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>
35 #include <linux/uaccess.h>
37 #define __EXEC_OBJECT_HAS_PIN (1<<31)
38 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
39 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
40 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
42 #define BATCH_OFFSET_BIAS (256*1024)
45 struct list_head vmas
;
48 struct i915_vma
*lut
[0];
49 struct hlist_head buckets
[0];
53 static struct eb_vmas
*
54 eb_create(struct drm_i915_gem_execbuffer2
*args
)
56 struct eb_vmas
*eb
= NULL
;
58 if (args
->flags
& I915_EXEC_HANDLE_LUT
) {
59 unsigned size
= args
->buffer_count
;
60 size
*= sizeof(struct i915_vma
*);
61 size
+= sizeof(struct eb_vmas
);
62 eb
= kmalloc(size
, GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
66 unsigned size
= args
->buffer_count
;
67 unsigned count
= PAGE_SIZE
/ sizeof(struct hlist_head
) / 2;
68 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE
/ sizeof(struct hlist_head
));
69 while (count
> 2*size
)
71 eb
= kzalloc(count
*sizeof(struct hlist_head
) +
72 sizeof(struct eb_vmas
),
79 eb
->and = -args
->buffer_count
;
81 INIT_LIST_HEAD(&eb
->vmas
);
86 eb_reset(struct eb_vmas
*eb
)
89 memset(eb
->buckets
, 0, (eb
->and+1)*sizeof(struct hlist_head
));
93 eb_lookup_vmas(struct eb_vmas
*eb
,
94 struct drm_i915_gem_exec_object2
*exec
,
95 const struct drm_i915_gem_execbuffer2
*args
,
96 struct i915_address_space
*vm
,
97 struct drm_file
*file
)
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
;
134 obj
= list_first_entry(&objects
,
135 struct drm_i915_gem_object
,
139 * NOTE: We can leak any vmas created here when something fails
140 * later on. But that's no issue since vma_unbind can deal with
141 * vmas which are not actually bound. And since only
142 * lookup_or_create exists as an interface to get at the vma
143 * from the (obj, vm) we don't run the risk of creating
144 * duplicated vmas for the same vm.
146 vma
= i915_gem_obj_lookup_or_create_vma(obj
, vm
);
148 DRM_DEBUG("Failed to lookup VMA\n");
153 /* Transfer ownership from the objects list to the vmas list. */
154 list_add_tail(&vma
->exec_list
, &eb
->vmas
);
155 list_del_init(&obj
->obj_exec_link
);
157 vma
->exec_entry
= &exec
[i
];
161 uint32_t handle
= args
->flags
& I915_EXEC_HANDLE_LUT
? i
: exec
[i
].handle
;
162 vma
->exec_handle
= handle
;
163 hlist_add_head(&vma
->exec_node
,
164 &eb
->buckets
[handle
& eb
->and]);
173 while (!list_empty(&objects
)) {
174 obj
= list_first_entry(&objects
,
175 struct drm_i915_gem_object
,
177 list_del_init(&obj
->obj_exec_link
);
178 drm_gem_object_unreference(&obj
->base
);
181 * Objects already transfered to the vmas list will be unreferenced by
188 static struct i915_vma
*eb_get_vma(struct eb_vmas
*eb
, unsigned long handle
)
191 if (handle
>= -eb
->and)
193 return eb
->lut
[handle
];
195 struct hlist_head
*head
;
196 struct hlist_node
*node
;
198 head
= &eb
->buckets
[handle
& eb
->and];
199 hlist_for_each(node
, head
) {
200 struct i915_vma
*vma
;
202 vma
= hlist_entry(node
, struct i915_vma
, exec_node
);
203 if (vma
->exec_handle
== handle
)
211 i915_gem_execbuffer_unreserve_vma(struct i915_vma
*vma
)
213 struct drm_i915_gem_exec_object2
*entry
;
214 struct drm_i915_gem_object
*obj
= vma
->obj
;
216 if (!drm_mm_node_allocated(&vma
->node
))
219 entry
= vma
->exec_entry
;
221 if (entry
->flags
& __EXEC_OBJECT_HAS_FENCE
)
222 i915_gem_object_unpin_fence(obj
);
224 if (entry
->flags
& __EXEC_OBJECT_HAS_PIN
)
227 entry
->flags
&= ~(__EXEC_OBJECT_HAS_FENCE
| __EXEC_OBJECT_HAS_PIN
);
230 static void eb_destroy(struct eb_vmas
*eb
)
232 while (!list_empty(&eb
->vmas
)) {
233 struct i915_vma
*vma
;
235 vma
= list_first_entry(&eb
->vmas
,
238 list_del_init(&vma
->exec_list
);
239 i915_gem_execbuffer_unreserve_vma(vma
);
240 drm_gem_object_unreference(&vma
->obj
->base
);
245 static inline int use_cpu_reloc(struct drm_i915_gem_object
*obj
)
247 return (HAS_LLC(obj
->base
.dev
) ||
248 obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
||
249 obj
->cache_level
!= I915_CACHE_NONE
);
253 relocate_entry_cpu(struct drm_i915_gem_object
*obj
,
254 struct drm_i915_gem_relocation_entry
*reloc
,
255 uint64_t target_offset
)
257 struct drm_device
*dev
= obj
->base
.dev
;
258 uint32_t page_offset
= offset_in_page(reloc
->offset
);
259 uint64_t delta
= reloc
->delta
+ target_offset
;
263 ret
= i915_gem_object_set_to_cpu_domain(obj
, true);
267 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
268 reloc
->offset
>> PAGE_SHIFT
));
269 *(uint32_t *)(vaddr
+ page_offset
) = lower_32_bits(delta
);
271 if (INTEL_INFO(dev
)->gen
>= 8) {
272 page_offset
= offset_in_page(page_offset
+ sizeof(uint32_t));
274 if (page_offset
== 0) {
275 kunmap_atomic(vaddr
);
276 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
277 (reloc
->offset
+ sizeof(uint32_t)) >> PAGE_SHIFT
));
280 *(uint32_t *)(vaddr
+ page_offset
) = upper_32_bits(delta
);
283 kunmap_atomic(vaddr
);
289 relocate_entry_gtt(struct drm_i915_gem_object
*obj
,
290 struct drm_i915_gem_relocation_entry
*reloc
,
291 uint64_t target_offset
)
293 struct drm_device
*dev
= obj
->base
.dev
;
294 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
295 uint64_t delta
= reloc
->delta
+ target_offset
;
297 void __iomem
*reloc_page
;
300 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
304 ret
= i915_gem_object_put_fence(obj
);
308 /* Map the page containing the relocation we're going to perform. */
309 offset
= i915_gem_obj_ggtt_offset(obj
);
310 offset
+= reloc
->offset
;
311 reloc_page
= io_mapping_map_atomic_wc(dev_priv
->gtt
.mappable
,
313 iowrite32(lower_32_bits(delta
), reloc_page
+ offset_in_page(offset
));
315 if (INTEL_INFO(dev
)->gen
>= 8) {
316 offset
+= sizeof(uint32_t);
318 if (offset_in_page(offset
) == 0) {
319 io_mapping_unmap_atomic(reloc_page
);
321 io_mapping_map_atomic_wc(dev_priv
->gtt
.mappable
,
325 iowrite32(upper_32_bits(delta
),
326 reloc_page
+ offset_in_page(offset
));
329 io_mapping_unmap_atomic(reloc_page
);
335 clflush_write32(void *addr
, uint32_t value
)
337 /* This is not a fast path, so KISS. */
338 drm_clflush_virt_range(addr
, sizeof(uint32_t));
339 *(uint32_t *)addr
= value
;
340 drm_clflush_virt_range(addr
, sizeof(uint32_t));
344 relocate_entry_clflush(struct drm_i915_gem_object
*obj
,
345 struct drm_i915_gem_relocation_entry
*reloc
,
346 uint64_t target_offset
)
348 struct drm_device
*dev
= obj
->base
.dev
;
349 uint32_t page_offset
= offset_in_page(reloc
->offset
);
350 uint64_t delta
= (int)reloc
->delta
+ target_offset
;
354 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
358 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
359 reloc
->offset
>> PAGE_SHIFT
));
360 clflush_write32(vaddr
+ page_offset
, lower_32_bits(delta
));
362 if (INTEL_INFO(dev
)->gen
>= 8) {
363 page_offset
= offset_in_page(page_offset
+ sizeof(uint32_t));
365 if (page_offset
== 0) {
366 kunmap_atomic(vaddr
);
367 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
368 (reloc
->offset
+ sizeof(uint32_t)) >> PAGE_SHIFT
));
371 clflush_write32(vaddr
+ page_offset
, upper_32_bits(delta
));
374 kunmap_atomic(vaddr
);
380 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object
*obj
,
382 struct drm_i915_gem_relocation_entry
*reloc
)
384 struct drm_device
*dev
= obj
->base
.dev
;
385 struct drm_gem_object
*target_obj
;
386 struct drm_i915_gem_object
*target_i915_obj
;
387 struct i915_vma
*target_vma
;
388 uint64_t target_offset
;
391 /* we've already hold a reference to all valid objects */
392 target_vma
= eb_get_vma(eb
, reloc
->target_handle
);
393 if (unlikely(target_vma
== NULL
))
395 target_i915_obj
= target_vma
->obj
;
396 target_obj
= &target_vma
->obj
->base
;
398 target_offset
= target_vma
->node
.start
;
400 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
401 * pipe_control writes because the gpu doesn't properly redirect them
402 * through the ppgtt for non_secure batchbuffers. */
403 if (unlikely(IS_GEN6(dev
) &&
404 reloc
->write_domain
== I915_GEM_DOMAIN_INSTRUCTION
)) {
405 ret
= i915_vma_bind(target_vma
, target_i915_obj
->cache_level
,
407 if (WARN_ONCE(ret
, "Unexpected failure to bind target VMA!"))
411 /* Validate that the target is in a valid r/w GPU domain */
412 if (unlikely(reloc
->write_domain
& (reloc
->write_domain
- 1))) {
413 DRM_DEBUG("reloc with multiple write domains: "
414 "obj %p target %d offset %d "
415 "read %08x write %08x",
416 obj
, reloc
->target_handle
,
419 reloc
->write_domain
);
422 if (unlikely((reloc
->write_domain
| reloc
->read_domains
)
423 & ~I915_GEM_GPU_DOMAINS
)) {
424 DRM_DEBUG("reloc with read/write non-GPU domains: "
425 "obj %p target %d offset %d "
426 "read %08x write %08x",
427 obj
, reloc
->target_handle
,
430 reloc
->write_domain
);
434 target_obj
->pending_read_domains
|= reloc
->read_domains
;
435 target_obj
->pending_write_domain
|= reloc
->write_domain
;
437 /* If the relocation already has the right value in it, no
438 * more work needs to be done.
440 if (target_offset
== reloc
->presumed_offset
)
443 /* Check that the relocation address is valid... */
444 if (unlikely(reloc
->offset
>
445 obj
->base
.size
- (INTEL_INFO(dev
)->gen
>= 8 ? 8 : 4))) {
446 DRM_DEBUG("Relocation beyond object bounds: "
447 "obj %p target %d offset %d size %d.\n",
448 obj
, reloc
->target_handle
,
450 (int) obj
->base
.size
);
453 if (unlikely(reloc
->offset
& 3)) {
454 DRM_DEBUG("Relocation not 4-byte aligned: "
455 "obj %p target %d offset %d.\n",
456 obj
, reloc
->target_handle
,
457 (int) reloc
->offset
);
461 /* We can't wait for rendering with pagefaults disabled */
462 if (obj
->active
&& pagefault_disabled())
465 if (use_cpu_reloc(obj
))
466 ret
= relocate_entry_cpu(obj
, reloc
, target_offset
);
467 else if (obj
->map_and_fenceable
)
468 ret
= relocate_entry_gtt(obj
, reloc
, target_offset
);
469 else if (cpu_has_clflush
)
470 ret
= relocate_entry_clflush(obj
, reloc
, target_offset
);
472 WARN_ONCE(1, "Impossible case in relocation handling\n");
479 /* and update the user's relocation entry */
480 reloc
->presumed_offset
= target_offset
;
486 i915_gem_execbuffer_relocate_vma(struct i915_vma
*vma
,
489 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
490 struct drm_i915_gem_relocation_entry stack_reloc
[N_RELOC(512)];
491 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
492 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
495 user_relocs
= to_user_ptr(entry
->relocs_ptr
);
497 remain
= entry
->relocation_count
;
499 struct drm_i915_gem_relocation_entry
*r
= stack_reloc
;
501 if (count
> ARRAY_SIZE(stack_reloc
))
502 count
= ARRAY_SIZE(stack_reloc
);
505 if (__copy_from_user_inatomic(r
, user_relocs
, count
*sizeof(r
[0])))
509 u64 offset
= r
->presumed_offset
;
511 ret
= i915_gem_execbuffer_relocate_entry(vma
->obj
, eb
, r
);
515 if (r
->presumed_offset
!= offset
&&
516 __copy_to_user_inatomic(&user_relocs
->presumed_offset
,
518 sizeof(r
->presumed_offset
))) {
532 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma
*vma
,
534 struct drm_i915_gem_relocation_entry
*relocs
)
536 const struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
539 for (i
= 0; i
< entry
->relocation_count
; i
++) {
540 ret
= i915_gem_execbuffer_relocate_entry(vma
->obj
, eb
, &relocs
[i
]);
549 i915_gem_execbuffer_relocate(struct eb_vmas
*eb
)
551 struct i915_vma
*vma
;
554 /* This is the fast path and we cannot handle a pagefault whilst
555 * holding the struct mutex lest the user pass in the relocations
556 * contained within a mmaped bo. For in such a case we, the page
557 * fault handler would call i915_gem_fault() and we would try to
558 * acquire the struct mutex again. Obviously this is bad and so
559 * lockdep complains vehemently.
562 list_for_each_entry(vma
, &eb
->vmas
, exec_list
) {
563 ret
= i915_gem_execbuffer_relocate_vma(vma
, eb
);
572 static bool only_mappable_for_reloc(unsigned int flags
)
574 return (flags
& (EXEC_OBJECT_NEEDS_FENCE
| __EXEC_OBJECT_NEEDS_MAP
)) ==
575 __EXEC_OBJECT_NEEDS_MAP
;
579 i915_gem_execbuffer_reserve_vma(struct i915_vma
*vma
,
580 struct intel_engine_cs
*ring
,
583 struct drm_i915_gem_object
*obj
= vma
->obj
;
584 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
589 if (entry
->flags
& EXEC_OBJECT_NEEDS_GTT
)
592 if (!drm_mm_node_allocated(&vma
->node
)) {
593 if (entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
)
594 flags
|= PIN_GLOBAL
| PIN_MAPPABLE
;
595 if (entry
->flags
& __EXEC_OBJECT_NEEDS_BIAS
)
596 flags
|= BATCH_OFFSET_BIAS
| PIN_OFFSET_BIAS
;
599 ret
= i915_gem_object_pin(obj
, vma
->vm
, entry
->alignment
, flags
);
600 if ((ret
== -ENOSPC
|| ret
== -E2BIG
) &&
601 only_mappable_for_reloc(entry
->flags
))
602 ret
= i915_gem_object_pin(obj
, vma
->vm
,
604 flags
& ~PIN_MAPPABLE
);
608 entry
->flags
|= __EXEC_OBJECT_HAS_PIN
;
610 if (entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
) {
611 ret
= i915_gem_object_get_fence(obj
);
615 if (i915_gem_object_pin_fence(obj
))
616 entry
->flags
|= __EXEC_OBJECT_HAS_FENCE
;
619 if (entry
->offset
!= vma
->node
.start
) {
620 entry
->offset
= vma
->node
.start
;
624 if (entry
->flags
& EXEC_OBJECT_WRITE
) {
625 obj
->base
.pending_read_domains
= I915_GEM_DOMAIN_RENDER
;
626 obj
->base
.pending_write_domain
= I915_GEM_DOMAIN_RENDER
;
633 need_reloc_mappable(struct i915_vma
*vma
)
635 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
637 if (entry
->relocation_count
== 0)
640 if (!i915_is_ggtt(vma
->vm
))
643 /* See also use_cpu_reloc() */
644 if (HAS_LLC(vma
->obj
->base
.dev
))
647 if (vma
->obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
)
654 eb_vma_misplaced(struct i915_vma
*vma
)
656 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
657 struct drm_i915_gem_object
*obj
= vma
->obj
;
659 WARN_ON(entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
&&
660 !i915_is_ggtt(vma
->vm
));
662 if (entry
->alignment
&&
663 vma
->node
.start
& (entry
->alignment
- 1))
666 if (entry
->flags
& __EXEC_OBJECT_NEEDS_BIAS
&&
667 vma
->node
.start
< BATCH_OFFSET_BIAS
)
670 /* avoid costly ping-pong once a batch bo ended up non-mappable */
671 if (entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
&& !obj
->map_and_fenceable
)
672 return !only_mappable_for_reloc(entry
->flags
);
678 i915_gem_execbuffer_reserve(struct intel_engine_cs
*ring
,
679 struct list_head
*vmas
,
682 struct drm_i915_gem_object
*obj
;
683 struct i915_vma
*vma
;
684 struct i915_address_space
*vm
;
685 struct list_head ordered_vmas
;
686 bool has_fenced_gpu_access
= INTEL_INFO(ring
->dev
)->gen
< 4;
689 i915_gem_retire_requests_ring(ring
);
691 vm
= list_first_entry(vmas
, struct i915_vma
, exec_list
)->vm
;
693 INIT_LIST_HEAD(&ordered_vmas
);
694 while (!list_empty(vmas
)) {
695 struct drm_i915_gem_exec_object2
*entry
;
696 bool need_fence
, need_mappable
;
698 vma
= list_first_entry(vmas
, struct i915_vma
, exec_list
);
700 entry
= vma
->exec_entry
;
702 if (!has_fenced_gpu_access
)
703 entry
->flags
&= ~EXEC_OBJECT_NEEDS_FENCE
;
705 entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
&&
706 obj
->tiling_mode
!= I915_TILING_NONE
;
707 need_mappable
= need_fence
|| need_reloc_mappable(vma
);
710 entry
->flags
|= __EXEC_OBJECT_NEEDS_MAP
;
711 list_move(&vma
->exec_list
, &ordered_vmas
);
713 list_move_tail(&vma
->exec_list
, &ordered_vmas
);
715 obj
->base
.pending_read_domains
= I915_GEM_GPU_DOMAINS
& ~I915_GEM_DOMAIN_COMMAND
;
716 obj
->base
.pending_write_domain
= 0;
718 list_splice(&ordered_vmas
, vmas
);
720 /* Attempt to pin all of the buffers into the GTT.
721 * This is done in 3 phases:
723 * 1a. Unbind all objects that do not match the GTT constraints for
724 * the execbuffer (fenceable, mappable, alignment etc).
725 * 1b. Increment pin count for already bound objects.
726 * 2. Bind new objects.
727 * 3. Decrement pin count.
729 * This avoid unnecessary unbinding of later objects in order to make
730 * room for the earlier objects *unless* we need to defragment.
736 /* Unbind any ill-fitting objects or pin. */
737 list_for_each_entry(vma
, vmas
, exec_list
) {
738 if (!drm_mm_node_allocated(&vma
->node
))
741 if (eb_vma_misplaced(vma
))
742 ret
= i915_vma_unbind(vma
);
744 ret
= i915_gem_execbuffer_reserve_vma(vma
, ring
, need_relocs
);
749 /* Bind fresh objects */
750 list_for_each_entry(vma
, vmas
, exec_list
) {
751 if (drm_mm_node_allocated(&vma
->node
))
754 ret
= i915_gem_execbuffer_reserve_vma(vma
, ring
, need_relocs
);
760 if (ret
!= -ENOSPC
|| retry
++)
763 /* Decrement pin count for bound objects */
764 list_for_each_entry(vma
, vmas
, exec_list
)
765 i915_gem_execbuffer_unreserve_vma(vma
);
767 ret
= i915_gem_evict_vm(vm
, true);
774 i915_gem_execbuffer_relocate_slow(struct drm_device
*dev
,
775 struct drm_i915_gem_execbuffer2
*args
,
776 struct drm_file
*file
,
777 struct intel_engine_cs
*ring
,
779 struct drm_i915_gem_exec_object2
*exec
)
781 struct drm_i915_gem_relocation_entry
*reloc
;
782 struct i915_address_space
*vm
;
783 struct i915_vma
*vma
;
787 unsigned count
= args
->buffer_count
;
789 vm
= list_first_entry(&eb
->vmas
, struct i915_vma
, exec_list
)->vm
;
791 /* We may process another execbuffer during the unlock... */
792 while (!list_empty(&eb
->vmas
)) {
793 vma
= list_first_entry(&eb
->vmas
, struct i915_vma
, exec_list
);
794 list_del_init(&vma
->exec_list
);
795 i915_gem_execbuffer_unreserve_vma(vma
);
796 drm_gem_object_unreference(&vma
->obj
->base
);
799 mutex_unlock(&dev
->struct_mutex
);
802 for (i
= 0; i
< count
; i
++)
803 total
+= exec
[i
].relocation_count
;
805 reloc_offset
= drm_malloc_ab(count
, sizeof(*reloc_offset
));
806 reloc
= drm_malloc_ab(total
, sizeof(*reloc
));
807 if (reloc
== NULL
|| reloc_offset
== NULL
) {
808 drm_free_large(reloc
);
809 drm_free_large(reloc_offset
);
810 mutex_lock(&dev
->struct_mutex
);
815 for (i
= 0; i
< count
; i
++) {
816 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
817 u64 invalid_offset
= (u64
)-1;
820 user_relocs
= to_user_ptr(exec
[i
].relocs_ptr
);
822 if (copy_from_user(reloc
+total
, user_relocs
,
823 exec
[i
].relocation_count
* sizeof(*reloc
))) {
825 mutex_lock(&dev
->struct_mutex
);
829 /* As we do not update the known relocation offsets after
830 * relocating (due to the complexities in lock handling),
831 * we need to mark them as invalid now so that we force the
832 * relocation processing next time. Just in case the target
833 * object is evicted and then rebound into its old
834 * presumed_offset before the next execbuffer - if that
835 * happened we would make the mistake of assuming that the
836 * relocations were valid.
838 for (j
= 0; j
< exec
[i
].relocation_count
; j
++) {
839 if (__copy_to_user(&user_relocs
[j
].presumed_offset
,
841 sizeof(invalid_offset
))) {
843 mutex_lock(&dev
->struct_mutex
);
848 reloc_offset
[i
] = total
;
849 total
+= exec
[i
].relocation_count
;
852 ret
= i915_mutex_lock_interruptible(dev
);
854 mutex_lock(&dev
->struct_mutex
);
858 /* reacquire the objects */
860 ret
= eb_lookup_vmas(eb
, exec
, args
, vm
, file
);
864 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
865 ret
= i915_gem_execbuffer_reserve(ring
, &eb
->vmas
, &need_relocs
);
869 list_for_each_entry(vma
, &eb
->vmas
, exec_list
) {
870 int offset
= vma
->exec_entry
- exec
;
871 ret
= i915_gem_execbuffer_relocate_vma_slow(vma
, eb
,
872 reloc
+ reloc_offset
[offset
]);
877 /* Leave the user relocations as are, this is the painfully slow path,
878 * and we want to avoid the complication of dropping the lock whilst
879 * having buffers reserved in the aperture and so causing spurious
880 * ENOSPC for random operations.
884 drm_free_large(reloc
);
885 drm_free_large(reloc_offset
);
890 i915_gem_execbuffer_move_to_gpu(struct intel_engine_cs
*ring
,
891 struct list_head
*vmas
)
893 const unsigned other_rings
= ~intel_ring_flag(ring
);
894 struct i915_vma
*vma
;
895 uint32_t flush_domains
= 0;
896 bool flush_chipset
= false;
899 list_for_each_entry(vma
, vmas
, exec_list
) {
900 struct drm_i915_gem_object
*obj
= vma
->obj
;
902 if (obj
->active
& other_rings
) {
903 ret
= i915_gem_object_sync(obj
, ring
);
908 if (obj
->base
.write_domain
& I915_GEM_DOMAIN_CPU
)
909 flush_chipset
|= i915_gem_clflush_object(obj
, false);
911 flush_domains
|= obj
->base
.write_domain
;
915 i915_gem_chipset_flush(ring
->dev
);
917 if (flush_domains
& I915_GEM_DOMAIN_GTT
)
920 /* Unconditionally invalidate gpu caches and ensure that we do flush
921 * any residual writes from the previous batch.
923 return intel_ring_invalidate_all_caches(ring
);
927 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2
*exec
)
929 if (exec
->flags
& __I915_EXEC_UNKNOWN_FLAGS
)
932 return ((exec
->batch_start_offset
| exec
->batch_len
) & 0x7) == 0;
936 validate_exec_list(struct drm_device
*dev
,
937 struct drm_i915_gem_exec_object2
*exec
,
940 unsigned relocs_total
= 0;
941 unsigned relocs_max
= UINT_MAX
/ sizeof(struct drm_i915_gem_relocation_entry
);
942 unsigned invalid_flags
;
945 invalid_flags
= __EXEC_OBJECT_UNKNOWN_FLAGS
;
946 if (USES_FULL_PPGTT(dev
))
947 invalid_flags
|= EXEC_OBJECT_NEEDS_GTT
;
949 for (i
= 0; i
< count
; i
++) {
950 char __user
*ptr
= to_user_ptr(exec
[i
].relocs_ptr
);
951 int length
; /* limited by fault_in_pages_readable() */
953 if (exec
[i
].flags
& invalid_flags
)
956 /* First check for malicious input causing overflow in
957 * the worst case where we need to allocate the entire
958 * relocation tree as a single array.
960 if (exec
[i
].relocation_count
> relocs_max
- relocs_total
)
962 relocs_total
+= exec
[i
].relocation_count
;
964 length
= exec
[i
].relocation_count
*
965 sizeof(struct drm_i915_gem_relocation_entry
);
967 * We must check that the entire relocation array is safe
968 * to read, but since we may need to update the presumed
969 * offsets during execution, check for full write access.
971 if (!access_ok(VERIFY_WRITE
, ptr
, length
))
974 if (likely(!i915
.prefault_disable
)) {
975 if (fault_in_multipages_readable(ptr
, length
))
983 static struct intel_context
*
984 i915_gem_validate_context(struct drm_device
*dev
, struct drm_file
*file
,
985 struct intel_engine_cs
*ring
, const u32 ctx_id
)
987 struct intel_context
*ctx
= NULL
;
988 struct i915_ctx_hang_stats
*hs
;
990 if (ring
->id
!= RCS
&& ctx_id
!= DEFAULT_CONTEXT_HANDLE
)
991 return ERR_PTR(-EINVAL
);
993 ctx
= i915_gem_context_get(file
->driver_priv
, ctx_id
);
997 hs
= &ctx
->hang_stats
;
999 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id
);
1000 return ERR_PTR(-EIO
);
1003 if (i915
.enable_execlists
&& !ctx
->engine
[ring
->id
].state
) {
1004 int ret
= intel_lr_context_deferred_create(ctx
, ring
);
1006 DRM_DEBUG("Could not create LRC %u: %d\n", ctx_id
, ret
);
1007 return ERR_PTR(ret
);
1015 i915_gem_execbuffer_move_to_active(struct list_head
*vmas
,
1016 struct intel_engine_cs
*ring
)
1018 struct drm_i915_gem_request
*req
= intel_ring_get_request(ring
);
1019 struct i915_vma
*vma
;
1021 list_for_each_entry(vma
, vmas
, exec_list
) {
1022 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
1023 struct drm_i915_gem_object
*obj
= vma
->obj
;
1024 u32 old_read
= obj
->base
.read_domains
;
1025 u32 old_write
= obj
->base
.write_domain
;
1027 obj
->dirty
= 1; /* be paranoid */
1028 obj
->base
.write_domain
= obj
->base
.pending_write_domain
;
1029 if (obj
->base
.write_domain
== 0)
1030 obj
->base
.pending_read_domains
|= obj
->base
.read_domains
;
1031 obj
->base
.read_domains
= obj
->base
.pending_read_domains
;
1033 i915_vma_move_to_active(vma
, ring
);
1034 if (obj
->base
.write_domain
) {
1035 i915_gem_request_assign(&obj
->last_write_req
, req
);
1037 intel_fb_obj_invalidate(obj
, ring
, ORIGIN_CS
);
1039 /* update for the implicit flush after a batch */
1040 obj
->base
.write_domain
&= ~I915_GEM_GPU_DOMAINS
;
1042 if (entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
) {
1043 i915_gem_request_assign(&obj
->last_fenced_req
, req
);
1044 if (entry
->flags
& __EXEC_OBJECT_HAS_FENCE
) {
1045 struct drm_i915_private
*dev_priv
= to_i915(ring
->dev
);
1046 list_move_tail(&dev_priv
->fence_regs
[obj
->fence_reg
].lru_list
,
1047 &dev_priv
->mm
.fence_list
);
1051 trace_i915_gem_object_change_domain(obj
, old_read
, old_write
);
1056 i915_gem_execbuffer_retire_commands(struct drm_device
*dev
,
1057 struct drm_file
*file
,
1058 struct intel_engine_cs
*ring
,
1059 struct drm_i915_gem_object
*obj
)
1061 /* Unconditionally force add_request to emit a full flush. */
1062 ring
->gpu_caches_dirty
= true;
1064 /* Add a breadcrumb for the completion of the batch buffer */
1065 (void)__i915_add_request(ring
, file
, obj
);
1069 i915_reset_gen7_sol_offsets(struct drm_device
*dev
,
1070 struct intel_engine_cs
*ring
)
1072 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1075 if (!IS_GEN7(dev
) || ring
!= &dev_priv
->ring
[RCS
]) {
1076 DRM_DEBUG("sol reset is gen7/rcs only\n");
1080 ret
= intel_ring_begin(ring
, 4 * 3);
1084 for (i
= 0; i
< 4; i
++) {
1085 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
1086 intel_ring_emit(ring
, GEN7_SO_WRITE_OFFSET(i
));
1087 intel_ring_emit(ring
, 0);
1090 intel_ring_advance(ring
);
1096 i915_emit_box(struct intel_engine_cs
*ring
,
1097 struct drm_clip_rect
*box
,
1102 if (box
->y2
<= box
->y1
|| box
->x2
<= box
->x1
||
1103 box
->y2
<= 0 || box
->x2
<= 0) {
1104 DRM_ERROR("Bad box %d,%d..%d,%d\n",
1105 box
->x1
, box
->y1
, box
->x2
, box
->y2
);
1109 if (INTEL_INFO(ring
->dev
)->gen
>= 4) {
1110 ret
= intel_ring_begin(ring
, 4);
1114 intel_ring_emit(ring
, GFX_OP_DRAWRECT_INFO_I965
);
1115 intel_ring_emit(ring
, (box
->x1
& 0xffff) | box
->y1
<< 16);
1116 intel_ring_emit(ring
, ((box
->x2
- 1) & 0xffff) | (box
->y2
- 1) << 16);
1117 intel_ring_emit(ring
, DR4
);
1119 ret
= intel_ring_begin(ring
, 6);
1123 intel_ring_emit(ring
, GFX_OP_DRAWRECT_INFO
);
1124 intel_ring_emit(ring
, DR1
);
1125 intel_ring_emit(ring
, (box
->x1
& 0xffff) | box
->y1
<< 16);
1126 intel_ring_emit(ring
, ((box
->x2
- 1) & 0xffff) | (box
->y2
- 1) << 16);
1127 intel_ring_emit(ring
, DR4
);
1128 intel_ring_emit(ring
, 0);
1130 intel_ring_advance(ring
);
1135 static struct drm_i915_gem_object
*
1136 i915_gem_execbuffer_parse(struct intel_engine_cs
*ring
,
1137 struct drm_i915_gem_exec_object2
*shadow_exec_entry
,
1139 struct drm_i915_gem_object
*batch_obj
,
1140 u32 batch_start_offset
,
1144 struct drm_i915_gem_object
*shadow_batch_obj
;
1145 struct i915_vma
*vma
;
1148 shadow_batch_obj
= i915_gem_batch_pool_get(&ring
->batch_pool
,
1149 PAGE_ALIGN(batch_len
));
1150 if (IS_ERR(shadow_batch_obj
))
1151 return shadow_batch_obj
;
1153 ret
= i915_parse_cmds(ring
,
1162 ret
= i915_gem_obj_ggtt_pin(shadow_batch_obj
, 0, 0);
1166 i915_gem_object_unpin_pages(shadow_batch_obj
);
1168 memset(shadow_exec_entry
, 0, sizeof(*shadow_exec_entry
));
1170 vma
= i915_gem_obj_to_ggtt(shadow_batch_obj
);
1171 vma
->exec_entry
= shadow_exec_entry
;
1172 vma
->exec_entry
->flags
= __EXEC_OBJECT_HAS_PIN
;
1173 drm_gem_object_reference(&shadow_batch_obj
->base
);
1174 list_add_tail(&vma
->exec_list
, &eb
->vmas
);
1176 shadow_batch_obj
->base
.pending_read_domains
= I915_GEM_DOMAIN_COMMAND
;
1178 return shadow_batch_obj
;
1181 i915_gem_object_unpin_pages(shadow_batch_obj
);
1182 if (ret
== -EACCES
) /* unhandled chained batch */
1185 return ERR_PTR(ret
);
1189 i915_gem_ringbuffer_submission(struct drm_device
*dev
, struct drm_file
*file
,
1190 struct intel_engine_cs
*ring
,
1191 struct intel_context
*ctx
,
1192 struct drm_i915_gem_execbuffer2
*args
,
1193 struct list_head
*vmas
,
1194 struct drm_i915_gem_object
*batch_obj
,
1195 u64 exec_start
, u32 dispatch_flags
)
1197 struct drm_clip_rect
*cliprects
= NULL
;
1198 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1204 if (args
->num_cliprects
!= 0) {
1205 if (ring
!= &dev_priv
->ring
[RCS
]) {
1206 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1210 if (INTEL_INFO(dev
)->gen
>= 5) {
1211 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1215 if (args
->num_cliprects
> UINT_MAX
/ sizeof(*cliprects
)) {
1216 DRM_DEBUG("execbuf with %u cliprects\n",
1217 args
->num_cliprects
);
1221 cliprects
= kcalloc(args
->num_cliprects
,
1224 if (cliprects
== NULL
) {
1229 if (copy_from_user(cliprects
,
1230 to_user_ptr(args
->cliprects_ptr
),
1231 sizeof(*cliprects
)*args
->num_cliprects
)) {
1236 if (args
->DR4
== 0xffffffff) {
1237 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1241 if (args
->DR1
|| args
->DR4
|| args
->cliprects_ptr
) {
1242 DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
1247 ret
= i915_gem_execbuffer_move_to_gpu(ring
, vmas
);
1251 ret
= i915_switch_context(ring
, ctx
);
1255 WARN(ctx
->ppgtt
&& ctx
->ppgtt
->pd_dirty_rings
& (1<<ring
->id
),
1256 "%s didn't clear reload\n", ring
->name
);
1258 instp_mode
= args
->flags
& I915_EXEC_CONSTANTS_MASK
;
1259 instp_mask
= I915_EXEC_CONSTANTS_MASK
;
1260 switch (instp_mode
) {
1261 case I915_EXEC_CONSTANTS_REL_GENERAL
:
1262 case I915_EXEC_CONSTANTS_ABSOLUTE
:
1263 case I915_EXEC_CONSTANTS_REL_SURFACE
:
1264 if (instp_mode
!= 0 && ring
!= &dev_priv
->ring
[RCS
]) {
1265 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1270 if (instp_mode
!= dev_priv
->relative_constants_mode
) {
1271 if (INTEL_INFO(dev
)->gen
< 4) {
1272 DRM_DEBUG("no rel constants on pre-gen4\n");
1277 if (INTEL_INFO(dev
)->gen
> 5 &&
1278 instp_mode
== I915_EXEC_CONSTANTS_REL_SURFACE
) {
1279 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1284 /* The HW changed the meaning on this bit on gen6 */
1285 if (INTEL_INFO(dev
)->gen
>= 6)
1286 instp_mask
&= ~I915_EXEC_CONSTANTS_REL_SURFACE
;
1290 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode
);
1295 if (ring
== &dev_priv
->ring
[RCS
] &&
1296 instp_mode
!= dev_priv
->relative_constants_mode
) {
1297 ret
= intel_ring_begin(ring
, 4);
1301 intel_ring_emit(ring
, MI_NOOP
);
1302 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
1303 intel_ring_emit(ring
, INSTPM
);
1304 intel_ring_emit(ring
, instp_mask
<< 16 | instp_mode
);
1305 intel_ring_advance(ring
);
1307 dev_priv
->relative_constants_mode
= instp_mode
;
1310 if (args
->flags
& I915_EXEC_GEN7_SOL_RESET
) {
1311 ret
= i915_reset_gen7_sol_offsets(dev
, ring
);
1316 exec_len
= args
->batch_len
;
1318 for (i
= 0; i
< args
->num_cliprects
; i
++) {
1319 ret
= i915_emit_box(ring
, &cliprects
[i
],
1320 args
->DR1
, args
->DR4
);
1324 ret
= ring
->dispatch_execbuffer(ring
,
1325 exec_start
, exec_len
,
1331 ret
= ring
->dispatch_execbuffer(ring
,
1332 exec_start
, exec_len
,
1338 trace_i915_gem_ring_dispatch(intel_ring_get_request(ring
), dispatch_flags
);
1340 i915_gem_execbuffer_move_to_active(vmas
, ring
);
1341 i915_gem_execbuffer_retire_commands(dev
, file
, ring
, batch_obj
);
1349 * Find one BSD ring to dispatch the corresponding BSD command.
1350 * The Ring ID is returned.
1352 static int gen8_dispatch_bsd_ring(struct drm_device
*dev
,
1353 struct drm_file
*file
)
1355 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1356 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
1358 /* Check whether the file_priv is using one ring */
1359 if (file_priv
->bsd_ring
)
1360 return file_priv
->bsd_ring
->id
;
1362 /* If no, use the ping-pong mechanism to select one ring */
1365 mutex_lock(&dev
->struct_mutex
);
1366 if (dev_priv
->mm
.bsd_ring_dispatch_index
== 0) {
1368 dev_priv
->mm
.bsd_ring_dispatch_index
= 1;
1371 dev_priv
->mm
.bsd_ring_dispatch_index
= 0;
1373 file_priv
->bsd_ring
= &dev_priv
->ring
[ring_id
];
1374 mutex_unlock(&dev
->struct_mutex
);
1379 static struct drm_i915_gem_object
*
1380 eb_get_batch(struct eb_vmas
*eb
)
1382 struct i915_vma
*vma
= list_entry(eb
->vmas
.prev
, typeof(*vma
), exec_list
);
1385 * SNA is doing fancy tricks with compressing batch buffers, which leads
1386 * to negative relocation deltas. Usually that works out ok since the
1387 * relocate address is still positive, except when the batch is placed
1388 * very low in the GTT. Ensure this doesn't happen.
1390 * Note that actual hangs have only been observed on gen7, but for
1391 * paranoia do it everywhere.
1393 vma
->exec_entry
->flags
|= __EXEC_OBJECT_NEEDS_BIAS
;
1399 i915_gem_do_execbuffer(struct drm_device
*dev
, void *data
,
1400 struct drm_file
*file
,
1401 struct drm_i915_gem_execbuffer2
*args
,
1402 struct drm_i915_gem_exec_object2
*exec
)
1404 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1406 struct drm_i915_gem_object
*batch_obj
;
1407 struct drm_i915_gem_exec_object2 shadow_exec_entry
;
1408 struct intel_engine_cs
*ring
;
1409 struct intel_context
*ctx
;
1410 struct i915_address_space
*vm
;
1411 const u32 ctx_id
= i915_execbuffer2_get_context_id(*args
);
1412 u64 exec_start
= args
->batch_start_offset
;
1417 if (!i915_gem_check_execbuffer(args
))
1420 ret
= validate_exec_list(dev
, exec
, args
->buffer_count
);
1425 if (args
->flags
& I915_EXEC_SECURE
) {
1426 if (!file
->is_master
|| !capable(CAP_SYS_ADMIN
))
1429 dispatch_flags
|= I915_DISPATCH_SECURE
;
1431 if (args
->flags
& I915_EXEC_IS_PINNED
)
1432 dispatch_flags
|= I915_DISPATCH_PINNED
;
1434 if ((args
->flags
& I915_EXEC_RING_MASK
) > LAST_USER_RING
) {
1435 DRM_DEBUG("execbuf with unknown ring: %d\n",
1436 (int)(args
->flags
& I915_EXEC_RING_MASK
));
1440 if (((args
->flags
& I915_EXEC_RING_MASK
) != I915_EXEC_BSD
) &&
1441 ((args
->flags
& I915_EXEC_BSD_MASK
) != 0)) {
1442 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1443 "bsd dispatch flags: %d\n", (int)(args
->flags
));
1447 if ((args
->flags
& I915_EXEC_RING_MASK
) == I915_EXEC_DEFAULT
)
1448 ring
= &dev_priv
->ring
[RCS
];
1449 else if ((args
->flags
& I915_EXEC_RING_MASK
) == I915_EXEC_BSD
) {
1450 if (HAS_BSD2(dev
)) {
1453 switch (args
->flags
& I915_EXEC_BSD_MASK
) {
1454 case I915_EXEC_BSD_DEFAULT
:
1455 ring_id
= gen8_dispatch_bsd_ring(dev
, file
);
1456 ring
= &dev_priv
->ring
[ring_id
];
1458 case I915_EXEC_BSD_RING1
:
1459 ring
= &dev_priv
->ring
[VCS
];
1461 case I915_EXEC_BSD_RING2
:
1462 ring
= &dev_priv
->ring
[VCS2
];
1465 DRM_DEBUG("execbuf with unknown bsd ring: %d\n",
1466 (int)(args
->flags
& I915_EXEC_BSD_MASK
));
1470 ring
= &dev_priv
->ring
[VCS
];
1472 ring
= &dev_priv
->ring
[(args
->flags
& I915_EXEC_RING_MASK
) - 1];
1474 if (!intel_ring_initialized(ring
)) {
1475 DRM_DEBUG("execbuf with invalid ring: %d\n",
1476 (int)(args
->flags
& I915_EXEC_RING_MASK
));
1480 if (args
->buffer_count
< 1) {
1481 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1485 intel_runtime_pm_get(dev_priv
);
1487 ret
= i915_mutex_lock_interruptible(dev
);
1491 ctx
= i915_gem_validate_context(dev
, file
, ring
, ctx_id
);
1493 mutex_unlock(&dev
->struct_mutex
);
1498 i915_gem_context_reference(ctx
);
1501 vm
= &ctx
->ppgtt
->base
;
1503 vm
= &dev_priv
->gtt
.base
;
1505 eb
= eb_create(args
);
1507 i915_gem_context_unreference(ctx
);
1508 mutex_unlock(&dev
->struct_mutex
);
1513 /* Look up object handles */
1514 ret
= eb_lookup_vmas(eb
, exec
, args
, vm
, file
);
1518 /* take note of the batch buffer before we might reorder the lists */
1519 batch_obj
= eb_get_batch(eb
);
1521 /* Move the objects en-masse into the GTT, evicting if necessary. */
1522 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
1523 ret
= i915_gem_execbuffer_reserve(ring
, &eb
->vmas
, &need_relocs
);
1527 /* The objects are in their final locations, apply the relocations. */
1529 ret
= i915_gem_execbuffer_relocate(eb
);
1531 if (ret
== -EFAULT
) {
1532 ret
= i915_gem_execbuffer_relocate_slow(dev
, args
, file
, ring
,
1534 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
1540 /* Set the pending read domains for the batch buffer to COMMAND */
1541 if (batch_obj
->base
.pending_write_domain
) {
1542 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1547 if (i915_needs_cmd_parser(ring
) && args
->batch_len
) {
1548 struct drm_i915_gem_object
*parsed_batch_obj
;
1550 parsed_batch_obj
= i915_gem_execbuffer_parse(ring
,
1554 args
->batch_start_offset
,
1557 if (IS_ERR(parsed_batch_obj
)) {
1558 ret
= PTR_ERR(parsed_batch_obj
);
1563 * parsed_batch_obj == batch_obj means batch not fully parsed:
1564 * Accept, but don't promote to secure.
1567 if (parsed_batch_obj
!= batch_obj
) {
1569 * Batch parsed and accepted:
1571 * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1572 * bit from MI_BATCH_BUFFER_START commands issued in
1573 * the dispatch_execbuffer implementations. We
1574 * specifically don't want that set on batches the
1575 * command parser has accepted.
1577 dispatch_flags
|= I915_DISPATCH_SECURE
;
1579 batch_obj
= parsed_batch_obj
;
1583 batch_obj
->base
.pending_read_domains
|= I915_GEM_DOMAIN_COMMAND
;
1585 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1586 * batch" bit. Hence we need to pin secure batches into the global gtt.
1587 * hsw should have this fixed, but bdw mucks it up again. */
1588 if (dispatch_flags
& I915_DISPATCH_SECURE
) {
1590 * So on first glance it looks freaky that we pin the batch here
1591 * outside of the reservation loop. But:
1592 * - The batch is already pinned into the relevant ppgtt, so we
1593 * already have the backing storage fully allocated.
1594 * - No other BO uses the global gtt (well contexts, but meh),
1595 * so we don't really have issues with multiple objects not
1596 * fitting due to fragmentation.
1597 * So this is actually safe.
1599 ret
= i915_gem_obj_ggtt_pin(batch_obj
, 0, 0);
1603 exec_start
+= i915_gem_obj_ggtt_offset(batch_obj
);
1605 exec_start
+= i915_gem_obj_offset(batch_obj
, vm
);
1607 ret
= dev_priv
->gt
.execbuf_submit(dev
, file
, ring
, ctx
, args
,
1608 &eb
->vmas
, batch_obj
, exec_start
,
1612 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1613 * batch vma for correctness. For less ugly and less fragility this
1614 * needs to be adjusted to also track the ggtt batch vma properly as
1617 if (dispatch_flags
& I915_DISPATCH_SECURE
)
1618 i915_gem_object_ggtt_unpin(batch_obj
);
1620 /* the request owns the ref now */
1621 i915_gem_context_unreference(ctx
);
1624 mutex_unlock(&dev
->struct_mutex
);
1627 /* intel_gpu_busy should also get a ref, so it will free when the device
1628 * is really idle. */
1629 intel_runtime_pm_put(dev_priv
);
1634 * Legacy execbuffer just creates an exec2 list from the original exec object
1635 * list array and passes it to the real function.
1638 i915_gem_execbuffer(struct drm_device
*dev
, void *data
,
1639 struct drm_file
*file
)
1641 struct drm_i915_gem_execbuffer
*args
= data
;
1642 struct drm_i915_gem_execbuffer2 exec2
;
1643 struct drm_i915_gem_exec_object
*exec_list
= NULL
;
1644 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1647 if (args
->buffer_count
< 1) {
1648 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1652 /* Copy in the exec list from userland */
1653 exec_list
= drm_malloc_ab(sizeof(*exec_list
), args
->buffer_count
);
1654 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
), args
->buffer_count
);
1655 if (exec_list
== NULL
|| exec2_list
== NULL
) {
1656 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1657 args
->buffer_count
);
1658 drm_free_large(exec_list
);
1659 drm_free_large(exec2_list
);
1662 ret
= copy_from_user(exec_list
,
1663 to_user_ptr(args
->buffers_ptr
),
1664 sizeof(*exec_list
) * args
->buffer_count
);
1666 DRM_DEBUG("copy %d exec entries failed %d\n",
1667 args
->buffer_count
, ret
);
1668 drm_free_large(exec_list
);
1669 drm_free_large(exec2_list
);
1673 for (i
= 0; i
< args
->buffer_count
; i
++) {
1674 exec2_list
[i
].handle
= exec_list
[i
].handle
;
1675 exec2_list
[i
].relocation_count
= exec_list
[i
].relocation_count
;
1676 exec2_list
[i
].relocs_ptr
= exec_list
[i
].relocs_ptr
;
1677 exec2_list
[i
].alignment
= exec_list
[i
].alignment
;
1678 exec2_list
[i
].offset
= exec_list
[i
].offset
;
1679 if (INTEL_INFO(dev
)->gen
< 4)
1680 exec2_list
[i
].flags
= EXEC_OBJECT_NEEDS_FENCE
;
1682 exec2_list
[i
].flags
= 0;
1685 exec2
.buffers_ptr
= args
->buffers_ptr
;
1686 exec2
.buffer_count
= args
->buffer_count
;
1687 exec2
.batch_start_offset
= args
->batch_start_offset
;
1688 exec2
.batch_len
= args
->batch_len
;
1689 exec2
.DR1
= args
->DR1
;
1690 exec2
.DR4
= args
->DR4
;
1691 exec2
.num_cliprects
= args
->num_cliprects
;
1692 exec2
.cliprects_ptr
= args
->cliprects_ptr
;
1693 exec2
.flags
= I915_EXEC_RENDER
;
1694 i915_execbuffer2_set_context_id(exec2
, 0);
1696 ret
= i915_gem_do_execbuffer(dev
, data
, file
, &exec2
, exec2_list
);
1698 struct drm_i915_gem_exec_object __user
*user_exec_list
=
1699 to_user_ptr(args
->buffers_ptr
);
1701 /* Copy the new buffer offsets back to the user's exec list. */
1702 for (i
= 0; i
< args
->buffer_count
; i
++) {
1703 ret
= __copy_to_user(&user_exec_list
[i
].offset
,
1704 &exec2_list
[i
].offset
,
1705 sizeof(user_exec_list
[i
].offset
));
1708 DRM_DEBUG("failed to copy %d exec entries "
1709 "back to user (%d)\n",
1710 args
->buffer_count
, ret
);
1716 drm_free_large(exec_list
);
1717 drm_free_large(exec2_list
);
1722 i915_gem_execbuffer2(struct drm_device
*dev
, void *data
,
1723 struct drm_file
*file
)
1725 struct drm_i915_gem_execbuffer2
*args
= data
;
1726 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1729 if (args
->buffer_count
< 1 ||
1730 args
->buffer_count
> UINT_MAX
/ sizeof(*exec2_list
)) {
1731 DRM_DEBUG("execbuf2 with %d buffers\n", args
->buffer_count
);
1735 if (args
->rsvd2
!= 0) {
1736 DRM_DEBUG("dirty rvsd2 field\n");
1740 exec2_list
= kmalloc(sizeof(*exec2_list
)*args
->buffer_count
,
1741 GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
1742 if (exec2_list
== NULL
)
1743 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
),
1744 args
->buffer_count
);
1745 if (exec2_list
== NULL
) {
1746 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1747 args
->buffer_count
);
1750 ret
= copy_from_user(exec2_list
,
1751 to_user_ptr(args
->buffers_ptr
),
1752 sizeof(*exec2_list
) * args
->buffer_count
);
1754 DRM_DEBUG("copy %d exec entries failed %d\n",
1755 args
->buffer_count
, ret
);
1756 drm_free_large(exec2_list
);
1760 ret
= i915_gem_do_execbuffer(dev
, data
, file
, args
, exec2_list
);
1762 /* Copy the new buffer offsets back to the user's exec list. */
1763 struct drm_i915_gem_exec_object2 __user
*user_exec_list
=
1764 to_user_ptr(args
->buffers_ptr
);
1767 for (i
= 0; i
< args
->buffer_count
; i
++) {
1768 ret
= __copy_to_user(&user_exec_list
[i
].offset
,
1769 &exec2_list
[i
].offset
,
1770 sizeof(user_exec_list
[i
].offset
));
1773 DRM_DEBUG("failed to copy %d exec entries "
1775 args
->buffer_count
);
1781 drm_free_large(exec2_list
);