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>
37 struct list_head objects
;
40 struct drm_i915_gem_object
*lut
[0];
41 struct hlist_head buckets
[0];
45 static struct eb_objects
*
46 eb_create(struct drm_i915_gem_execbuffer2
*args
)
48 struct eb_objects
*eb
= NULL
;
50 if (args
->flags
& I915_EXEC_HANDLE_LUT
) {
51 int size
= args
->buffer_count
;
52 size
*= sizeof(struct drm_i915_gem_object
*);
53 size
+= sizeof(struct eb_objects
);
54 eb
= kmalloc(size
, GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
58 int size
= args
->buffer_count
;
59 int count
= PAGE_SIZE
/ sizeof(struct hlist_head
) / 2;
60 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE
/ sizeof(struct hlist_head
));
61 while (count
> 2*size
)
63 eb
= kzalloc(count
*sizeof(struct hlist_head
) +
64 sizeof(struct eb_objects
),
71 eb
->and = -args
->buffer_count
;
73 INIT_LIST_HEAD(&eb
->objects
);
78 eb_reset(struct eb_objects
*eb
)
81 memset(eb
->buckets
, 0, (eb
->and+1)*sizeof(struct hlist_head
));
85 eb_lookup_objects(struct eb_objects
*eb
,
86 struct drm_i915_gem_exec_object2
*exec
,
87 const struct drm_i915_gem_execbuffer2
*args
,
88 struct drm_file
*file
)
92 spin_lock(&file
->table_lock
);
93 for (i
= 0; i
< args
->buffer_count
; i
++) {
94 struct drm_i915_gem_object
*obj
;
96 obj
= to_intel_bo(idr_find(&file
->object_idr
, exec
[i
].handle
));
98 spin_unlock(&file
->table_lock
);
99 DRM_DEBUG("Invalid object handle %d at index %d\n",
104 if (!list_empty(&obj
->exec_list
)) {
105 spin_unlock(&file
->table_lock
);
106 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
107 obj
, exec
[i
].handle
, i
);
111 drm_gem_object_reference(&obj
->base
);
112 list_add_tail(&obj
->exec_list
, &eb
->objects
);
114 obj
->exec_entry
= &exec
[i
];
118 uint32_t handle
= args
->flags
& I915_EXEC_HANDLE_LUT
? i
: exec
[i
].handle
;
119 obj
->exec_handle
= handle
;
120 hlist_add_head(&obj
->exec_node
,
121 &eb
->buckets
[handle
& eb
->and]);
124 spin_unlock(&file
->table_lock
);
129 static struct drm_i915_gem_object
*
130 eb_get_object(struct eb_objects
*eb
, unsigned long handle
)
133 if (handle
>= -eb
->and)
135 return eb
->lut
[handle
];
137 struct hlist_head
*head
;
138 struct hlist_node
*node
;
140 head
= &eb
->buckets
[handle
& eb
->and];
141 hlist_for_each(node
, head
) {
142 struct drm_i915_gem_object
*obj
;
144 obj
= hlist_entry(node
, struct drm_i915_gem_object
, exec_node
);
145 if (obj
->exec_handle
== handle
)
153 eb_destroy(struct eb_objects
*eb
)
155 while (!list_empty(&eb
->objects
)) {
156 struct drm_i915_gem_object
*obj
;
158 obj
= list_first_entry(&eb
->objects
,
159 struct drm_i915_gem_object
,
161 list_del_init(&obj
->exec_list
);
162 drm_gem_object_unreference(&obj
->base
);
167 static inline int use_cpu_reloc(struct drm_i915_gem_object
*obj
)
169 return (obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
||
170 !obj
->map_and_fenceable
||
171 obj
->cache_level
!= I915_CACHE_NONE
);
175 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object
*obj
,
176 struct eb_objects
*eb
,
177 struct drm_i915_gem_relocation_entry
*reloc
,
178 struct i915_address_space
*vm
)
180 struct drm_device
*dev
= obj
->base
.dev
;
181 struct drm_gem_object
*target_obj
;
182 struct drm_i915_gem_object
*target_i915_obj
;
183 uint32_t target_offset
;
186 /* we've already hold a reference to all valid objects */
187 target_obj
= &eb_get_object(eb
, reloc
->target_handle
)->base
;
188 if (unlikely(target_obj
== NULL
))
191 target_i915_obj
= to_intel_bo(target_obj
);
192 target_offset
= i915_gem_obj_ggtt_offset(target_i915_obj
);
194 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
195 * pipe_control writes because the gpu doesn't properly redirect them
196 * through the ppgtt for non_secure batchbuffers. */
197 if (unlikely(IS_GEN6(dev
) &&
198 reloc
->write_domain
== I915_GEM_DOMAIN_INSTRUCTION
&&
199 !target_i915_obj
->has_global_gtt_mapping
)) {
200 i915_gem_gtt_bind_object(target_i915_obj
,
201 target_i915_obj
->cache_level
);
204 /* Validate that the target is in a valid r/w GPU domain */
205 if (unlikely(reloc
->write_domain
& (reloc
->write_domain
- 1))) {
206 DRM_DEBUG("reloc with multiple write domains: "
207 "obj %p target %d offset %d "
208 "read %08x write %08x",
209 obj
, reloc
->target_handle
,
212 reloc
->write_domain
);
215 if (unlikely((reloc
->write_domain
| reloc
->read_domains
)
216 & ~I915_GEM_GPU_DOMAINS
)) {
217 DRM_DEBUG("reloc with read/write non-GPU domains: "
218 "obj %p target %d offset %d "
219 "read %08x write %08x",
220 obj
, reloc
->target_handle
,
223 reloc
->write_domain
);
227 target_obj
->pending_read_domains
|= reloc
->read_domains
;
228 target_obj
->pending_write_domain
|= reloc
->write_domain
;
230 /* If the relocation already has the right value in it, no
231 * more work needs to be done.
233 if (target_offset
== reloc
->presumed_offset
)
236 /* Check that the relocation address is valid... */
237 if (unlikely(reloc
->offset
> obj
->base
.size
- 4)) {
238 DRM_DEBUG("Relocation beyond object bounds: "
239 "obj %p target %d offset %d size %d.\n",
240 obj
, reloc
->target_handle
,
242 (int) obj
->base
.size
);
245 if (unlikely(reloc
->offset
& 3)) {
246 DRM_DEBUG("Relocation not 4-byte aligned: "
247 "obj %p target %d offset %d.\n",
248 obj
, reloc
->target_handle
,
249 (int) reloc
->offset
);
253 /* We can't wait for rendering with pagefaults disabled */
254 if (obj
->active
&& in_atomic())
257 reloc
->delta
+= target_offset
;
258 if (use_cpu_reloc(obj
)) {
259 uint32_t page_offset
= offset_in_page(reloc
->offset
);
262 ret
= i915_gem_object_set_to_cpu_domain(obj
, 1);
266 vaddr
= kmap_atomic(i915_gem_object_get_page(obj
,
267 reloc
->offset
>> PAGE_SHIFT
));
268 *(uint32_t *)(vaddr
+ page_offset
) = reloc
->delta
;
269 kunmap_atomic(vaddr
);
271 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
272 uint32_t __iomem
*reloc_entry
;
273 void __iomem
*reloc_page
;
275 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
279 ret
= i915_gem_object_put_fence(obj
);
283 /* Map the page containing the relocation we're going to perform. */
284 reloc
->offset
+= i915_gem_obj_ggtt_offset(obj
);
285 reloc_page
= io_mapping_map_atomic_wc(dev_priv
->gtt
.mappable
,
286 reloc
->offset
& PAGE_MASK
);
287 reloc_entry
= (uint32_t __iomem
*)
288 (reloc_page
+ offset_in_page(reloc
->offset
));
289 iowrite32(reloc
->delta
, reloc_entry
);
290 io_mapping_unmap_atomic(reloc_page
);
293 /* and update the user's relocation entry */
294 reloc
->presumed_offset
= target_offset
;
300 i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object
*obj
,
301 struct eb_objects
*eb
,
302 struct i915_address_space
*vm
)
304 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
305 struct drm_i915_gem_relocation_entry stack_reloc
[N_RELOC(512)];
306 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
307 struct drm_i915_gem_exec_object2
*entry
= obj
->exec_entry
;
310 user_relocs
= to_user_ptr(entry
->relocs_ptr
);
312 remain
= entry
->relocation_count
;
314 struct drm_i915_gem_relocation_entry
*r
= stack_reloc
;
316 if (count
> ARRAY_SIZE(stack_reloc
))
317 count
= ARRAY_SIZE(stack_reloc
);
320 if (__copy_from_user_inatomic(r
, user_relocs
, count
*sizeof(r
[0])))
324 u64 offset
= r
->presumed_offset
;
326 ret
= i915_gem_execbuffer_relocate_entry(obj
, eb
, r
,
331 if (r
->presumed_offset
!= offset
&&
332 __copy_to_user_inatomic(&user_relocs
->presumed_offset
,
334 sizeof(r
->presumed_offset
))) {
348 i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object
*obj
,
349 struct eb_objects
*eb
,
350 struct drm_i915_gem_relocation_entry
*relocs
,
351 struct i915_address_space
*vm
)
353 const struct drm_i915_gem_exec_object2
*entry
= obj
->exec_entry
;
356 for (i
= 0; i
< entry
->relocation_count
; i
++) {
357 ret
= i915_gem_execbuffer_relocate_entry(obj
, eb
, &relocs
[i
],
367 i915_gem_execbuffer_relocate(struct eb_objects
*eb
,
368 struct i915_address_space
*vm
)
370 struct drm_i915_gem_object
*obj
;
373 /* This is the fast path and we cannot handle a pagefault whilst
374 * holding the struct mutex lest the user pass in the relocations
375 * contained within a mmaped bo. For in such a case we, the page
376 * fault handler would call i915_gem_fault() and we would try to
377 * acquire the struct mutex again. Obviously this is bad and so
378 * lockdep complains vehemently.
381 list_for_each_entry(obj
, &eb
->objects
, exec_list
) {
382 ret
= i915_gem_execbuffer_relocate_object(obj
, eb
, vm
);
391 #define __EXEC_OBJECT_HAS_PIN (1<<31)
392 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
395 need_reloc_mappable(struct drm_i915_gem_object
*obj
)
397 struct drm_i915_gem_exec_object2
*entry
= obj
->exec_entry
;
398 return entry
->relocation_count
&& !use_cpu_reloc(obj
);
402 i915_gem_execbuffer_reserve_object(struct drm_i915_gem_object
*obj
,
403 struct intel_ring_buffer
*ring
,
404 struct i915_address_space
*vm
,
407 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
408 struct drm_i915_gem_exec_object2
*entry
= obj
->exec_entry
;
409 bool has_fenced_gpu_access
= INTEL_INFO(ring
->dev
)->gen
< 4;
410 bool need_fence
, need_mappable
;
414 has_fenced_gpu_access
&&
415 entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
&&
416 obj
->tiling_mode
!= I915_TILING_NONE
;
417 need_mappable
= need_fence
|| need_reloc_mappable(obj
);
419 ret
= i915_gem_object_pin(obj
, vm
, entry
->alignment
, need_mappable
,
424 entry
->flags
|= __EXEC_OBJECT_HAS_PIN
;
426 if (has_fenced_gpu_access
) {
427 if (entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
) {
428 ret
= i915_gem_object_get_fence(obj
);
432 if (i915_gem_object_pin_fence(obj
))
433 entry
->flags
|= __EXEC_OBJECT_HAS_FENCE
;
435 obj
->pending_fenced_gpu_access
= true;
439 /* Ensure ppgtt mapping exists if needed */
440 if (dev_priv
->mm
.aliasing_ppgtt
&& !obj
->has_aliasing_ppgtt_mapping
) {
441 i915_ppgtt_bind_object(dev_priv
->mm
.aliasing_ppgtt
,
442 obj
, obj
->cache_level
);
444 obj
->has_aliasing_ppgtt_mapping
= 1;
447 if (entry
->offset
!= i915_gem_obj_offset(obj
, vm
)) {
448 entry
->offset
= i915_gem_obj_offset(obj
, vm
);
452 if (entry
->flags
& EXEC_OBJECT_WRITE
) {
453 obj
->base
.pending_read_domains
= I915_GEM_DOMAIN_RENDER
;
454 obj
->base
.pending_write_domain
= I915_GEM_DOMAIN_RENDER
;
457 if (entry
->flags
& EXEC_OBJECT_NEEDS_GTT
&&
458 !obj
->has_global_gtt_mapping
)
459 i915_gem_gtt_bind_object(obj
, obj
->cache_level
);
465 i915_gem_execbuffer_unreserve_object(struct drm_i915_gem_object
*obj
)
467 struct drm_i915_gem_exec_object2
*entry
;
469 if (!i915_gem_obj_bound_any(obj
))
472 entry
= obj
->exec_entry
;
474 if (entry
->flags
& __EXEC_OBJECT_HAS_FENCE
)
475 i915_gem_object_unpin_fence(obj
);
477 if (entry
->flags
& __EXEC_OBJECT_HAS_PIN
)
478 i915_gem_object_unpin(obj
);
480 entry
->flags
&= ~(__EXEC_OBJECT_HAS_FENCE
| __EXEC_OBJECT_HAS_PIN
);
484 i915_gem_execbuffer_reserve(struct intel_ring_buffer
*ring
,
485 struct list_head
*objects
,
486 struct i915_address_space
*vm
,
489 struct drm_i915_gem_object
*obj
;
490 struct list_head ordered_objects
;
491 bool has_fenced_gpu_access
= INTEL_INFO(ring
->dev
)->gen
< 4;
494 INIT_LIST_HEAD(&ordered_objects
);
495 while (!list_empty(objects
)) {
496 struct drm_i915_gem_exec_object2
*entry
;
497 bool need_fence
, need_mappable
;
499 obj
= list_first_entry(objects
,
500 struct drm_i915_gem_object
,
502 entry
= obj
->exec_entry
;
505 has_fenced_gpu_access
&&
506 entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
&&
507 obj
->tiling_mode
!= I915_TILING_NONE
;
508 need_mappable
= need_fence
|| need_reloc_mappable(obj
);
511 list_move(&obj
->exec_list
, &ordered_objects
);
513 list_move_tail(&obj
->exec_list
, &ordered_objects
);
515 obj
->base
.pending_read_domains
= I915_GEM_GPU_DOMAINS
& ~I915_GEM_DOMAIN_COMMAND
;
516 obj
->base
.pending_write_domain
= 0;
517 obj
->pending_fenced_gpu_access
= false;
519 list_splice(&ordered_objects
, objects
);
521 /* Attempt to pin all of the buffers into the GTT.
522 * This is done in 3 phases:
524 * 1a. Unbind all objects that do not match the GTT constraints for
525 * the execbuffer (fenceable, mappable, alignment etc).
526 * 1b. Increment pin count for already bound objects.
527 * 2. Bind new objects.
528 * 3. Decrement pin count.
530 * This avoid unnecessary unbinding of later objects in order to make
531 * room for the earlier objects *unless* we need to defragment.
537 /* Unbind any ill-fitting objects or pin. */
538 list_for_each_entry(obj
, objects
, exec_list
) {
539 struct drm_i915_gem_exec_object2
*entry
= obj
->exec_entry
;
540 bool need_fence
, need_mappable
;
543 if (!i915_gem_obj_bound(obj
, vm
))
546 obj_offset
= i915_gem_obj_offset(obj
, vm
);
548 has_fenced_gpu_access
&&
549 entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
&&
550 obj
->tiling_mode
!= I915_TILING_NONE
;
551 need_mappable
= need_fence
|| need_reloc_mappable(obj
);
553 WARN_ON((need_mappable
|| need_fence
) &&
556 if ((entry
->alignment
&&
557 obj_offset
& (entry
->alignment
- 1)) ||
558 (need_mappable
&& !obj
->map_and_fenceable
))
559 ret
= i915_vma_unbind(i915_gem_obj_to_vma(obj
, vm
));
561 ret
= i915_gem_execbuffer_reserve_object(obj
, ring
, vm
, need_relocs
);
566 /* Bind fresh objects */
567 list_for_each_entry(obj
, objects
, exec_list
) {
568 if (i915_gem_obj_bound(obj
, vm
))
571 ret
= i915_gem_execbuffer_reserve_object(obj
, ring
, vm
, need_relocs
);
576 err
: /* Decrement pin count for bound objects */
577 list_for_each_entry(obj
, objects
, exec_list
)
578 i915_gem_execbuffer_unreserve_object(obj
);
580 if (ret
!= -ENOSPC
|| retry
++)
583 ret
= i915_gem_evict_everything(ring
->dev
);
590 i915_gem_execbuffer_relocate_slow(struct drm_device
*dev
,
591 struct drm_i915_gem_execbuffer2
*args
,
592 struct drm_file
*file
,
593 struct intel_ring_buffer
*ring
,
594 struct eb_objects
*eb
,
595 struct drm_i915_gem_exec_object2
*exec
,
596 struct i915_address_space
*vm
)
598 struct drm_i915_gem_relocation_entry
*reloc
;
599 struct drm_i915_gem_object
*obj
;
603 int count
= args
->buffer_count
;
605 /* We may process another execbuffer during the unlock... */
606 while (!list_empty(&eb
->objects
)) {
607 obj
= list_first_entry(&eb
->objects
,
608 struct drm_i915_gem_object
,
610 list_del_init(&obj
->exec_list
);
611 drm_gem_object_unreference(&obj
->base
);
614 mutex_unlock(&dev
->struct_mutex
);
617 for (i
= 0; i
< count
; i
++)
618 total
+= exec
[i
].relocation_count
;
620 reloc_offset
= drm_malloc_ab(count
, sizeof(*reloc_offset
));
621 reloc
= drm_malloc_ab(total
, sizeof(*reloc
));
622 if (reloc
== NULL
|| reloc_offset
== NULL
) {
623 drm_free_large(reloc
);
624 drm_free_large(reloc_offset
);
625 mutex_lock(&dev
->struct_mutex
);
630 for (i
= 0; i
< count
; i
++) {
631 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
632 u64 invalid_offset
= (u64
)-1;
635 user_relocs
= to_user_ptr(exec
[i
].relocs_ptr
);
637 if (copy_from_user(reloc
+total
, user_relocs
,
638 exec
[i
].relocation_count
* sizeof(*reloc
))) {
640 mutex_lock(&dev
->struct_mutex
);
644 /* As we do not update the known relocation offsets after
645 * relocating (due to the complexities in lock handling),
646 * we need to mark them as invalid now so that we force the
647 * relocation processing next time. Just in case the target
648 * object is evicted and then rebound into its old
649 * presumed_offset before the next execbuffer - if that
650 * happened we would make the mistake of assuming that the
651 * relocations were valid.
653 for (j
= 0; j
< exec
[i
].relocation_count
; j
++) {
654 if (copy_to_user(&user_relocs
[j
].presumed_offset
,
656 sizeof(invalid_offset
))) {
658 mutex_lock(&dev
->struct_mutex
);
663 reloc_offset
[i
] = total
;
664 total
+= exec
[i
].relocation_count
;
667 ret
= i915_mutex_lock_interruptible(dev
);
669 mutex_lock(&dev
->struct_mutex
);
673 /* reacquire the objects */
675 ret
= eb_lookup_objects(eb
, exec
, args
, file
);
679 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
680 ret
= i915_gem_execbuffer_reserve(ring
, &eb
->objects
, vm
, &need_relocs
);
684 list_for_each_entry(obj
, &eb
->objects
, exec_list
) {
685 int offset
= obj
->exec_entry
- exec
;
686 ret
= i915_gem_execbuffer_relocate_object_slow(obj
, eb
,
687 reloc
+ reloc_offset
[offset
],
693 /* Leave the user relocations as are, this is the painfully slow path,
694 * and we want to avoid the complication of dropping the lock whilst
695 * having buffers reserved in the aperture and so causing spurious
696 * ENOSPC for random operations.
700 drm_free_large(reloc
);
701 drm_free_large(reloc_offset
);
706 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer
*ring
,
707 struct list_head
*objects
)
709 struct drm_i915_gem_object
*obj
;
710 uint32_t flush_domains
= 0;
711 bool flush_chipset
= false;
714 list_for_each_entry(obj
, objects
, exec_list
) {
715 ret
= i915_gem_object_sync(obj
, ring
);
719 if (obj
->base
.write_domain
& I915_GEM_DOMAIN_CPU
)
720 flush_chipset
|= i915_gem_clflush_object(obj
, false);
722 flush_domains
|= obj
->base
.write_domain
;
726 i915_gem_chipset_flush(ring
->dev
);
728 if (flush_domains
& I915_GEM_DOMAIN_GTT
)
731 /* Unconditionally invalidate gpu caches and ensure that we do flush
732 * any residual writes from the previous batch.
734 return intel_ring_invalidate_all_caches(ring
);
738 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2
*exec
)
740 if (exec
->flags
& __I915_EXEC_UNKNOWN_FLAGS
)
743 return ((exec
->batch_start_offset
| exec
->batch_len
) & 0x7) == 0;
747 validate_exec_list(struct drm_i915_gem_exec_object2
*exec
,
751 int relocs_total
= 0;
752 int relocs_max
= INT_MAX
/ sizeof(struct drm_i915_gem_relocation_entry
);
754 for (i
= 0; i
< count
; i
++) {
755 char __user
*ptr
= to_user_ptr(exec
[i
].relocs_ptr
);
756 int length
; /* limited by fault_in_pages_readable() */
758 if (exec
[i
].flags
& __EXEC_OBJECT_UNKNOWN_FLAGS
)
761 /* First check for malicious input causing overflow in
762 * the worst case where we need to allocate the entire
763 * relocation tree as a single array.
765 if (exec
[i
].relocation_count
> relocs_max
- relocs_total
)
767 relocs_total
+= exec
[i
].relocation_count
;
769 length
= exec
[i
].relocation_count
*
770 sizeof(struct drm_i915_gem_relocation_entry
);
772 * We must check that the entire relocation array is safe
773 * to read, but since we may need to update the presumed
774 * offsets during execution, check for full write access.
776 if (!access_ok(VERIFY_WRITE
, ptr
, length
))
779 if (likely(!i915_prefault_disable
)) {
780 if (fault_in_multipages_readable(ptr
, length
))
789 i915_gem_execbuffer_move_to_active(struct list_head
*objects
,
790 struct i915_address_space
*vm
,
791 struct intel_ring_buffer
*ring
)
793 struct drm_i915_gem_object
*obj
;
795 list_for_each_entry(obj
, objects
, exec_list
) {
796 u32 old_read
= obj
->base
.read_domains
;
797 u32 old_write
= obj
->base
.write_domain
;
799 obj
->base
.write_domain
= obj
->base
.pending_write_domain
;
800 if (obj
->base
.write_domain
== 0)
801 obj
->base
.pending_read_domains
|= obj
->base
.read_domains
;
802 obj
->base
.read_domains
= obj
->base
.pending_read_domains
;
803 obj
->fenced_gpu_access
= obj
->pending_fenced_gpu_access
;
805 /* FIXME: This lookup gets fixed later <-- danvet */
806 list_move_tail(&i915_gem_obj_to_vma(obj
, vm
)->mm_list
, &vm
->active_list
);
807 i915_gem_object_move_to_active(obj
, ring
);
808 if (obj
->base
.write_domain
) {
810 obj
->last_write_seqno
= intel_ring_get_seqno(ring
);
811 if (obj
->pin_count
) /* check for potential scanout */
812 intel_mark_fb_busy(obj
, ring
);
815 trace_i915_gem_object_change_domain(obj
, old_read
, old_write
);
820 i915_gem_execbuffer_retire_commands(struct drm_device
*dev
,
821 struct drm_file
*file
,
822 struct intel_ring_buffer
*ring
,
823 struct drm_i915_gem_object
*obj
)
825 /* Unconditionally force add_request to emit a full flush. */
826 ring
->gpu_caches_dirty
= true;
828 /* Add a breadcrumb for the completion of the batch buffer */
829 (void)__i915_add_request(ring
, file
, obj
, NULL
);
833 i915_reset_gen7_sol_offsets(struct drm_device
*dev
,
834 struct intel_ring_buffer
*ring
)
836 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
839 if (!IS_GEN7(dev
) || ring
!= &dev_priv
->ring
[RCS
])
842 ret
= intel_ring_begin(ring
, 4 * 3);
846 for (i
= 0; i
< 4; i
++) {
847 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
848 intel_ring_emit(ring
, GEN7_SO_WRITE_OFFSET(i
));
849 intel_ring_emit(ring
, 0);
852 intel_ring_advance(ring
);
858 i915_gem_do_execbuffer(struct drm_device
*dev
, void *data
,
859 struct drm_file
*file
,
860 struct drm_i915_gem_execbuffer2
*args
,
861 struct drm_i915_gem_exec_object2
*exec
,
862 struct i915_address_space
*vm
)
864 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
865 struct eb_objects
*eb
;
866 struct drm_i915_gem_object
*batch_obj
;
867 struct drm_clip_rect
*cliprects
= NULL
;
868 struct intel_ring_buffer
*ring
;
869 u32 ctx_id
= i915_execbuffer2_get_context_id(*args
);
870 u32 exec_start
, exec_len
;
875 if (!i915_gem_check_execbuffer(args
))
878 ret
= validate_exec_list(exec
, args
->buffer_count
);
883 if (args
->flags
& I915_EXEC_SECURE
) {
884 if (!file
->is_master
|| !capable(CAP_SYS_ADMIN
))
887 flags
|= I915_DISPATCH_SECURE
;
889 if (args
->flags
& I915_EXEC_IS_PINNED
)
890 flags
|= I915_DISPATCH_PINNED
;
892 switch (args
->flags
& I915_EXEC_RING_MASK
) {
893 case I915_EXEC_DEFAULT
:
894 case I915_EXEC_RENDER
:
895 ring
= &dev_priv
->ring
[RCS
];
898 ring
= &dev_priv
->ring
[VCS
];
899 if (ctx_id
!= DEFAULT_CONTEXT_ID
) {
900 DRM_DEBUG("Ring %s doesn't support contexts\n",
906 ring
= &dev_priv
->ring
[BCS
];
907 if (ctx_id
!= DEFAULT_CONTEXT_ID
) {
908 DRM_DEBUG("Ring %s doesn't support contexts\n",
913 case I915_EXEC_VEBOX
:
914 ring
= &dev_priv
->ring
[VECS
];
915 if (ctx_id
!= DEFAULT_CONTEXT_ID
) {
916 DRM_DEBUG("Ring %s doesn't support contexts\n",
923 DRM_DEBUG("execbuf with unknown ring: %d\n",
924 (int)(args
->flags
& I915_EXEC_RING_MASK
));
927 if (!intel_ring_initialized(ring
)) {
928 DRM_DEBUG("execbuf with invalid ring: %d\n",
929 (int)(args
->flags
& I915_EXEC_RING_MASK
));
933 mode
= args
->flags
& I915_EXEC_CONSTANTS_MASK
;
934 mask
= I915_EXEC_CONSTANTS_MASK
;
936 case I915_EXEC_CONSTANTS_REL_GENERAL
:
937 case I915_EXEC_CONSTANTS_ABSOLUTE
:
938 case I915_EXEC_CONSTANTS_REL_SURFACE
:
939 if (ring
== &dev_priv
->ring
[RCS
] &&
940 mode
!= dev_priv
->relative_constants_mode
) {
941 if (INTEL_INFO(dev
)->gen
< 4)
944 if (INTEL_INFO(dev
)->gen
> 5 &&
945 mode
== I915_EXEC_CONSTANTS_REL_SURFACE
)
948 /* The HW changed the meaning on this bit on gen6 */
949 if (INTEL_INFO(dev
)->gen
>= 6)
950 mask
&= ~I915_EXEC_CONSTANTS_REL_SURFACE
;
954 DRM_DEBUG("execbuf with unknown constants: %d\n", mode
);
958 if (args
->buffer_count
< 1) {
959 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
963 if (args
->num_cliprects
!= 0) {
964 if (ring
!= &dev_priv
->ring
[RCS
]) {
965 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
969 if (INTEL_INFO(dev
)->gen
>= 5) {
970 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
974 if (args
->num_cliprects
> UINT_MAX
/ sizeof(*cliprects
)) {
975 DRM_DEBUG("execbuf with %u cliprects\n",
976 args
->num_cliprects
);
980 cliprects
= kmalloc(args
->num_cliprects
* sizeof(*cliprects
),
982 if (cliprects
== NULL
) {
987 if (copy_from_user(cliprects
,
988 to_user_ptr(args
->cliprects_ptr
),
989 sizeof(*cliprects
)*args
->num_cliprects
)) {
995 ret
= i915_mutex_lock_interruptible(dev
);
999 if (dev_priv
->ums
.mm_suspended
) {
1000 mutex_unlock(&dev
->struct_mutex
);
1005 eb
= eb_create(args
);
1007 mutex_unlock(&dev
->struct_mutex
);
1012 /* Look up object handles */
1013 ret
= eb_lookup_objects(eb
, exec
, args
, file
);
1017 /* take note of the batch buffer before we might reorder the lists */
1018 batch_obj
= list_entry(eb
->objects
.prev
,
1019 struct drm_i915_gem_object
,
1022 /* Move the objects en-masse into the GTT, evicting if necessary. */
1023 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
1024 ret
= i915_gem_execbuffer_reserve(ring
, &eb
->objects
, vm
, &need_relocs
);
1028 /* The objects are in their final locations, apply the relocations. */
1030 ret
= i915_gem_execbuffer_relocate(eb
, vm
);
1032 if (ret
== -EFAULT
) {
1033 ret
= i915_gem_execbuffer_relocate_slow(dev
, args
, file
, ring
,
1035 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
1041 /* Set the pending read domains for the batch buffer to COMMAND */
1042 if (batch_obj
->base
.pending_write_domain
) {
1043 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1047 batch_obj
->base
.pending_read_domains
|= I915_GEM_DOMAIN_COMMAND
;
1049 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1050 * batch" bit. Hence we need to pin secure batches into the global gtt.
1051 * hsw should have this fixed, but let's be paranoid and do it
1052 * unconditionally for now. */
1053 if (flags
& I915_DISPATCH_SECURE
&& !batch_obj
->has_global_gtt_mapping
)
1054 i915_gem_gtt_bind_object(batch_obj
, batch_obj
->cache_level
);
1056 ret
= i915_gem_execbuffer_move_to_gpu(ring
, &eb
->objects
);
1060 ret
= i915_switch_context(ring
, file
, ctx_id
);
1064 if (ring
== &dev_priv
->ring
[RCS
] &&
1065 mode
!= dev_priv
->relative_constants_mode
) {
1066 ret
= intel_ring_begin(ring
, 4);
1070 intel_ring_emit(ring
, MI_NOOP
);
1071 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
1072 intel_ring_emit(ring
, INSTPM
);
1073 intel_ring_emit(ring
, mask
<< 16 | mode
);
1074 intel_ring_advance(ring
);
1076 dev_priv
->relative_constants_mode
= mode
;
1079 if (args
->flags
& I915_EXEC_GEN7_SOL_RESET
) {
1080 ret
= i915_reset_gen7_sol_offsets(dev
, ring
);
1085 exec_start
= i915_gem_obj_offset(batch_obj
, vm
) +
1086 args
->batch_start_offset
;
1087 exec_len
= args
->batch_len
;
1089 for (i
= 0; i
< args
->num_cliprects
; i
++) {
1090 ret
= i915_emit_box(dev
, &cliprects
[i
],
1091 args
->DR1
, args
->DR4
);
1095 ret
= ring
->dispatch_execbuffer(ring
,
1096 exec_start
, exec_len
,
1102 ret
= ring
->dispatch_execbuffer(ring
,
1103 exec_start
, exec_len
,
1109 trace_i915_gem_ring_dispatch(ring
, intel_ring_get_seqno(ring
), flags
);
1111 i915_gem_execbuffer_move_to_active(&eb
->objects
, vm
, ring
);
1112 i915_gem_execbuffer_retire_commands(dev
, file
, ring
, batch_obj
);
1117 mutex_unlock(&dev
->struct_mutex
);
1125 * Legacy execbuffer just creates an exec2 list from the original exec object
1126 * list array and passes it to the real function.
1129 i915_gem_execbuffer(struct drm_device
*dev
, void *data
,
1130 struct drm_file
*file
)
1132 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1133 struct drm_i915_gem_execbuffer
*args
= data
;
1134 struct drm_i915_gem_execbuffer2 exec2
;
1135 struct drm_i915_gem_exec_object
*exec_list
= NULL
;
1136 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1139 if (args
->buffer_count
< 1) {
1140 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1144 /* Copy in the exec list from userland */
1145 exec_list
= drm_malloc_ab(sizeof(*exec_list
), args
->buffer_count
);
1146 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
), args
->buffer_count
);
1147 if (exec_list
== NULL
|| exec2_list
== NULL
) {
1148 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1149 args
->buffer_count
);
1150 drm_free_large(exec_list
);
1151 drm_free_large(exec2_list
);
1154 ret
= copy_from_user(exec_list
,
1155 to_user_ptr(args
->buffers_ptr
),
1156 sizeof(*exec_list
) * args
->buffer_count
);
1158 DRM_DEBUG("copy %d exec entries failed %d\n",
1159 args
->buffer_count
, ret
);
1160 drm_free_large(exec_list
);
1161 drm_free_large(exec2_list
);
1165 for (i
= 0; i
< args
->buffer_count
; i
++) {
1166 exec2_list
[i
].handle
= exec_list
[i
].handle
;
1167 exec2_list
[i
].relocation_count
= exec_list
[i
].relocation_count
;
1168 exec2_list
[i
].relocs_ptr
= exec_list
[i
].relocs_ptr
;
1169 exec2_list
[i
].alignment
= exec_list
[i
].alignment
;
1170 exec2_list
[i
].offset
= exec_list
[i
].offset
;
1171 if (INTEL_INFO(dev
)->gen
< 4)
1172 exec2_list
[i
].flags
= EXEC_OBJECT_NEEDS_FENCE
;
1174 exec2_list
[i
].flags
= 0;
1177 exec2
.buffers_ptr
= args
->buffers_ptr
;
1178 exec2
.buffer_count
= args
->buffer_count
;
1179 exec2
.batch_start_offset
= args
->batch_start_offset
;
1180 exec2
.batch_len
= args
->batch_len
;
1181 exec2
.DR1
= args
->DR1
;
1182 exec2
.DR4
= args
->DR4
;
1183 exec2
.num_cliprects
= args
->num_cliprects
;
1184 exec2
.cliprects_ptr
= args
->cliprects_ptr
;
1185 exec2
.flags
= I915_EXEC_RENDER
;
1186 i915_execbuffer2_set_context_id(exec2
, 0);
1188 ret
= i915_gem_do_execbuffer(dev
, data
, file
, &exec2
, exec2_list
,
1189 &dev_priv
->gtt
.base
);
1191 /* Copy the new buffer offsets back to the user's exec list. */
1192 for (i
= 0; i
< args
->buffer_count
; i
++)
1193 exec_list
[i
].offset
= exec2_list
[i
].offset
;
1194 /* ... and back out to userspace */
1195 ret
= copy_to_user(to_user_ptr(args
->buffers_ptr
),
1197 sizeof(*exec_list
) * args
->buffer_count
);
1200 DRM_DEBUG("failed to copy %d exec entries "
1201 "back to user (%d)\n",
1202 args
->buffer_count
, ret
);
1206 drm_free_large(exec_list
);
1207 drm_free_large(exec2_list
);
1212 i915_gem_execbuffer2(struct drm_device
*dev
, void *data
,
1213 struct drm_file
*file
)
1215 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1216 struct drm_i915_gem_execbuffer2
*args
= data
;
1217 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1220 if (args
->buffer_count
< 1 ||
1221 args
->buffer_count
> UINT_MAX
/ sizeof(*exec2_list
)) {
1222 DRM_DEBUG("execbuf2 with %d buffers\n", args
->buffer_count
);
1226 exec2_list
= kmalloc(sizeof(*exec2_list
)*args
->buffer_count
,
1227 GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
1228 if (exec2_list
== NULL
)
1229 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
),
1230 args
->buffer_count
);
1231 if (exec2_list
== NULL
) {
1232 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1233 args
->buffer_count
);
1236 ret
= copy_from_user(exec2_list
,
1237 to_user_ptr(args
->buffers_ptr
),
1238 sizeof(*exec2_list
) * args
->buffer_count
);
1240 DRM_DEBUG("copy %d exec entries failed %d\n",
1241 args
->buffer_count
, ret
);
1242 drm_free_large(exec2_list
);
1246 ret
= i915_gem_do_execbuffer(dev
, data
, file
, args
, exec2_list
,
1247 &dev_priv
->gtt
.base
);
1249 /* Copy the new buffer offsets back to the user's exec list. */
1250 ret
= copy_to_user(to_user_ptr(args
->buffers_ptr
),
1252 sizeof(*exec2_list
) * args
->buffer_count
);
1255 DRM_DEBUG("failed to copy %d exec entries "
1256 "back to user (%d)\n",
1257 args
->buffer_count
, ret
);
1261 drm_free_large(exec2_list
);