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)
40 struct list_head vmas
;
43 struct i915_vma
*lut
[0];
44 struct hlist_head buckets
[0];
48 static struct eb_vmas
*
49 eb_create(struct drm_i915_gem_execbuffer2
*args
, struct i915_address_space
*vm
)
51 struct eb_vmas
*eb
= NULL
;
53 if (args
->flags
& I915_EXEC_HANDLE_LUT
) {
54 unsigned size
= args
->buffer_count
;
55 size
*= sizeof(struct i915_vma
*);
56 size
+= sizeof(struct eb_vmas
);
57 eb
= kmalloc(size
, GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
61 unsigned size
= args
->buffer_count
;
62 unsigned count
= PAGE_SIZE
/ sizeof(struct hlist_head
) / 2;
63 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE
/ sizeof(struct hlist_head
));
64 while (count
> 2*size
)
66 eb
= kzalloc(count
*sizeof(struct hlist_head
) +
67 sizeof(struct eb_vmas
),
74 eb
->and = -args
->buffer_count
;
76 INIT_LIST_HEAD(&eb
->vmas
);
81 eb_reset(struct eb_vmas
*eb
)
84 memset(eb
->buckets
, 0, (eb
->and+1)*sizeof(struct hlist_head
));
88 eb_lookup_vmas(struct eb_vmas
*eb
,
89 struct drm_i915_gem_exec_object2
*exec
,
90 const struct drm_i915_gem_execbuffer2
*args
,
91 struct i915_address_space
*vm
,
92 struct drm_file
*file
)
94 struct drm_i915_gem_object
*obj
;
95 struct list_head objects
;
98 INIT_LIST_HEAD(&objects
);
99 spin_lock(&file
->table_lock
);
100 /* Grab a reference to the object and release the lock so we can lookup
101 * or create the VMA without using GFP_ATOMIC */
102 for (i
= 0; i
< args
->buffer_count
; i
++) {
103 obj
= to_intel_bo(idr_find(&file
->object_idr
, exec
[i
].handle
));
105 spin_unlock(&file
->table_lock
);
106 DRM_DEBUG("Invalid object handle %d at index %d\n",
112 if (!list_empty(&obj
->obj_exec_link
)) {
113 spin_unlock(&file
->table_lock
);
114 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
115 obj
, exec
[i
].handle
, i
);
120 drm_gem_object_reference(&obj
->base
);
121 list_add_tail(&obj
->obj_exec_link
, &objects
);
123 spin_unlock(&file
->table_lock
);
126 list_for_each_entry(obj
, &objects
, obj_exec_link
) {
127 struct i915_vma
*vma
;
130 * NOTE: We can leak any vmas created here when something fails
131 * later on. But that's no issue since vma_unbind can deal with
132 * vmas which are not actually bound. And since only
133 * lookup_or_create exists as an interface to get at the vma
134 * from the (obj, vm) we don't run the risk of creating
135 * duplicated vmas for the same vm.
137 vma
= i915_gem_obj_lookup_or_create_vma(obj
, vm
);
139 DRM_DEBUG("Failed to lookup VMA\n");
144 list_add_tail(&vma
->exec_list
, &eb
->vmas
);
146 vma
->exec_entry
= &exec
[i
];
150 uint32_t handle
= args
->flags
& I915_EXEC_HANDLE_LUT
? i
: exec
[i
].handle
;
151 vma
->exec_handle
= handle
;
152 hlist_add_head(&vma
->exec_node
,
153 &eb
->buckets
[handle
& eb
->and]);
160 while (!list_empty(&objects
)) {
161 obj
= list_first_entry(&objects
,
162 struct drm_i915_gem_object
,
164 list_del_init(&obj
->obj_exec_link
);
166 drm_gem_object_unreference(&obj
->base
);
171 static struct i915_vma
*eb_get_vma(struct eb_vmas
*eb
, unsigned long handle
)
174 if (handle
>= -eb
->and)
176 return eb
->lut
[handle
];
178 struct hlist_head
*head
;
179 struct hlist_node
*node
;
181 head
= &eb
->buckets
[handle
& eb
->and];
182 hlist_for_each(node
, head
) {
183 struct i915_vma
*vma
;
185 vma
= hlist_entry(node
, struct i915_vma
, exec_node
);
186 if (vma
->exec_handle
== handle
)
194 i915_gem_execbuffer_unreserve_vma(struct i915_vma
*vma
)
196 struct drm_i915_gem_exec_object2
*entry
;
197 struct drm_i915_gem_object
*obj
= vma
->obj
;
199 if (!drm_mm_node_allocated(&vma
->node
))
202 entry
= vma
->exec_entry
;
204 if (entry
->flags
& __EXEC_OBJECT_HAS_FENCE
)
205 i915_gem_object_unpin_fence(obj
);
207 if (entry
->flags
& __EXEC_OBJECT_HAS_PIN
)
208 i915_gem_object_unpin(obj
);
210 entry
->flags
&= ~(__EXEC_OBJECT_HAS_FENCE
| __EXEC_OBJECT_HAS_PIN
);
213 static void eb_destroy(struct eb_vmas
*eb
)
215 while (!list_empty(&eb
->vmas
)) {
216 struct i915_vma
*vma
;
218 vma
= list_first_entry(&eb
->vmas
,
221 list_del_init(&vma
->exec_list
);
222 i915_gem_execbuffer_unreserve_vma(vma
);
223 drm_gem_object_unreference(&vma
->obj
->base
);
228 static inline int use_cpu_reloc(struct drm_i915_gem_object
*obj
)
230 return (HAS_LLC(obj
->base
.dev
) ||
231 obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
||
232 !obj
->map_and_fenceable
||
233 obj
->cache_level
!= I915_CACHE_NONE
);
237 relocate_entry_cpu(struct drm_i915_gem_object
*obj
,
238 struct drm_i915_gem_relocation_entry
*reloc
)
240 struct drm_device
*dev
= obj
->base
.dev
;
241 uint32_t page_offset
= offset_in_page(reloc
->offset
);
245 ret
= i915_gem_object_set_to_cpu_domain(obj
, true);
249 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
250 reloc
->offset
>> PAGE_SHIFT
));
251 *(uint32_t *)(vaddr
+ page_offset
) = reloc
->delta
;
253 if (INTEL_INFO(dev
)->gen
>= 8) {
254 page_offset
= offset_in_page(page_offset
+ sizeof(uint32_t));
256 if (page_offset
== 0) {
257 kunmap_atomic(vaddr
);
258 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
259 (reloc
->offset
+ sizeof(uint32_t)) >> PAGE_SHIFT
));
262 *(uint32_t *)(vaddr
+ page_offset
) = 0;
265 kunmap_atomic(vaddr
);
271 relocate_entry_gtt(struct drm_i915_gem_object
*obj
,
272 struct drm_i915_gem_relocation_entry
*reloc
)
274 struct drm_device
*dev
= obj
->base
.dev
;
275 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
276 uint32_t __iomem
*reloc_entry
;
277 void __iomem
*reloc_page
;
280 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
284 ret
= i915_gem_object_put_fence(obj
);
288 /* Map the page containing the relocation we're going to perform. */
289 reloc
->offset
+= i915_gem_obj_ggtt_offset(obj
);
290 reloc_page
= io_mapping_map_atomic_wc(dev_priv
->gtt
.mappable
,
291 reloc
->offset
& PAGE_MASK
);
292 reloc_entry
= (uint32_t __iomem
*)
293 (reloc_page
+ offset_in_page(reloc
->offset
));
294 iowrite32(reloc
->delta
, reloc_entry
);
296 if (INTEL_INFO(dev
)->gen
>= 8) {
299 if (offset_in_page(reloc
->offset
+ sizeof(uint32_t)) == 0) {
300 io_mapping_unmap_atomic(reloc_page
);
301 reloc_page
= io_mapping_map_atomic_wc(
302 dev_priv
->gtt
.mappable
,
303 reloc
->offset
+ sizeof(uint32_t));
304 reloc_entry
= reloc_page
;
307 iowrite32(0, reloc_entry
);
310 io_mapping_unmap_atomic(reloc_page
);
316 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object
*obj
,
318 struct drm_i915_gem_relocation_entry
*reloc
,
319 struct i915_address_space
*vm
)
321 struct drm_device
*dev
= obj
->base
.dev
;
322 struct drm_gem_object
*target_obj
;
323 struct drm_i915_gem_object
*target_i915_obj
;
324 struct i915_vma
*target_vma
;
325 uint32_t target_offset
;
328 /* we've already hold a reference to all valid objects */
329 target_vma
= eb_get_vma(eb
, reloc
->target_handle
);
330 if (unlikely(target_vma
== NULL
))
332 target_i915_obj
= target_vma
->obj
;
333 target_obj
= &target_vma
->obj
->base
;
335 target_offset
= i915_gem_obj_ggtt_offset(target_i915_obj
);
337 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
338 * pipe_control writes because the gpu doesn't properly redirect them
339 * through the ppgtt for non_secure batchbuffers. */
340 if (unlikely(IS_GEN6(dev
) &&
341 reloc
->write_domain
== I915_GEM_DOMAIN_INSTRUCTION
&&
342 !target_i915_obj
->has_global_gtt_mapping
)) {
343 i915_gem_gtt_bind_object(target_i915_obj
,
344 target_i915_obj
->cache_level
);
347 /* Validate that the target is in a valid r/w GPU domain */
348 if (unlikely(reloc
->write_domain
& (reloc
->write_domain
- 1))) {
349 DRM_DEBUG("reloc with multiple write domains: "
350 "obj %p target %d offset %d "
351 "read %08x write %08x",
352 obj
, reloc
->target_handle
,
355 reloc
->write_domain
);
358 if (unlikely((reloc
->write_domain
| reloc
->read_domains
)
359 & ~I915_GEM_GPU_DOMAINS
)) {
360 DRM_DEBUG("reloc with read/write non-GPU domains: "
361 "obj %p target %d offset %d "
362 "read %08x write %08x",
363 obj
, reloc
->target_handle
,
366 reloc
->write_domain
);
370 target_obj
->pending_read_domains
|= reloc
->read_domains
;
371 target_obj
->pending_write_domain
|= reloc
->write_domain
;
373 /* If the relocation already has the right value in it, no
374 * more work needs to be done.
376 if (target_offset
== reloc
->presumed_offset
)
379 /* Check that the relocation address is valid... */
380 if (unlikely(reloc
->offset
>
381 obj
->base
.size
- (INTEL_INFO(dev
)->gen
>= 8 ? 8 : 4))) {
382 DRM_DEBUG("Relocation beyond object bounds: "
383 "obj %p target %d offset %d size %d.\n",
384 obj
, reloc
->target_handle
,
386 (int) obj
->base
.size
);
389 if (unlikely(reloc
->offset
& 3)) {
390 DRM_DEBUG("Relocation not 4-byte aligned: "
391 "obj %p target %d offset %d.\n",
392 obj
, reloc
->target_handle
,
393 (int) reloc
->offset
);
397 /* We can't wait for rendering with pagefaults disabled */
398 if (obj
->active
&& in_atomic())
401 reloc
->delta
+= target_offset
;
402 if (use_cpu_reloc(obj
))
403 ret
= relocate_entry_cpu(obj
, reloc
);
405 ret
= relocate_entry_gtt(obj
, reloc
);
410 /* and update the user's relocation entry */
411 reloc
->presumed_offset
= target_offset
;
417 i915_gem_execbuffer_relocate_vma(struct i915_vma
*vma
,
420 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
421 struct drm_i915_gem_relocation_entry stack_reloc
[N_RELOC(512)];
422 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
423 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
426 user_relocs
= to_user_ptr(entry
->relocs_ptr
);
428 remain
= entry
->relocation_count
;
430 struct drm_i915_gem_relocation_entry
*r
= stack_reloc
;
432 if (count
> ARRAY_SIZE(stack_reloc
))
433 count
= ARRAY_SIZE(stack_reloc
);
436 if (__copy_from_user_inatomic(r
, user_relocs
, count
*sizeof(r
[0])))
440 u64 offset
= r
->presumed_offset
;
442 ret
= i915_gem_execbuffer_relocate_entry(vma
->obj
, eb
, r
,
447 if (r
->presumed_offset
!= offset
&&
448 __copy_to_user_inatomic(&user_relocs
->presumed_offset
,
450 sizeof(r
->presumed_offset
))) {
464 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma
*vma
,
466 struct drm_i915_gem_relocation_entry
*relocs
)
468 const struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
471 for (i
= 0; i
< entry
->relocation_count
; i
++) {
472 ret
= i915_gem_execbuffer_relocate_entry(vma
->obj
, eb
, &relocs
[i
],
482 i915_gem_execbuffer_relocate(struct eb_vmas
*eb
,
483 struct i915_address_space
*vm
)
485 struct i915_vma
*vma
;
488 /* This is the fast path and we cannot handle a pagefault whilst
489 * holding the struct mutex lest the user pass in the relocations
490 * contained within a mmaped bo. For in such a case we, the page
491 * fault handler would call i915_gem_fault() and we would try to
492 * acquire the struct mutex again. Obviously this is bad and so
493 * lockdep complains vehemently.
496 list_for_each_entry(vma
, &eb
->vmas
, exec_list
) {
497 ret
= i915_gem_execbuffer_relocate_vma(vma
, eb
);
507 need_reloc_mappable(struct i915_vma
*vma
)
509 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
510 return entry
->relocation_count
&& !use_cpu_reloc(vma
->obj
) &&
511 i915_is_ggtt(vma
->vm
);
515 i915_gem_execbuffer_reserve_vma(struct i915_vma
*vma
,
516 struct intel_ring_buffer
*ring
,
519 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
520 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
521 bool has_fenced_gpu_access
= INTEL_INFO(ring
->dev
)->gen
< 4;
522 bool need_fence
, need_mappable
;
523 struct drm_i915_gem_object
*obj
= vma
->obj
;
527 has_fenced_gpu_access
&&
528 entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
&&
529 obj
->tiling_mode
!= I915_TILING_NONE
;
530 need_mappable
= need_fence
|| need_reloc_mappable(vma
);
532 ret
= i915_gem_object_pin(obj
, vma
->vm
, entry
->alignment
, need_mappable
,
537 entry
->flags
|= __EXEC_OBJECT_HAS_PIN
;
539 if (has_fenced_gpu_access
) {
540 if (entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
) {
541 ret
= i915_gem_object_get_fence(obj
);
545 if (i915_gem_object_pin_fence(obj
))
546 entry
->flags
|= __EXEC_OBJECT_HAS_FENCE
;
548 obj
->pending_fenced_gpu_access
= true;
552 /* Ensure ppgtt mapping exists if needed */
553 if (dev_priv
->mm
.aliasing_ppgtt
&& !obj
->has_aliasing_ppgtt_mapping
) {
554 i915_ppgtt_bind_object(dev_priv
->mm
.aliasing_ppgtt
,
555 obj
, obj
->cache_level
);
557 obj
->has_aliasing_ppgtt_mapping
= 1;
560 if (entry
->offset
!= vma
->node
.start
) {
561 entry
->offset
= vma
->node
.start
;
565 if (entry
->flags
& EXEC_OBJECT_WRITE
) {
566 obj
->base
.pending_read_domains
= I915_GEM_DOMAIN_RENDER
;
567 obj
->base
.pending_write_domain
= I915_GEM_DOMAIN_RENDER
;
570 if (entry
->flags
& EXEC_OBJECT_NEEDS_GTT
&&
571 !obj
->has_global_gtt_mapping
)
572 i915_gem_gtt_bind_object(obj
, obj
->cache_level
);
578 i915_gem_execbuffer_reserve(struct intel_ring_buffer
*ring
,
579 struct list_head
*vmas
,
582 struct drm_i915_gem_object
*obj
;
583 struct i915_vma
*vma
;
584 struct i915_address_space
*vm
;
585 struct list_head ordered_vmas
;
586 bool has_fenced_gpu_access
= INTEL_INFO(ring
->dev
)->gen
< 4;
589 if (list_empty(vmas
))
592 vm
= list_first_entry(vmas
, struct i915_vma
, exec_list
)->vm
;
594 INIT_LIST_HEAD(&ordered_vmas
);
595 while (!list_empty(vmas
)) {
596 struct drm_i915_gem_exec_object2
*entry
;
597 bool need_fence
, need_mappable
;
599 vma
= list_first_entry(vmas
, struct i915_vma
, exec_list
);
601 entry
= vma
->exec_entry
;
604 has_fenced_gpu_access
&&
605 entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
&&
606 obj
->tiling_mode
!= I915_TILING_NONE
;
607 need_mappable
= need_fence
|| need_reloc_mappable(vma
);
610 list_move(&vma
->exec_list
, &ordered_vmas
);
612 list_move_tail(&vma
->exec_list
, &ordered_vmas
);
614 obj
->base
.pending_read_domains
= I915_GEM_GPU_DOMAINS
& ~I915_GEM_DOMAIN_COMMAND
;
615 obj
->base
.pending_write_domain
= 0;
616 obj
->pending_fenced_gpu_access
= false;
618 list_splice(&ordered_vmas
, vmas
);
620 /* Attempt to pin all of the buffers into the GTT.
621 * This is done in 3 phases:
623 * 1a. Unbind all objects that do not match the GTT constraints for
624 * the execbuffer (fenceable, mappable, alignment etc).
625 * 1b. Increment pin count for already bound objects.
626 * 2. Bind new objects.
627 * 3. Decrement pin count.
629 * This avoid unnecessary unbinding of later objects in order to make
630 * room for the earlier objects *unless* we need to defragment.
636 /* Unbind any ill-fitting objects or pin. */
637 list_for_each_entry(vma
, vmas
, exec_list
) {
638 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
639 bool need_fence
, need_mappable
;
643 if (!drm_mm_node_allocated(&vma
->node
))
647 has_fenced_gpu_access
&&
648 entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
&&
649 obj
->tiling_mode
!= I915_TILING_NONE
;
650 need_mappable
= need_fence
|| need_reloc_mappable(vma
);
652 WARN_ON((need_mappable
|| need_fence
) &&
653 !i915_is_ggtt(vma
->vm
));
655 if ((entry
->alignment
&&
656 vma
->node
.start
& (entry
->alignment
- 1)) ||
657 (need_mappable
&& !obj
->map_and_fenceable
))
658 ret
= i915_vma_unbind(vma
);
660 ret
= i915_gem_execbuffer_reserve_vma(vma
, ring
, need_relocs
);
665 /* Bind fresh objects */
666 list_for_each_entry(vma
, vmas
, exec_list
) {
667 if (drm_mm_node_allocated(&vma
->node
))
670 ret
= i915_gem_execbuffer_reserve_vma(vma
, ring
, need_relocs
);
676 if (ret
!= -ENOSPC
|| retry
++)
679 /* Decrement pin count for bound objects */
680 list_for_each_entry(vma
, vmas
, exec_list
)
681 i915_gem_execbuffer_unreserve_vma(vma
);
683 ret
= i915_gem_evict_vm(vm
, true);
690 i915_gem_execbuffer_relocate_slow(struct drm_device
*dev
,
691 struct drm_i915_gem_execbuffer2
*args
,
692 struct drm_file
*file
,
693 struct intel_ring_buffer
*ring
,
695 struct drm_i915_gem_exec_object2
*exec
)
697 struct drm_i915_gem_relocation_entry
*reloc
;
698 struct i915_address_space
*vm
;
699 struct i915_vma
*vma
;
703 unsigned count
= args
->buffer_count
;
705 if (WARN_ON(list_empty(&eb
->vmas
)))
708 vm
= list_first_entry(&eb
->vmas
, struct i915_vma
, exec_list
)->vm
;
710 /* We may process another execbuffer during the unlock... */
711 while (!list_empty(&eb
->vmas
)) {
712 vma
= list_first_entry(&eb
->vmas
, struct i915_vma
, exec_list
);
713 list_del_init(&vma
->exec_list
);
714 i915_gem_execbuffer_unreserve_vma(vma
);
715 drm_gem_object_unreference(&vma
->obj
->base
);
718 mutex_unlock(&dev
->struct_mutex
);
721 for (i
= 0; i
< count
; i
++)
722 total
+= exec
[i
].relocation_count
;
724 reloc_offset
= drm_malloc_ab(count
, sizeof(*reloc_offset
));
725 reloc
= drm_malloc_ab(total
, sizeof(*reloc
));
726 if (reloc
== NULL
|| reloc_offset
== NULL
) {
727 drm_free_large(reloc
);
728 drm_free_large(reloc_offset
);
729 mutex_lock(&dev
->struct_mutex
);
734 for (i
= 0; i
< count
; i
++) {
735 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
736 u64 invalid_offset
= (u64
)-1;
739 user_relocs
= to_user_ptr(exec
[i
].relocs_ptr
);
741 if (copy_from_user(reloc
+total
, user_relocs
,
742 exec
[i
].relocation_count
* sizeof(*reloc
))) {
744 mutex_lock(&dev
->struct_mutex
);
748 /* As we do not update the known relocation offsets after
749 * relocating (due to the complexities in lock handling),
750 * we need to mark them as invalid now so that we force the
751 * relocation processing next time. Just in case the target
752 * object is evicted and then rebound into its old
753 * presumed_offset before the next execbuffer - if that
754 * happened we would make the mistake of assuming that the
755 * relocations were valid.
757 for (j
= 0; j
< exec
[i
].relocation_count
; j
++) {
758 if (copy_to_user(&user_relocs
[j
].presumed_offset
,
760 sizeof(invalid_offset
))) {
762 mutex_lock(&dev
->struct_mutex
);
767 reloc_offset
[i
] = total
;
768 total
+= exec
[i
].relocation_count
;
771 ret
= i915_mutex_lock_interruptible(dev
);
773 mutex_lock(&dev
->struct_mutex
);
777 /* reacquire the objects */
779 ret
= eb_lookup_vmas(eb
, exec
, args
, vm
, file
);
783 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
784 ret
= i915_gem_execbuffer_reserve(ring
, &eb
->vmas
, &need_relocs
);
788 list_for_each_entry(vma
, &eb
->vmas
, exec_list
) {
789 int offset
= vma
->exec_entry
- exec
;
790 ret
= i915_gem_execbuffer_relocate_vma_slow(vma
, eb
,
791 reloc
+ reloc_offset
[offset
]);
796 /* Leave the user relocations as are, this is the painfully slow path,
797 * and we want to avoid the complication of dropping the lock whilst
798 * having buffers reserved in the aperture and so causing spurious
799 * ENOSPC for random operations.
803 drm_free_large(reloc
);
804 drm_free_large(reloc_offset
);
809 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer
*ring
,
810 struct list_head
*vmas
)
812 struct i915_vma
*vma
;
813 uint32_t flush_domains
= 0;
814 bool flush_chipset
= false;
817 list_for_each_entry(vma
, vmas
, exec_list
) {
818 struct drm_i915_gem_object
*obj
= vma
->obj
;
819 ret
= i915_gem_object_sync(obj
, ring
);
823 if (obj
->base
.write_domain
& I915_GEM_DOMAIN_CPU
)
824 flush_chipset
|= i915_gem_clflush_object(obj
, false);
826 flush_domains
|= obj
->base
.write_domain
;
830 i915_gem_chipset_flush(ring
->dev
);
832 if (flush_domains
& I915_GEM_DOMAIN_GTT
)
835 /* Unconditionally invalidate gpu caches and ensure that we do flush
836 * any residual writes from the previous batch.
838 return intel_ring_invalidate_all_caches(ring
);
842 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2
*exec
)
844 if (exec
->flags
& __I915_EXEC_UNKNOWN_FLAGS
)
847 return ((exec
->batch_start_offset
| exec
->batch_len
) & 0x7) == 0;
851 validate_exec_list(struct drm_i915_gem_exec_object2
*exec
,
855 unsigned relocs_total
= 0;
856 unsigned relocs_max
= UINT_MAX
/ sizeof(struct drm_i915_gem_relocation_entry
);
858 for (i
= 0; i
< count
; i
++) {
859 char __user
*ptr
= to_user_ptr(exec
[i
].relocs_ptr
);
860 int length
; /* limited by fault_in_pages_readable() */
862 if (exec
[i
].flags
& __EXEC_OBJECT_UNKNOWN_FLAGS
)
865 /* First check for malicious input causing overflow in
866 * the worst case where we need to allocate the entire
867 * relocation tree as a single array.
869 if (exec
[i
].relocation_count
> relocs_max
- relocs_total
)
871 relocs_total
+= exec
[i
].relocation_count
;
873 length
= exec
[i
].relocation_count
*
874 sizeof(struct drm_i915_gem_relocation_entry
);
876 * We must check that the entire relocation array is safe
877 * to read, but since we may need to update the presumed
878 * offsets during execution, check for full write access.
880 if (!access_ok(VERIFY_WRITE
, ptr
, length
))
883 if (likely(!i915_prefault_disable
)) {
884 if (fault_in_multipages_readable(ptr
, length
))
893 i915_gem_execbuffer_move_to_active(struct list_head
*vmas
,
894 struct intel_ring_buffer
*ring
)
896 struct i915_vma
*vma
;
898 list_for_each_entry(vma
, vmas
, exec_list
) {
899 struct drm_i915_gem_object
*obj
= vma
->obj
;
900 u32 old_read
= obj
->base
.read_domains
;
901 u32 old_write
= obj
->base
.write_domain
;
903 obj
->base
.write_domain
= obj
->base
.pending_write_domain
;
904 if (obj
->base
.write_domain
== 0)
905 obj
->base
.pending_read_domains
|= obj
->base
.read_domains
;
906 obj
->base
.read_domains
= obj
->base
.pending_read_domains
;
907 obj
->fenced_gpu_access
= obj
->pending_fenced_gpu_access
;
909 i915_vma_move_to_active(vma
, ring
);
910 if (obj
->base
.write_domain
) {
912 obj
->last_write_seqno
= intel_ring_get_seqno(ring
);
913 if (obj
->pin_count
) /* check for potential scanout */
914 intel_mark_fb_busy(obj
, ring
);
917 trace_i915_gem_object_change_domain(obj
, old_read
, old_write
);
922 i915_gem_execbuffer_retire_commands(struct drm_device
*dev
,
923 struct drm_file
*file
,
924 struct intel_ring_buffer
*ring
,
925 struct drm_i915_gem_object
*obj
)
927 /* Unconditionally force add_request to emit a full flush. */
928 ring
->gpu_caches_dirty
= true;
930 /* Add a breadcrumb for the completion of the batch buffer */
931 (void)__i915_add_request(ring
, file
, obj
, NULL
);
935 i915_reset_gen7_sol_offsets(struct drm_device
*dev
,
936 struct intel_ring_buffer
*ring
)
938 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
941 if (!IS_GEN7(dev
) || ring
!= &dev_priv
->ring
[RCS
])
944 ret
= intel_ring_begin(ring
, 4 * 3);
948 for (i
= 0; i
< 4; i
++) {
949 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
950 intel_ring_emit(ring
, GEN7_SO_WRITE_OFFSET(i
));
951 intel_ring_emit(ring
, 0);
954 intel_ring_advance(ring
);
960 i915_gem_do_execbuffer(struct drm_device
*dev
, void *data
,
961 struct drm_file
*file
,
962 struct drm_i915_gem_execbuffer2
*args
,
963 struct drm_i915_gem_exec_object2
*exec
,
964 struct i915_address_space
*vm
)
966 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
968 struct drm_i915_gem_object
*batch_obj
;
969 struct drm_clip_rect
*cliprects
= NULL
;
970 struct intel_ring_buffer
*ring
;
971 struct i915_ctx_hang_stats
*hs
;
972 u32 ctx_id
= i915_execbuffer2_get_context_id(*args
);
973 u32 exec_start
, exec_len
;
978 if (!i915_gem_check_execbuffer(args
))
981 ret
= validate_exec_list(exec
, args
->buffer_count
);
986 if (args
->flags
& I915_EXEC_SECURE
) {
987 if (!file
->is_master
|| !capable(CAP_SYS_ADMIN
))
990 flags
|= I915_DISPATCH_SECURE
;
992 if (args
->flags
& I915_EXEC_IS_PINNED
)
993 flags
|= I915_DISPATCH_PINNED
;
995 switch (args
->flags
& I915_EXEC_RING_MASK
) {
996 case I915_EXEC_DEFAULT
:
997 case I915_EXEC_RENDER
:
998 ring
= &dev_priv
->ring
[RCS
];
1001 ring
= &dev_priv
->ring
[VCS
];
1002 if (ctx_id
!= DEFAULT_CONTEXT_ID
) {
1003 DRM_DEBUG("Ring %s doesn't support contexts\n",
1009 ring
= &dev_priv
->ring
[BCS
];
1010 if (ctx_id
!= DEFAULT_CONTEXT_ID
) {
1011 DRM_DEBUG("Ring %s doesn't support contexts\n",
1016 case I915_EXEC_VEBOX
:
1017 ring
= &dev_priv
->ring
[VECS
];
1018 if (ctx_id
!= DEFAULT_CONTEXT_ID
) {
1019 DRM_DEBUG("Ring %s doesn't support contexts\n",
1026 DRM_DEBUG("execbuf with unknown ring: %d\n",
1027 (int)(args
->flags
& I915_EXEC_RING_MASK
));
1030 if (!intel_ring_initialized(ring
)) {
1031 DRM_DEBUG("execbuf with invalid ring: %d\n",
1032 (int)(args
->flags
& I915_EXEC_RING_MASK
));
1036 mode
= args
->flags
& I915_EXEC_CONSTANTS_MASK
;
1037 mask
= I915_EXEC_CONSTANTS_MASK
;
1039 case I915_EXEC_CONSTANTS_REL_GENERAL
:
1040 case I915_EXEC_CONSTANTS_ABSOLUTE
:
1041 case I915_EXEC_CONSTANTS_REL_SURFACE
:
1042 if (ring
== &dev_priv
->ring
[RCS
] &&
1043 mode
!= dev_priv
->relative_constants_mode
) {
1044 if (INTEL_INFO(dev
)->gen
< 4)
1047 if (INTEL_INFO(dev
)->gen
> 5 &&
1048 mode
== I915_EXEC_CONSTANTS_REL_SURFACE
)
1051 /* The HW changed the meaning on this bit on gen6 */
1052 if (INTEL_INFO(dev
)->gen
>= 6)
1053 mask
&= ~I915_EXEC_CONSTANTS_REL_SURFACE
;
1057 DRM_DEBUG("execbuf with unknown constants: %d\n", mode
);
1061 if (args
->buffer_count
< 1) {
1062 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1066 if (args
->num_cliprects
!= 0) {
1067 if (ring
!= &dev_priv
->ring
[RCS
]) {
1068 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1072 if (INTEL_INFO(dev
)->gen
>= 5) {
1073 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1077 if (args
->num_cliprects
> UINT_MAX
/ sizeof(*cliprects
)) {
1078 DRM_DEBUG("execbuf with %u cliprects\n",
1079 args
->num_cliprects
);
1083 cliprects
= kcalloc(args
->num_cliprects
,
1086 if (cliprects
== NULL
) {
1091 if (copy_from_user(cliprects
,
1092 to_user_ptr(args
->cliprects_ptr
),
1093 sizeof(*cliprects
)*args
->num_cliprects
)) {
1099 ret
= i915_mutex_lock_interruptible(dev
);
1103 if (dev_priv
->ums
.mm_suspended
) {
1104 mutex_unlock(&dev
->struct_mutex
);
1109 eb
= eb_create(args
, vm
);
1111 mutex_unlock(&dev
->struct_mutex
);
1116 /* Look up object handles */
1117 ret
= eb_lookup_vmas(eb
, exec
, args
, vm
, file
);
1121 /* take note of the batch buffer before we might reorder the lists */
1122 batch_obj
= list_entry(eb
->vmas
.prev
, struct i915_vma
, exec_list
)->obj
;
1124 /* Move the objects en-masse into the GTT, evicting if necessary. */
1125 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
1126 ret
= i915_gem_execbuffer_reserve(ring
, &eb
->vmas
, &need_relocs
);
1130 /* The objects are in their final locations, apply the relocations. */
1132 ret
= i915_gem_execbuffer_relocate(eb
, vm
);
1134 if (ret
== -EFAULT
) {
1135 ret
= i915_gem_execbuffer_relocate_slow(dev
, args
, file
, ring
,
1137 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
1143 /* Set the pending read domains for the batch buffer to COMMAND */
1144 if (batch_obj
->base
.pending_write_domain
) {
1145 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1149 batch_obj
->base
.pending_read_domains
|= I915_GEM_DOMAIN_COMMAND
;
1151 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1152 * batch" bit. Hence we need to pin secure batches into the global gtt.
1153 * hsw should have this fixed, but bdw mucks it up again. */
1154 if (flags
& I915_DISPATCH_SECURE
&& !batch_obj
->has_global_gtt_mapping
)
1155 i915_gem_gtt_bind_object(batch_obj
, batch_obj
->cache_level
);
1157 ret
= i915_gem_execbuffer_move_to_gpu(ring
, &eb
->vmas
);
1161 hs
= i915_gem_context_get_hang_stats(dev
, file
, ctx_id
);
1172 ret
= i915_switch_context(ring
, file
, ctx_id
);
1176 if (ring
== &dev_priv
->ring
[RCS
] &&
1177 mode
!= dev_priv
->relative_constants_mode
) {
1178 ret
= intel_ring_begin(ring
, 4);
1182 intel_ring_emit(ring
, MI_NOOP
);
1183 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
1184 intel_ring_emit(ring
, INSTPM
);
1185 intel_ring_emit(ring
, mask
<< 16 | mode
);
1186 intel_ring_advance(ring
);
1188 dev_priv
->relative_constants_mode
= mode
;
1191 if (args
->flags
& I915_EXEC_GEN7_SOL_RESET
) {
1192 ret
= i915_reset_gen7_sol_offsets(dev
, ring
);
1197 exec_start
= i915_gem_obj_offset(batch_obj
, vm
) +
1198 args
->batch_start_offset
;
1199 exec_len
= args
->batch_len
;
1201 for (i
= 0; i
< args
->num_cliprects
; i
++) {
1202 ret
= i915_emit_box(dev
, &cliprects
[i
],
1203 args
->DR1
, args
->DR4
);
1207 ret
= ring
->dispatch_execbuffer(ring
,
1208 exec_start
, exec_len
,
1214 ret
= ring
->dispatch_execbuffer(ring
,
1215 exec_start
, exec_len
,
1221 trace_i915_gem_ring_dispatch(ring
, intel_ring_get_seqno(ring
), flags
);
1223 i915_gem_execbuffer_move_to_active(&eb
->vmas
, ring
);
1224 i915_gem_execbuffer_retire_commands(dev
, file
, ring
, batch_obj
);
1229 mutex_unlock(&dev
->struct_mutex
);
1237 * Legacy execbuffer just creates an exec2 list from the original exec object
1238 * list array and passes it to the real function.
1241 i915_gem_execbuffer(struct drm_device
*dev
, void *data
,
1242 struct drm_file
*file
)
1244 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1245 struct drm_i915_gem_execbuffer
*args
= data
;
1246 struct drm_i915_gem_execbuffer2 exec2
;
1247 struct drm_i915_gem_exec_object
*exec_list
= NULL
;
1248 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1251 if (args
->buffer_count
< 1) {
1252 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1256 /* Copy in the exec list from userland */
1257 exec_list
= drm_malloc_ab(sizeof(*exec_list
), args
->buffer_count
);
1258 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
), args
->buffer_count
);
1259 if (exec_list
== NULL
|| exec2_list
== NULL
) {
1260 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1261 args
->buffer_count
);
1262 drm_free_large(exec_list
);
1263 drm_free_large(exec2_list
);
1266 ret
= copy_from_user(exec_list
,
1267 to_user_ptr(args
->buffers_ptr
),
1268 sizeof(*exec_list
) * args
->buffer_count
);
1270 DRM_DEBUG("copy %d exec entries failed %d\n",
1271 args
->buffer_count
, ret
);
1272 drm_free_large(exec_list
);
1273 drm_free_large(exec2_list
);
1277 for (i
= 0; i
< args
->buffer_count
; i
++) {
1278 exec2_list
[i
].handle
= exec_list
[i
].handle
;
1279 exec2_list
[i
].relocation_count
= exec_list
[i
].relocation_count
;
1280 exec2_list
[i
].relocs_ptr
= exec_list
[i
].relocs_ptr
;
1281 exec2_list
[i
].alignment
= exec_list
[i
].alignment
;
1282 exec2_list
[i
].offset
= exec_list
[i
].offset
;
1283 if (INTEL_INFO(dev
)->gen
< 4)
1284 exec2_list
[i
].flags
= EXEC_OBJECT_NEEDS_FENCE
;
1286 exec2_list
[i
].flags
= 0;
1289 exec2
.buffers_ptr
= args
->buffers_ptr
;
1290 exec2
.buffer_count
= args
->buffer_count
;
1291 exec2
.batch_start_offset
= args
->batch_start_offset
;
1292 exec2
.batch_len
= args
->batch_len
;
1293 exec2
.DR1
= args
->DR1
;
1294 exec2
.DR4
= args
->DR4
;
1295 exec2
.num_cliprects
= args
->num_cliprects
;
1296 exec2
.cliprects_ptr
= args
->cliprects_ptr
;
1297 exec2
.flags
= I915_EXEC_RENDER
;
1298 i915_execbuffer2_set_context_id(exec2
, 0);
1300 ret
= i915_gem_do_execbuffer(dev
, data
, file
, &exec2
, exec2_list
,
1301 &dev_priv
->gtt
.base
);
1303 /* Copy the new buffer offsets back to the user's exec list. */
1304 for (i
= 0; i
< args
->buffer_count
; i
++)
1305 exec_list
[i
].offset
= exec2_list
[i
].offset
;
1306 /* ... and back out to userspace */
1307 ret
= copy_to_user(to_user_ptr(args
->buffers_ptr
),
1309 sizeof(*exec_list
) * args
->buffer_count
);
1312 DRM_DEBUG("failed to copy %d exec entries "
1313 "back to user (%d)\n",
1314 args
->buffer_count
, ret
);
1318 drm_free_large(exec_list
);
1319 drm_free_large(exec2_list
);
1324 i915_gem_execbuffer2(struct drm_device
*dev
, void *data
,
1325 struct drm_file
*file
)
1327 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1328 struct drm_i915_gem_execbuffer2
*args
= data
;
1329 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1332 if (args
->buffer_count
< 1 ||
1333 args
->buffer_count
> UINT_MAX
/ sizeof(*exec2_list
)) {
1334 DRM_DEBUG("execbuf2 with %d buffers\n", args
->buffer_count
);
1338 exec2_list
= kmalloc(sizeof(*exec2_list
)*args
->buffer_count
,
1339 GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
1340 if (exec2_list
== NULL
)
1341 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
),
1342 args
->buffer_count
);
1343 if (exec2_list
== NULL
) {
1344 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1345 args
->buffer_count
);
1348 ret
= copy_from_user(exec2_list
,
1349 to_user_ptr(args
->buffers_ptr
),
1350 sizeof(*exec2_list
) * args
->buffer_count
);
1352 DRM_DEBUG("copy %d exec entries failed %d\n",
1353 args
->buffer_count
, ret
);
1354 drm_free_large(exec2_list
);
1358 ret
= i915_gem_do_execbuffer(dev
, data
, file
, args
, exec2_list
,
1359 &dev_priv
->gtt
.base
);
1361 /* Copy the new buffer offsets back to the user's exec list. */
1362 ret
= copy_to_user(to_user_ptr(args
->buffers_ptr
),
1364 sizeof(*exec2_list
) * args
->buffer_count
);
1367 DRM_DEBUG("failed to copy %d exec entries "
1368 "back to user (%d)\n",
1369 args
->buffer_count
, ret
);
1373 drm_free_large(exec2_list
);