec_list
[i
].offset
,
1559 &exec2_list
[i
].offset
,
1560 sizeof(user_exec_list
[i
].offset
));
1563 DRM_DEBUG("failed to copy %d exec entries "
1565 args
->buffer_count
);
1571 drm_free_large(exec2_list
);