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>
29 #include <linux/dma_remapping.h>
30 #include <linux/reservation.h>
31 #include <linux/sync_file.h>
32 #include <linux/uaccess.h>
35 #include <drm/i915_drm.h>
38 #include "i915_gem_clflush.h"
39 #include "i915_trace.h"
40 #include "intel_drv.h"
41 #include "intel_frontbuffer.h"
43 #define DBG_USE_CPU_RELOC 0 /* -1 force GTT relocs; 1 force CPU relocs */
45 #define __EXEC_OBJECT_HAS_PIN (1<<31)
46 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
47 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
48 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
49 #define __EXEC_OBJECT_INTERNAL_FLAGS (0xf<<28) /* all of the above */
51 #define BATCH_OFFSET_BIAS (256*1024)
53 struct i915_execbuffer_params
{
54 struct drm_device
*dev
;
55 struct drm_file
*file
;
56 struct i915_vma
*batch
;
58 u32 args_batch_start_offset
;
59 struct intel_engine_cs
*engine
;
60 struct i915_gem_context
*ctx
;
61 struct drm_i915_gem_request
*request
;
65 struct drm_i915_private
*i915
;
66 struct list_head vmas
;
69 struct i915_vma
*lut
[0];
70 struct hlist_head buckets
[0];
74 static struct eb_vmas
*
75 eb_create(struct drm_i915_private
*i915
,
76 struct drm_i915_gem_execbuffer2
*args
)
78 struct eb_vmas
*eb
= NULL
;
80 if (args
->flags
& I915_EXEC_HANDLE_LUT
) {
81 unsigned size
= args
->buffer_count
;
82 size
*= sizeof(struct i915_vma
*);
83 size
+= sizeof(struct eb_vmas
);
84 eb
= kmalloc(size
, GFP_TEMPORARY
| __GFP_NOWARN
| __GFP_NORETRY
);
88 unsigned size
= args
->buffer_count
;
89 unsigned count
= PAGE_SIZE
/ sizeof(struct hlist_head
) / 2;
90 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE
/ sizeof(struct hlist_head
));
91 while (count
> 2*size
)
93 eb
= kzalloc(count
*sizeof(struct hlist_head
) +
94 sizeof(struct eb_vmas
),
101 eb
->and = -args
->buffer_count
;
104 INIT_LIST_HEAD(&eb
->vmas
);
109 eb_reset(struct eb_vmas
*eb
)
112 memset(eb
->buckets
, 0, (eb
->and+1)*sizeof(struct hlist_head
));
115 static struct i915_vma
*
116 eb_get_batch(struct eb_vmas
*eb
)
118 struct i915_vma
*vma
= list_entry(eb
->vmas
.prev
, typeof(*vma
), exec_list
);
121 * SNA is doing fancy tricks with compressing batch buffers, which leads
122 * to negative relocation deltas. Usually that works out ok since the
123 * relocate address is still positive, except when the batch is placed
124 * very low in the GTT. Ensure this doesn't happen.
126 * Note that actual hangs have only been observed on gen7, but for
127 * paranoia do it everywhere.
129 if ((vma
->exec_entry
->flags
& EXEC_OBJECT_PINNED
) == 0)
130 vma
->exec_entry
->flags
|= __EXEC_OBJECT_NEEDS_BIAS
;
136 eb_lookup_vmas(struct eb_vmas
*eb
,
137 struct drm_i915_gem_exec_object2
*exec
,
138 const struct drm_i915_gem_execbuffer2
*args
,
139 struct i915_address_space
*vm
,
140 struct drm_file
*file
)
142 struct drm_i915_gem_object
*obj
;
143 struct list_head objects
;
146 INIT_LIST_HEAD(&objects
);
147 spin_lock(&file
->table_lock
);
148 /* Grab a reference to the object and release the lock so we can lookup
149 * or create the VMA without using GFP_ATOMIC */
150 for (i
= 0; i
< args
->buffer_count
; i
++) {
151 obj
= to_intel_bo(idr_find(&file
->object_idr
, exec
[i
].handle
));
153 spin_unlock(&file
->table_lock
);
154 DRM_DEBUG("Invalid object handle %d at index %d\n",
160 if (!list_empty(&obj
->obj_exec_link
)) {
161 spin_unlock(&file
->table_lock
);
162 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
163 obj
, exec
[i
].handle
, i
);
168 i915_gem_object_get(obj
);
169 list_add_tail(&obj
->obj_exec_link
, &objects
);
171 spin_unlock(&file
->table_lock
);
174 while (!list_empty(&objects
)) {
175 struct i915_vma
*vma
;
177 obj
= list_first_entry(&objects
,
178 struct drm_i915_gem_object
,
182 * NOTE: We can leak any vmas created here when something fails
183 * later on. But that's no issue since vma_unbind can deal with
184 * vmas which are not actually bound. And since only
185 * lookup_or_create exists as an interface to get at the vma
186 * from the (obj, vm) we don't run the risk of creating
187 * duplicated vmas for the same vm.
189 vma
= i915_vma_instance(obj
, vm
, NULL
);
190 if (unlikely(IS_ERR(vma
))) {
191 DRM_DEBUG("Failed to lookup VMA\n");
196 /* Transfer ownership from the objects list to the vmas list. */
197 list_add_tail(&vma
->exec_list
, &eb
->vmas
);
198 list_del_init(&obj
->obj_exec_link
);
200 vma
->exec_entry
= &exec
[i
];
204 uint32_t handle
= args
->flags
& I915_EXEC_HANDLE_LUT
? i
: exec
[i
].handle
;
205 vma
->exec_handle
= handle
;
206 hlist_add_head(&vma
->exec_node
,
207 &eb
->buckets
[handle
& eb
->and]);
216 while (!list_empty(&objects
)) {
217 obj
= list_first_entry(&objects
,
218 struct drm_i915_gem_object
,
220 list_del_init(&obj
->obj_exec_link
);
221 i915_gem_object_put(obj
);
224 * Objects already transfered to the vmas list will be unreferenced by
231 static struct i915_vma
*eb_get_vma(struct eb_vmas
*eb
, unsigned long handle
)
234 if (handle
>= -eb
->and)
236 return eb
->lut
[handle
];
238 struct hlist_head
*head
;
239 struct i915_vma
*vma
;
241 head
= &eb
->buckets
[handle
& eb
->and];
242 hlist_for_each_entry(vma
, head
, exec_node
) {
243 if (vma
->exec_handle
== handle
)
251 i915_gem_execbuffer_unreserve_vma(struct i915_vma
*vma
)
253 struct drm_i915_gem_exec_object2
*entry
;
255 if (!drm_mm_node_allocated(&vma
->node
))
258 entry
= vma
->exec_entry
;
260 if (entry
->flags
& __EXEC_OBJECT_HAS_FENCE
)
261 i915_vma_unpin_fence(vma
);
263 if (entry
->flags
& __EXEC_OBJECT_HAS_PIN
)
264 __i915_vma_unpin(vma
);
266 entry
->flags
&= ~(__EXEC_OBJECT_HAS_FENCE
| __EXEC_OBJECT_HAS_PIN
);
269 static void eb_destroy(struct eb_vmas
*eb
)
271 while (!list_empty(&eb
->vmas
)) {
272 struct i915_vma
*vma
;
274 vma
= list_first_entry(&eb
->vmas
,
277 list_del_init(&vma
->exec_list
);
278 i915_gem_execbuffer_unreserve_vma(vma
);
279 vma
->exec_entry
= NULL
;
285 static inline int use_cpu_reloc(struct drm_i915_gem_object
*obj
)
287 if (!i915_gem_object_has_struct_page(obj
))
290 if (DBG_USE_CPU_RELOC
)
291 return DBG_USE_CPU_RELOC
> 0;
293 return (HAS_LLC(to_i915(obj
->base
.dev
)) ||
294 obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
||
295 obj
->cache_level
!= I915_CACHE_NONE
);
298 /* Used to convert any address to canonical form.
299 * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
300 * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
301 * addresses to be in a canonical form:
302 * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
303 * canonical form [63:48] == [47]."
305 #define GEN8_HIGH_ADDRESS_BIT 47
306 static inline uint64_t gen8_canonical_addr(uint64_t address
)
308 return sign_extend64(address
, GEN8_HIGH_ADDRESS_BIT
);
311 static inline uint64_t gen8_noncanonical_addr(uint64_t address
)
313 return address
& ((1ULL << (GEN8_HIGH_ADDRESS_BIT
+ 1)) - 1);
316 static inline uint64_t
317 relocation_target(const struct drm_i915_gem_relocation_entry
*reloc
,
318 uint64_t target_offset
)
320 return gen8_canonical_addr((int)reloc
->delta
+ target_offset
);
324 struct drm_i915_private
*i915
;
325 struct drm_mm_node node
;
328 bool use_64bit_reloc
;
331 static void reloc_cache_init(struct reloc_cache
*cache
,
332 struct drm_i915_private
*i915
)
337 /* Must be a variable in the struct to allow GCC to unroll. */
338 cache
->use_64bit_reloc
= HAS_64BIT_RELOC(i915
);
339 cache
->node
.allocated
= false;
342 static inline void *unmask_page(unsigned long p
)
344 return (void *)(uintptr_t)(p
& PAGE_MASK
);
347 static inline unsigned int unmask_flags(unsigned long p
)
349 return p
& ~PAGE_MASK
;
352 #define KMAP 0x4 /* after CLFLUSH_FLAGS */
354 static void reloc_cache_fini(struct reloc_cache
*cache
)
361 vaddr
= unmask_page(cache
->vaddr
);
362 if (cache
->vaddr
& KMAP
) {
363 if (cache
->vaddr
& CLFLUSH_AFTER
)
366 kunmap_atomic(vaddr
);
367 i915_gem_obj_finish_shmem_access((struct drm_i915_gem_object
*)cache
->node
.mm
);
370 io_mapping_unmap_atomic((void __iomem
*)vaddr
);
371 if (cache
->node
.allocated
) {
372 struct i915_ggtt
*ggtt
= &cache
->i915
->ggtt
;
374 ggtt
->base
.clear_range(&ggtt
->base
,
377 drm_mm_remove_node(&cache
->node
);
379 i915_vma_unpin((struct i915_vma
*)cache
->node
.mm
);
384 static void *reloc_kmap(struct drm_i915_gem_object
*obj
,
385 struct reloc_cache
*cache
,
391 kunmap_atomic(unmask_page(cache
->vaddr
));
393 unsigned int flushes
;
396 ret
= i915_gem_obj_prepare_shmem_write(obj
, &flushes
);
400 BUILD_BUG_ON(KMAP
& CLFLUSH_FLAGS
);
401 BUILD_BUG_ON((KMAP
| CLFLUSH_FLAGS
) & PAGE_MASK
);
403 cache
->vaddr
= flushes
| KMAP
;
404 cache
->node
.mm
= (void *)obj
;
409 vaddr
= kmap_atomic(i915_gem_object_get_dirty_page(obj
, page
));
410 cache
->vaddr
= unmask_flags(cache
->vaddr
) | (unsigned long)vaddr
;
416 static void *reloc_iomap(struct drm_i915_gem_object
*obj
,
417 struct reloc_cache
*cache
,
420 struct i915_ggtt
*ggtt
= &cache
->i915
->ggtt
;
421 unsigned long offset
;
425 io_mapping_unmap_atomic((void __force __iomem
*) unmask_page(cache
->vaddr
));
427 struct i915_vma
*vma
;
430 if (use_cpu_reloc(obj
))
433 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
437 vma
= i915_gem_object_ggtt_pin(obj
, NULL
, 0, 0,
438 PIN_MAPPABLE
| PIN_NONBLOCK
);
440 memset(&cache
->node
, 0, sizeof(cache
->node
));
441 ret
= drm_mm_insert_node_in_range
442 (&ggtt
->base
.mm
, &cache
->node
,
443 PAGE_SIZE
, 0, I915_COLOR_UNEVICTABLE
,
444 0, ggtt
->mappable_end
,
446 if (ret
) /* no inactive aperture space, use cpu reloc */
449 ret
= i915_vma_put_fence(vma
);
455 cache
->node
.start
= vma
->node
.start
;
456 cache
->node
.mm
= (void *)vma
;
460 offset
= cache
->node
.start
;
461 if (cache
->node
.allocated
) {
463 ggtt
->base
.insert_page(&ggtt
->base
,
464 i915_gem_object_get_dma_address(obj
, page
),
465 offset
, I915_CACHE_NONE
, 0);
467 offset
+= page
<< PAGE_SHIFT
;
470 vaddr
= (void __force
*) io_mapping_map_atomic_wc(&cache
->i915
->ggtt
.mappable
, offset
);
472 cache
->vaddr
= (unsigned long)vaddr
;
477 static void *reloc_vaddr(struct drm_i915_gem_object
*obj
,
478 struct reloc_cache
*cache
,
483 if (cache
->page
== page
) {
484 vaddr
= unmask_page(cache
->vaddr
);
487 if ((cache
->vaddr
& KMAP
) == 0)
488 vaddr
= reloc_iomap(obj
, cache
, page
);
490 vaddr
= reloc_kmap(obj
, cache
, page
);
496 static void clflush_write32(u32
*addr
, u32 value
, unsigned int flushes
)
498 if (unlikely(flushes
& (CLFLUSH_BEFORE
| CLFLUSH_AFTER
))) {
499 if (flushes
& CLFLUSH_BEFORE
) {
506 /* Writes to the same cacheline are serialised by the CPU
507 * (including clflush). On the write path, we only require
508 * that it hits memory in an orderly fashion and place
509 * mb barriers at the start and end of the relocation phase
510 * to ensure ordering of clflush wrt to the system.
512 if (flushes
& CLFLUSH_AFTER
)
519 relocate_entry(struct drm_i915_gem_object
*obj
,
520 const struct drm_i915_gem_relocation_entry
*reloc
,
521 struct reloc_cache
*cache
,
524 u64 offset
= reloc
->offset
;
525 bool wide
= cache
->use_64bit_reloc
;
528 target_offset
= relocation_target(reloc
, target_offset
);
530 vaddr
= reloc_vaddr(obj
, cache
, offset
>> PAGE_SHIFT
);
532 return PTR_ERR(vaddr
);
534 clflush_write32(vaddr
+ offset_in_page(offset
),
535 lower_32_bits(target_offset
),
539 offset
+= sizeof(u32
);
540 target_offset
>>= 32;
549 i915_gem_execbuffer_relocate_entry(struct i915_vma
*vma
,
551 struct drm_i915_gem_relocation_entry
*reloc
,
552 struct reloc_cache
*cache
)
554 struct drm_i915_gem_object
*obj
= vma
->obj
;
555 struct drm_i915_private
*dev_priv
= to_i915(obj
->base
.dev
);
556 struct drm_gem_object
*target_obj
;
557 struct drm_i915_gem_object
*target_i915_obj
;
558 struct i915_vma
*target_vma
;
559 uint64_t target_offset
;
562 /* we've already hold a reference to all valid objects */
563 target_vma
= eb_get_vma(eb
, reloc
->target_handle
);
564 if (unlikely(target_vma
== NULL
))
566 target_i915_obj
= target_vma
->obj
;
567 target_obj
= &target_vma
->obj
->base
;
569 target_offset
= gen8_canonical_addr(target_vma
->node
.start
);
571 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
572 * pipe_control writes because the gpu doesn't properly redirect them
573 * through the ppgtt for non_secure batchbuffers. */
574 if (unlikely(IS_GEN6(dev_priv
) &&
575 reloc
->write_domain
== I915_GEM_DOMAIN_INSTRUCTION
)) {
576 ret
= i915_vma_bind(target_vma
, target_i915_obj
->cache_level
,
578 if (WARN_ONCE(ret
, "Unexpected failure to bind target VMA!"))
582 /* Validate that the target is in a valid r/w GPU domain */
583 if (unlikely(reloc
->write_domain
& (reloc
->write_domain
- 1))) {
584 DRM_DEBUG("reloc with multiple write domains: "
585 "obj %p target %d offset %d "
586 "read %08x write %08x",
587 obj
, reloc
->target_handle
,
590 reloc
->write_domain
);
593 if (unlikely((reloc
->write_domain
| reloc
->read_domains
)
594 & ~I915_GEM_GPU_DOMAINS
)) {
595 DRM_DEBUG("reloc with read/write non-GPU domains: "
596 "obj %p target %d offset %d "
597 "read %08x write %08x",
598 obj
, reloc
->target_handle
,
601 reloc
->write_domain
);
605 target_obj
->pending_read_domains
|= reloc
->read_domains
;
606 target_obj
->pending_write_domain
|= reloc
->write_domain
;
608 /* If the relocation already has the right value in it, no
609 * more work needs to be done.
611 if (target_offset
== reloc
->presumed_offset
)
614 /* Check that the relocation address is valid... */
615 if (unlikely(reloc
->offset
>
616 obj
->base
.size
- (cache
->use_64bit_reloc
? 8 : 4))) {
617 DRM_DEBUG("Relocation beyond object bounds: "
618 "obj %p target %d offset %d size %d.\n",
619 obj
, reloc
->target_handle
,
621 (int) obj
->base
.size
);
624 if (unlikely(reloc
->offset
& 3)) {
625 DRM_DEBUG("Relocation not 4-byte aligned: "
626 "obj %p target %d offset %d.\n",
627 obj
, reloc
->target_handle
,
628 (int) reloc
->offset
);
633 * If we write into the object, we need to force the synchronisation
634 * barrier, either with an asynchronous clflush or if we executed the
635 * patching using the GPU (though that should be serialised by the
636 * timeline). To be completely sure, and since we are required to
637 * do relocations we are already stalling, disable the user's opt
638 * of our synchronisation.
640 vma
->exec_entry
->flags
&= ~EXEC_OBJECT_ASYNC
;
642 ret
= relocate_entry(obj
, reloc
, cache
, target_offset
);
646 /* and update the user's relocation entry */
647 reloc
->presumed_offset
= target_offset
;
652 i915_gem_execbuffer_relocate_vma(struct i915_vma
*vma
,
655 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
656 struct drm_i915_gem_relocation_entry stack_reloc
[N_RELOC(512)];
657 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
658 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
659 struct reloc_cache cache
;
662 user_relocs
= u64_to_user_ptr(entry
->relocs_ptr
);
663 reloc_cache_init(&cache
, eb
->i915
);
665 remain
= entry
->relocation_count
;
667 struct drm_i915_gem_relocation_entry
*r
= stack_reloc
;
668 unsigned long unwritten
;
671 count
= min_t(unsigned int, remain
, ARRAY_SIZE(stack_reloc
));
674 /* This is the fast path and we cannot handle a pagefault
675 * whilst holding the struct mutex lest the user pass in the
676 * relocations contained within a mmaped bo. For in such a case
677 * we, the page fault handler would call i915_gem_fault() and
678 * we would try to acquire the struct mutex again. Obviously
679 * this is bad and so lockdep complains vehemently.
682 unwritten
= __copy_from_user_inatomic(r
, user_relocs
, count
*sizeof(r
[0]));
684 if (unlikely(unwritten
)) {
690 u64 offset
= r
->presumed_offset
;
692 ret
= i915_gem_execbuffer_relocate_entry(vma
, eb
, r
, &cache
);
696 if (r
->presumed_offset
!= offset
) {
698 unwritten
= __put_user(r
->presumed_offset
,
699 &user_relocs
->presumed_offset
);
701 if (unlikely(unwritten
)) {
702 /* Note that reporting an error now
703 * leaves everything in an inconsistent
704 * state as we have *already* changed
705 * the relocation value inside the
706 * object. As we have not changed the
707 * reloc.presumed_offset or will not
708 * change the execobject.offset, on the
709 * call we may not rewrite the value
710 * inside the object, leaving it
711 * dangling and causing a GPU hang.
724 reloc_cache_fini(&cache
);
730 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma
*vma
,
732 struct drm_i915_gem_relocation_entry
*relocs
)
734 const struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
735 struct reloc_cache cache
;
738 reloc_cache_init(&cache
, eb
->i915
);
739 for (i
= 0; i
< entry
->relocation_count
; i
++) {
740 ret
= i915_gem_execbuffer_relocate_entry(vma
, eb
, &relocs
[i
], &cache
);
744 reloc_cache_fini(&cache
);
750 i915_gem_execbuffer_relocate(struct eb_vmas
*eb
)
752 struct i915_vma
*vma
;
755 list_for_each_entry(vma
, &eb
->vmas
, exec_list
) {
756 ret
= i915_gem_execbuffer_relocate_vma(vma
, eb
);
764 static bool only_mappable_for_reloc(unsigned int flags
)
766 return (flags
& (EXEC_OBJECT_NEEDS_FENCE
| __EXEC_OBJECT_NEEDS_MAP
)) ==
767 __EXEC_OBJECT_NEEDS_MAP
;
771 i915_gem_execbuffer_reserve_vma(struct i915_vma
*vma
,
772 struct intel_engine_cs
*engine
,
775 struct drm_i915_gem_object
*obj
= vma
->obj
;
776 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
781 if (entry
->flags
& EXEC_OBJECT_NEEDS_GTT
)
784 if (!drm_mm_node_allocated(&vma
->node
)) {
785 /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
786 * limit address to the first 4GBs for unflagged objects.
788 if ((entry
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) == 0)
789 flags
|= PIN_ZONE_4G
;
790 if (entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
)
791 flags
|= PIN_GLOBAL
| PIN_MAPPABLE
;
792 if (entry
->flags
& __EXEC_OBJECT_NEEDS_BIAS
)
793 flags
|= BATCH_OFFSET_BIAS
| PIN_OFFSET_BIAS
;
794 if (entry
->flags
& EXEC_OBJECT_PINNED
)
795 flags
|= entry
->offset
| PIN_OFFSET_FIXED
;
796 if ((flags
& PIN_MAPPABLE
) == 0)
800 ret
= i915_vma_pin(vma
,
804 if ((ret
== -ENOSPC
|| ret
== -E2BIG
) &&
805 only_mappable_for_reloc(entry
->flags
))
806 ret
= i915_vma_pin(vma
,
809 flags
& ~PIN_MAPPABLE
);
813 entry
->flags
|= __EXEC_OBJECT_HAS_PIN
;
815 if (entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
) {
816 ret
= i915_vma_get_fence(vma
);
820 if (i915_vma_pin_fence(vma
))
821 entry
->flags
|= __EXEC_OBJECT_HAS_FENCE
;
824 if (entry
->offset
!= vma
->node
.start
) {
825 entry
->offset
= vma
->node
.start
;
829 if (entry
->flags
& EXEC_OBJECT_WRITE
) {
830 obj
->base
.pending_read_domains
= I915_GEM_DOMAIN_RENDER
;
831 obj
->base
.pending_write_domain
= I915_GEM_DOMAIN_RENDER
;
838 need_reloc_mappable(struct i915_vma
*vma
)
840 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
842 if (entry
->relocation_count
== 0)
845 if (!i915_vma_is_ggtt(vma
))
848 /* See also use_cpu_reloc() */
849 if (HAS_LLC(to_i915(vma
->obj
->base
.dev
)))
852 if (vma
->obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
)
859 eb_vma_misplaced(struct i915_vma
*vma
)
861 struct drm_i915_gem_exec_object2
*entry
= vma
->exec_entry
;
863 WARN_ON(entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
&&
864 !i915_vma_is_ggtt(vma
));
866 if (entry
->alignment
&& !IS_ALIGNED(vma
->node
.start
, entry
->alignment
))
869 if (vma
->node
.size
< entry
->pad_to_size
)
872 if (entry
->flags
& EXEC_OBJECT_PINNED
&&
873 vma
->node
.start
!= entry
->offset
)
876 if (entry
->flags
& __EXEC_OBJECT_NEEDS_BIAS
&&
877 vma
->node
.start
< BATCH_OFFSET_BIAS
)
880 /* avoid costly ping-pong once a batch bo ended up non-mappable */
881 if (entry
->flags
& __EXEC_OBJECT_NEEDS_MAP
&&
882 !i915_vma_is_map_and_fenceable(vma
))
883 return !only_mappable_for_reloc(entry
->flags
);
885 if ((entry
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) == 0 &&
886 (vma
->node
.start
+ vma
->node
.size
- 1) >> 32)
893 i915_gem_execbuffer_reserve(struct intel_engine_cs
*engine
,
894 struct list_head
*vmas
,
895 struct i915_gem_context
*ctx
,
898 struct drm_i915_gem_object
*obj
;
899 struct i915_vma
*vma
;
900 struct i915_address_space
*vm
;
901 struct list_head ordered_vmas
;
902 struct list_head pinned_vmas
;
903 bool has_fenced_gpu_access
= INTEL_GEN(engine
->i915
) < 4;
904 bool needs_unfenced_map
= INTEL_INFO(engine
->i915
)->unfenced_needs_alignment
;
907 vm
= list_first_entry(vmas
, struct i915_vma
, exec_list
)->vm
;
909 INIT_LIST_HEAD(&ordered_vmas
);
910 INIT_LIST_HEAD(&pinned_vmas
);
911 while (!list_empty(vmas
)) {
912 struct drm_i915_gem_exec_object2
*entry
;
913 bool need_fence
, need_mappable
;
915 vma
= list_first_entry(vmas
, struct i915_vma
, exec_list
);
917 entry
= vma
->exec_entry
;
919 if (ctx
->flags
& CONTEXT_NO_ZEROMAP
)
920 entry
->flags
|= __EXEC_OBJECT_NEEDS_BIAS
;
922 if (!has_fenced_gpu_access
)
923 entry
->flags
&= ~EXEC_OBJECT_NEEDS_FENCE
;
925 (entry
->flags
& EXEC_OBJECT_NEEDS_FENCE
||
926 needs_unfenced_map
) &&
927 i915_gem_object_is_tiled(obj
);
928 need_mappable
= need_fence
|| need_reloc_mappable(vma
);
930 if (entry
->flags
& EXEC_OBJECT_PINNED
)
931 list_move_tail(&vma
->exec_list
, &pinned_vmas
);
932 else if (need_mappable
) {
933 entry
->flags
|= __EXEC_OBJECT_NEEDS_MAP
;
934 list_move(&vma
->exec_list
, &ordered_vmas
);
936 list_move_tail(&vma
->exec_list
, &ordered_vmas
);
938 obj
->base
.pending_read_domains
= I915_GEM_GPU_DOMAINS
& ~I915_GEM_DOMAIN_COMMAND
;
939 obj
->base
.pending_write_domain
= 0;
941 list_splice(&ordered_vmas
, vmas
);
942 list_splice(&pinned_vmas
, vmas
);
944 /* Attempt to pin all of the buffers into the GTT.
945 * This is done in 3 phases:
947 * 1a. Unbind all objects that do not match the GTT constraints for
948 * the execbuffer (fenceable, mappable, alignment etc).
949 * 1b. Increment pin count for already bound objects.
950 * 2. Bind new objects.
951 * 3. Decrement pin count.
953 * This avoid unnecessary unbinding of later objects in order to make
954 * room for the earlier objects *unless* we need to defragment.
960 /* Unbind any ill-fitting objects or pin. */
961 list_for_each_entry(vma
, vmas
, exec_list
) {
962 if (!drm_mm_node_allocated(&vma
->node
))
965 if (eb_vma_misplaced(vma
))
966 ret
= i915_vma_unbind(vma
);
968 ret
= i915_gem_execbuffer_reserve_vma(vma
,
975 /* Bind fresh objects */
976 list_for_each_entry(vma
, vmas
, exec_list
) {
977 if (drm_mm_node_allocated(&vma
->node
))
980 ret
= i915_gem_execbuffer_reserve_vma(vma
, engine
,
987 if (ret
!= -ENOSPC
|| retry
++)
990 /* Decrement pin count for bound objects */
991 list_for_each_entry(vma
, vmas
, exec_list
)
992 i915_gem_execbuffer_unreserve_vma(vma
);
994 ret
= i915_gem_evict_vm(vm
, true);
1001 i915_gem_execbuffer_relocate_slow(struct drm_device
*dev
,
1002 struct drm_i915_gem_execbuffer2
*args
,
1003 struct drm_file
*file
,
1004 struct intel_engine_cs
*engine
,
1006 struct drm_i915_gem_exec_object2
*exec
,
1007 struct i915_gem_context
*ctx
)
1009 struct drm_i915_gem_relocation_entry
*reloc
;
1010 struct i915_address_space
*vm
;
1011 struct i915_vma
*vma
;
1015 unsigned count
= args
->buffer_count
;
1017 vm
= list_first_entry(&eb
->vmas
, struct i915_vma
, exec_list
)->vm
;
1019 /* We may process another execbuffer during the unlock... */
1020 while (!list_empty(&eb
->vmas
)) {
1021 vma
= list_first_entry(&eb
->vmas
, struct i915_vma
, exec_list
);
1022 list_del_init(&vma
->exec_list
);
1023 i915_gem_execbuffer_unreserve_vma(vma
);
1027 mutex_unlock(&dev
->struct_mutex
);
1030 for (i
= 0; i
< count
; i
++)
1031 total
+= exec
[i
].relocation_count
;
1033 reloc_offset
= drm_malloc_ab(count
, sizeof(*reloc_offset
));
1034 reloc
= drm_malloc_ab(total
, sizeof(*reloc
));
1035 if (reloc
== NULL
|| reloc_offset
== NULL
) {
1036 drm_free_large(reloc
);
1037 drm_free_large(reloc_offset
);
1038 mutex_lock(&dev
->struct_mutex
);
1043 for (i
= 0; i
< count
; i
++) {
1044 struct drm_i915_gem_relocation_entry __user
*user_relocs
;
1045 u64 invalid_offset
= (u64
)-1;
1048 user_relocs
= u64_to_user_ptr(exec
[i
].relocs_ptr
);
1050 if (copy_from_user(reloc
+total
, user_relocs
,
1051 exec
[i
].relocation_count
* sizeof(*reloc
))) {
1053 mutex_lock(&dev
->struct_mutex
);
1057 /* As we do not update the known relocation offsets after
1058 * relocating (due to the complexities in lock handling),
1059 * we need to mark them as invalid now so that we force the
1060 * relocation processing next time. Just in case the target
1061 * object is evicted and then rebound into its old
1062 * presumed_offset before the next execbuffer - if that
1063 * happened we would make the mistake of assuming that the
1064 * relocations were valid.
1066 for (j
= 0; j
< exec
[i
].relocation_count
; j
++) {
1067 if (__copy_to_user(&user_relocs
[j
].presumed_offset
,
1069 sizeof(invalid_offset
))) {
1071 mutex_lock(&dev
->struct_mutex
);
1076 reloc_offset
[i
] = total
;
1077 total
+= exec
[i
].relocation_count
;
1080 ret
= i915_mutex_lock_interruptible(dev
);
1082 mutex_lock(&dev
->struct_mutex
);
1086 /* reacquire the objects */
1088 ret
= eb_lookup_vmas(eb
, exec
, args
, vm
, file
);
1092 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
1093 ret
= i915_gem_execbuffer_reserve(engine
, &eb
->vmas
, ctx
,
1098 list_for_each_entry(vma
, &eb
->vmas
, exec_list
) {
1099 int offset
= vma
->exec_entry
- exec
;
1100 ret
= i915_gem_execbuffer_relocate_vma_slow(vma
, eb
,
1101 reloc
+ reloc_offset
[offset
]);
1106 /* Leave the user relocations as are, this is the painfully slow path,
1107 * and we want to avoid the complication of dropping the lock whilst
1108 * having buffers reserved in the aperture and so causing spurious
1109 * ENOSPC for random operations.
1113 drm_free_large(reloc
);
1114 drm_free_large(reloc_offset
);
1119 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request
*req
,
1120 struct list_head
*vmas
)
1122 struct i915_vma
*vma
;
1125 list_for_each_entry(vma
, vmas
, exec_list
) {
1126 struct drm_i915_gem_object
*obj
= vma
->obj
;
1128 if (vma
->exec_entry
->flags
& EXEC_OBJECT_ASYNC
)
1131 if (obj
->base
.write_domain
& I915_GEM_DOMAIN_CPU
) {
1132 i915_gem_clflush_object(obj
, 0);
1133 obj
->base
.write_domain
= 0;
1136 ret
= i915_gem_request_await_object
1137 (req
, obj
, obj
->base
.pending_write_domain
);
1142 /* Unconditionally flush any chipset caches (for streaming writes). */
1143 i915_gem_chipset_flush(req
->engine
->i915
);
1145 /* Unconditionally invalidate GPU caches and TLBs. */
1146 return req
->engine
->emit_flush(req
, EMIT_INVALIDATE
);
1150 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2
*exec
)
1152 if (exec
->flags
& __I915_EXEC_UNKNOWN_FLAGS
)
1155 /* Kernel clipping was a DRI1 misfeature */
1156 if (exec
->num_cliprects
|| exec
->cliprects_ptr
)
1159 if (exec
->DR4
== 0xffffffff) {
1160 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1163 if (exec
->DR1
|| exec
->DR4
)
1166 if ((exec
->batch_start_offset
| exec
->batch_len
) & 0x7)
1173 validate_exec_list(struct drm_device
*dev
,
1174 struct drm_i915_gem_exec_object2
*exec
,
1177 unsigned relocs_total
= 0;
1178 unsigned relocs_max
= UINT_MAX
/ sizeof(struct drm_i915_gem_relocation_entry
);
1179 unsigned invalid_flags
;
1182 /* INTERNAL flags must not overlap with external ones */
1183 BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS
& ~__EXEC_OBJECT_UNKNOWN_FLAGS
);
1185 invalid_flags
= __EXEC_OBJECT_UNKNOWN_FLAGS
;
1186 if (USES_FULL_PPGTT(dev
))
1187 invalid_flags
|= EXEC_OBJECT_NEEDS_GTT
;
1189 for (i
= 0; i
< count
; i
++) {
1190 char __user
*ptr
= u64_to_user_ptr(exec
[i
].relocs_ptr
);
1191 int length
; /* limited by fault_in_pages_readable() */
1193 if (exec
[i
].flags
& invalid_flags
)
1196 /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
1197 * any non-page-aligned or non-canonical addresses.
1199 if (exec
[i
].flags
& EXEC_OBJECT_PINNED
) {
1200 if (exec
[i
].offset
!=
1201 gen8_canonical_addr(exec
[i
].offset
& PAGE_MASK
))
1205 /* From drm_mm perspective address space is continuous,
1206 * so from this point we're always using non-canonical
1209 exec
[i
].offset
= gen8_noncanonical_addr(exec
[i
].offset
);
1211 if (exec
[i
].alignment
&& !is_power_of_2(exec
[i
].alignment
))
1214 /* pad_to_size was once a reserved field, so sanitize it */
1215 if (exec
[i
].flags
& EXEC_OBJECT_PAD_TO_SIZE
) {
1216 if (offset_in_page(exec
[i
].pad_to_size
))
1219 exec
[i
].pad_to_size
= 0;
1222 /* First check for malicious input causing overflow in
1223 * the worst case where we need to allocate the entire
1224 * relocation tree as a single array.
1226 if (exec
[i
].relocation_count
> relocs_max
- relocs_total
)
1228 relocs_total
+= exec
[i
].relocation_count
;
1230 length
= exec
[i
].relocation_count
*
1231 sizeof(struct drm_i915_gem_relocation_entry
);
1233 * We must check that the entire relocation array is safe
1234 * to read, but since we may need to update the presumed
1235 * offsets during execution, check for full write access.
1237 if (!access_ok(VERIFY_WRITE
, ptr
, length
))
1240 if (likely(!i915
.prefault_disable
)) {
1241 if (fault_in_pages_readable(ptr
, length
))
1249 static struct i915_gem_context
*
1250 i915_gem_validate_context(struct drm_device
*dev
, struct drm_file
*file
,
1251 struct intel_engine_cs
*engine
, const u32 ctx_id
)
1253 struct i915_gem_context
*ctx
;
1255 ctx
= i915_gem_context_lookup(file
->driver_priv
, ctx_id
);
1259 if (i915_gem_context_is_banned(ctx
)) {
1260 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id
);
1261 return ERR_PTR(-EIO
);
1267 static bool gpu_write_needs_clflush(struct drm_i915_gem_object
*obj
)
1269 return !(obj
->cache_level
== I915_CACHE_NONE
||
1270 obj
->cache_level
== I915_CACHE_WT
);
1273 void i915_vma_move_to_active(struct i915_vma
*vma
,
1274 struct drm_i915_gem_request
*req
,
1277 struct drm_i915_gem_object
*obj
= vma
->obj
;
1278 const unsigned int idx
= req
->engine
->id
;
1280 lockdep_assert_held(&req
->i915
->drm
.struct_mutex
);
1281 GEM_BUG_ON(!drm_mm_node_allocated(&vma
->node
));
1283 /* Add a reference if we're newly entering the active list.
1284 * The order in which we add operations to the retirement queue is
1285 * vital here: mark_active adds to the start of the callback list,
1286 * such that subsequent callbacks are called first. Therefore we
1287 * add the active reference first and queue for it to be dropped
1290 if (!i915_vma_is_active(vma
))
1291 obj
->active_count
++;
1292 i915_vma_set_active(vma
, idx
);
1293 i915_gem_active_set(&vma
->last_read
[idx
], req
);
1294 list_move_tail(&vma
->vm_link
, &vma
->vm
->active_list
);
1296 if (flags
& EXEC_OBJECT_WRITE
) {
1297 if (intel_fb_obj_invalidate(obj
, ORIGIN_CS
))
1298 i915_gem_active_set(&obj
->frontbuffer_write
, req
);
1300 /* update for the implicit flush after a batch */
1301 obj
->base
.write_domain
&= ~I915_GEM_GPU_DOMAINS
;
1302 if (!obj
->cache_dirty
&& gpu_write_needs_clflush(obj
))
1303 obj
->cache_dirty
= true;
1306 if (flags
& EXEC_OBJECT_NEEDS_FENCE
)
1307 i915_gem_active_set(&vma
->last_fence
, req
);
1310 static void eb_export_fence(struct drm_i915_gem_object
*obj
,
1311 struct drm_i915_gem_request
*req
,
1314 struct reservation_object
*resv
= obj
->resv
;
1316 /* Ignore errors from failing to allocate the new fence, we can't
1317 * handle an error right now. Worst case should be missed
1318 * synchronisation leading to rendering corruption.
1320 reservation_object_lock(resv
, NULL
);
1321 if (flags
& EXEC_OBJECT_WRITE
)
1322 reservation_object_add_excl_fence(resv
, &req
->fence
);
1323 else if (reservation_object_reserve_shared(resv
) == 0)
1324 reservation_object_add_shared_fence(resv
, &req
->fence
);
1325 reservation_object_unlock(resv
);
1329 i915_gem_execbuffer_move_to_active(struct list_head
*vmas
,
1330 struct drm_i915_gem_request
*req
)
1332 struct i915_vma
*vma
;
1334 list_for_each_entry(vma
, vmas
, exec_list
) {
1335 struct drm_i915_gem_object
*obj
= vma
->obj
;
1337 obj
->base
.write_domain
= obj
->base
.pending_write_domain
;
1338 if (obj
->base
.write_domain
)
1339 vma
->exec_entry
->flags
|= EXEC_OBJECT_WRITE
;
1341 obj
->base
.pending_read_domains
|= obj
->base
.read_domains
;
1342 obj
->base
.read_domains
= obj
->base
.pending_read_domains
;
1344 i915_vma_move_to_active(vma
, req
, vma
->exec_entry
->flags
);
1345 eb_export_fence(obj
, req
, vma
->exec_entry
->flags
);
1350 i915_reset_gen7_sol_offsets(struct drm_i915_gem_request
*req
)
1355 if (!IS_GEN7(req
->i915
) || req
->engine
->id
!= RCS
) {
1356 DRM_DEBUG("sol reset is gen7/rcs only\n");
1360 cs
= intel_ring_begin(req
, 4 * 3);
1364 for (i
= 0; i
< 4; i
++) {
1365 *cs
++ = MI_LOAD_REGISTER_IMM(1);
1366 *cs
++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i
));
1370 intel_ring_advance(req
, cs
);
1375 static struct i915_vma
*
1376 i915_gem_execbuffer_parse(struct intel_engine_cs
*engine
,
1377 struct drm_i915_gem_exec_object2
*shadow_exec_entry
,
1378 struct drm_i915_gem_object
*batch_obj
,
1380 u32 batch_start_offset
,
1384 struct drm_i915_gem_object
*shadow_batch_obj
;
1385 struct i915_vma
*vma
;
1388 shadow_batch_obj
= i915_gem_batch_pool_get(&engine
->batch_pool
,
1389 PAGE_ALIGN(batch_len
));
1390 if (IS_ERR(shadow_batch_obj
))
1391 return ERR_CAST(shadow_batch_obj
);
1393 ret
= intel_engine_cmd_parser(engine
,
1400 if (ret
== -EACCES
) /* unhandled chained batch */
1407 vma
= i915_gem_object_ggtt_pin(shadow_batch_obj
, NULL
, 0, 0, 0);
1411 memset(shadow_exec_entry
, 0, sizeof(*shadow_exec_entry
));
1413 vma
->exec_entry
= shadow_exec_entry
;
1414 vma
->exec_entry
->flags
= __EXEC_OBJECT_HAS_PIN
;
1415 i915_gem_object_get(shadow_batch_obj
);
1416 list_add_tail(&vma
->exec_list
, &eb
->vmas
);
1419 i915_gem_object_unpin_pages(shadow_batch_obj
);
1424 add_to_client(struct drm_i915_gem_request
*req
,
1425 struct drm_file
*file
)
1427 req
->file_priv
= file
->driver_priv
;
1428 list_add_tail(&req
->client_link
, &req
->file_priv
->mm
.request_list
);
1432 execbuf_submit(struct i915_execbuffer_params
*params
,
1433 struct drm_i915_gem_execbuffer2
*args
,
1434 struct list_head
*vmas
)
1436 u64 exec_start
, exec_len
;
1439 ret
= i915_gem_execbuffer_move_to_gpu(params
->request
, vmas
);
1443 ret
= i915_switch_context(params
->request
);
1447 if (args
->flags
& I915_EXEC_CONSTANTS_MASK
) {
1448 DRM_DEBUG("I915_EXEC_CONSTANTS_* unsupported\n");
1452 if (args
->flags
& I915_EXEC_GEN7_SOL_RESET
) {
1453 ret
= i915_reset_gen7_sol_offsets(params
->request
);
1458 exec_len
= args
->batch_len
;
1459 exec_start
= params
->batch
->node
.start
+
1460 params
->args_batch_start_offset
;
1463 exec_len
= params
->batch
->size
- params
->args_batch_start_offset
;
1465 ret
= params
->engine
->emit_bb_start(params
->request
,
1466 exec_start
, exec_len
,
1467 params
->dispatch_flags
);
1471 i915_gem_execbuffer_move_to_active(vmas
, params
->request
);
1477 * Find one BSD ring to dispatch the corresponding BSD command.
1478 * The engine index is returned.
1481 gen8_dispatch_bsd_engine(struct drm_i915_private
*dev_priv
,
1482 struct drm_file
*file
)
1484 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
1486 /* Check whether the file_priv has already selected one ring. */
1487 if ((int)file_priv
->bsd_engine
< 0)
1488 file_priv
->bsd_engine
= atomic_fetch_xor(1,
1489 &dev_priv
->mm
.bsd_engine_dispatch_index
);
1491 return file_priv
->bsd_engine
;
1494 #define I915_USER_RINGS (4)
1496 static const enum intel_engine_id user_ring_map
[I915_USER_RINGS
+ 1] = {
1497 [I915_EXEC_DEFAULT
] = RCS
,
1498 [I915_EXEC_RENDER
] = RCS
,
1499 [I915_EXEC_BLT
] = BCS
,
1500 [I915_EXEC_BSD
] = VCS
,
1501 [I915_EXEC_VEBOX
] = VECS
1504 static struct intel_engine_cs
*
1505 eb_select_engine(struct drm_i915_private
*dev_priv
,
1506 struct drm_file
*file
,
1507 struct drm_i915_gem_execbuffer2
*args
)
1509 unsigned int user_ring_id
= args
->flags
& I915_EXEC_RING_MASK
;
1510 struct intel_engine_cs
*engine
;
1512 if (user_ring_id
> I915_USER_RINGS
) {
1513 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id
);
1517 if ((user_ring_id
!= I915_EXEC_BSD
) &&
1518 ((args
->flags
& I915_EXEC_BSD_MASK
) != 0)) {
1519 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1520 "bsd dispatch flags: %d\n", (int)(args
->flags
));
1524 if (user_ring_id
== I915_EXEC_BSD
&& HAS_BSD2(dev_priv
)) {
1525 unsigned int bsd_idx
= args
->flags
& I915_EXEC_BSD_MASK
;
1527 if (bsd_idx
== I915_EXEC_BSD_DEFAULT
) {
1528 bsd_idx
= gen8_dispatch_bsd_engine(dev_priv
, file
);
1529 } else if (bsd_idx
>= I915_EXEC_BSD_RING1
&&
1530 bsd_idx
<= I915_EXEC_BSD_RING2
) {
1531 bsd_idx
>>= I915_EXEC_BSD_SHIFT
;
1534 DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
1539 engine
= dev_priv
->engine
[_VCS(bsd_idx
)];
1541 engine
= dev_priv
->engine
[user_ring_map
[user_ring_id
]];
1545 DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id
);
1553 i915_gem_do_execbuffer(struct drm_device
*dev
, void *data
,
1554 struct drm_file
*file
,
1555 struct drm_i915_gem_execbuffer2
*args
,
1556 struct drm_i915_gem_exec_object2
*exec
)
1558 struct drm_i915_private
*dev_priv
= to_i915(dev
);
1559 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
1561 struct drm_i915_gem_exec_object2 shadow_exec_entry
;
1562 struct intel_engine_cs
*engine
;
1563 struct i915_gem_context
*ctx
;
1564 struct i915_address_space
*vm
;
1565 struct i915_execbuffer_params params_master
; /* XXX: will be removed later */
1566 struct i915_execbuffer_params
*params
= ¶ms_master
;
1567 const u32 ctx_id
= i915_execbuffer2_get_context_id(*args
);
1569 struct dma_fence
*in_fence
= NULL
;
1570 struct sync_file
*out_fence
= NULL
;
1571 int out_fence_fd
= -1;
1575 if (!i915_gem_check_execbuffer(args
))
1578 ret
= validate_exec_list(dev
, exec
, args
->buffer_count
);
1583 if (args
->flags
& I915_EXEC_SECURE
) {
1584 if (!drm_is_current_master(file
) || !capable(CAP_SYS_ADMIN
))
1587 dispatch_flags
|= I915_DISPATCH_SECURE
;
1589 if (args
->flags
& I915_EXEC_IS_PINNED
)
1590 dispatch_flags
|= I915_DISPATCH_PINNED
;
1592 engine
= eb_select_engine(dev_priv
, file
, args
);
1596 if (args
->buffer_count
< 1) {
1597 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1601 if (args
->flags
& I915_EXEC_RESOURCE_STREAMER
) {
1602 if (!HAS_RESOURCE_STREAMER(dev_priv
)) {
1603 DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1606 if (engine
->id
!= RCS
) {
1607 DRM_DEBUG("RS is not available on %s\n",
1612 dispatch_flags
|= I915_DISPATCH_RS
;
1615 if (args
->flags
& I915_EXEC_FENCE_IN
) {
1616 in_fence
= sync_file_get_fence(lower_32_bits(args
->rsvd2
));
1621 if (args
->flags
& I915_EXEC_FENCE_OUT
) {
1622 out_fence_fd
= get_unused_fd_flags(O_CLOEXEC
);
1623 if (out_fence_fd
< 0) {
1629 /* Take a local wakeref for preparing to dispatch the execbuf as
1630 * we expect to access the hardware fairly frequently in the
1631 * process. Upon first dispatch, we acquire another prolonged
1632 * wakeref that we hold until the GPU has been idle for at least
1635 intel_runtime_pm_get(dev_priv
);
1637 ret
= i915_mutex_lock_interruptible(dev
);
1641 ctx
= i915_gem_validate_context(dev
, file
, engine
, ctx_id
);
1643 mutex_unlock(&dev
->struct_mutex
);
1648 i915_gem_context_get(ctx
);
1651 vm
= &ctx
->ppgtt
->base
;
1655 memset(¶ms_master
, 0x00, sizeof(params_master
));
1657 eb
= eb_create(dev_priv
, args
);
1659 i915_gem_context_put(ctx
);
1660 mutex_unlock(&dev
->struct_mutex
);
1665 /* Look up object handles */
1666 ret
= eb_lookup_vmas(eb
, exec
, args
, vm
, file
);
1670 /* take note of the batch buffer before we might reorder the lists */
1671 params
->batch
= eb_get_batch(eb
);
1673 /* Move the objects en-masse into the GTT, evicting if necessary. */
1674 need_relocs
= (args
->flags
& I915_EXEC_NO_RELOC
) == 0;
1675 ret
= i915_gem_execbuffer_reserve(engine
, &eb
->vmas
, ctx
,
1680 /* The objects are in their final locations, apply the relocations. */
1682 ret
= i915_gem_execbuffer_relocate(eb
);
1684 if (ret
== -EFAULT
) {
1685 ret
= i915_gem_execbuffer_relocate_slow(dev
, args
, file
,
1688 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
1694 /* Set the pending read domains for the batch buffer to COMMAND */
1695 if (params
->batch
->obj
->base
.pending_write_domain
) {
1696 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1700 if (args
->batch_start_offset
> params
->batch
->size
||
1701 args
->batch_len
> params
->batch
->size
- args
->batch_start_offset
) {
1702 DRM_DEBUG("Attempting to use out-of-bounds batch\n");
1707 params
->args_batch_start_offset
= args
->batch_start_offset
;
1708 if (engine
->needs_cmd_parser
&& args
->batch_len
) {
1709 struct i915_vma
*vma
;
1711 vma
= i915_gem_execbuffer_parse(engine
, &shadow_exec_entry
,
1714 args
->batch_start_offset
,
1716 drm_is_current_master(file
));
1724 * Batch parsed and accepted:
1726 * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1727 * bit from MI_BATCH_BUFFER_START commands issued in
1728 * the dispatch_execbuffer implementations. We
1729 * specifically don't want that set on batches the
1730 * command parser has accepted.
1732 dispatch_flags
|= I915_DISPATCH_SECURE
;
1733 params
->args_batch_start_offset
= 0;
1734 params
->batch
= vma
;
1738 params
->batch
->obj
->base
.pending_read_domains
|= I915_GEM_DOMAIN_COMMAND
;
1740 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1741 * batch" bit. Hence we need to pin secure batches into the global gtt.
1742 * hsw should have this fixed, but bdw mucks it up again. */
1743 if (dispatch_flags
& I915_DISPATCH_SECURE
) {
1744 struct drm_i915_gem_object
*obj
= params
->batch
->obj
;
1745 struct i915_vma
*vma
;
1748 * So on first glance it looks freaky that we pin the batch here
1749 * outside of the reservation loop. But:
1750 * - The batch is already pinned into the relevant ppgtt, so we
1751 * already have the backing storage fully allocated.
1752 * - No other BO uses the global gtt (well contexts, but meh),
1753 * so we don't really have issues with multiple objects not
1754 * fitting due to fragmentation.
1755 * So this is actually safe.
1757 vma
= i915_gem_object_ggtt_pin(obj
, NULL
, 0, 0, 0);
1763 params
->batch
= vma
;
1766 /* Allocate a request for this batch buffer nice and early. */
1767 params
->request
= i915_gem_request_alloc(engine
, ctx
);
1768 if (IS_ERR(params
->request
)) {
1769 ret
= PTR_ERR(params
->request
);
1770 goto err_batch_unpin
;
1774 ret
= i915_gem_request_await_dma_fence(params
->request
,
1780 if (out_fence_fd
!= -1) {
1781 out_fence
= sync_file_create(¶ms
->request
->fence
);
1788 /* Whilst this request exists, batch_obj will be on the
1789 * active_list, and so will hold the active reference. Only when this
1790 * request is retired will the the batch_obj be moved onto the
1791 * inactive_list and lose its active reference. Hence we do not need
1792 * to explicitly hold another reference here.
1794 params
->request
->batch
= params
->batch
;
1797 * Save assorted stuff away to pass through to *_submission().
1798 * NB: This data should be 'persistent' and not local as it will
1799 * kept around beyond the duration of the IOCTL once the GPU
1800 * scheduler arrives.
1803 params
->file
= file
;
1804 params
->engine
= engine
;
1805 params
->dispatch_flags
= dispatch_flags
;
1808 trace_i915_gem_request_queue(params
->request
, dispatch_flags
);
1810 ret
= execbuf_submit(params
, args
, &eb
->vmas
);
1812 __i915_add_request(params
->request
, ret
== 0);
1813 add_to_client(params
->request
, file
);
1817 fd_install(out_fence_fd
, out_fence
->file
);
1818 args
->rsvd2
&= GENMASK_ULL(0, 31); /* keep in-fence */
1819 args
->rsvd2
|= (u64
)out_fence_fd
<< 32;
1822 fput(out_fence
->file
);
1828 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1829 * batch vma for correctness. For less ugly and less fragility this
1830 * needs to be adjusted to also track the ggtt batch vma properly as
1833 if (dispatch_flags
& I915_DISPATCH_SECURE
)
1834 i915_vma_unpin(params
->batch
);
1836 /* the request owns the ref now */
1837 i915_gem_context_put(ctx
);
1840 mutex_unlock(&dev
->struct_mutex
);
1843 /* intel_gpu_busy should also get a ref, so it will free when the device
1844 * is really idle. */
1845 intel_runtime_pm_put(dev_priv
);
1846 if (out_fence_fd
!= -1)
1847 put_unused_fd(out_fence_fd
);
1849 dma_fence_put(in_fence
);
1854 * Legacy execbuffer just creates an exec2 list from the original exec object
1855 * list array and passes it to the real function.
1858 i915_gem_execbuffer(struct drm_device
*dev
, void *data
,
1859 struct drm_file
*file
)
1861 struct drm_i915_gem_execbuffer
*args
= data
;
1862 struct drm_i915_gem_execbuffer2 exec2
;
1863 struct drm_i915_gem_exec_object
*exec_list
= NULL
;
1864 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1867 if (args
->buffer_count
< 1) {
1868 DRM_DEBUG("execbuf with %d buffers\n", args
->buffer_count
);
1872 /* Copy in the exec list from userland */
1873 exec_list
= drm_malloc_ab(sizeof(*exec_list
), args
->buffer_count
);
1874 exec2_list
= drm_malloc_ab(sizeof(*exec2_list
), args
->buffer_count
);
1875 if (exec_list
== NULL
|| exec2_list
== NULL
) {
1876 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1877 args
->buffer_count
);
1878 drm_free_large(exec_list
);
1879 drm_free_large(exec2_list
);
1882 ret
= copy_from_user(exec_list
,
1883 u64_to_user_ptr(args
->buffers_ptr
),
1884 sizeof(*exec_list
) * args
->buffer_count
);
1886 DRM_DEBUG("copy %d exec entries failed %d\n",
1887 args
->buffer_count
, ret
);
1888 drm_free_large(exec_list
);
1889 drm_free_large(exec2_list
);
1893 for (i
= 0; i
< args
->buffer_count
; i
++) {
1894 exec2_list
[i
].handle
= exec_list
[i
].handle
;
1895 exec2_list
[i
].relocation_count
= exec_list
[i
].relocation_count
;
1896 exec2_list
[i
].relocs_ptr
= exec_list
[i
].relocs_ptr
;
1897 exec2_list
[i
].alignment
= exec_list
[i
].alignment
;
1898 exec2_list
[i
].offset
= exec_list
[i
].offset
;
1899 if (INTEL_GEN(to_i915(dev
)) < 4)
1900 exec2_list
[i
].flags
= EXEC_OBJECT_NEEDS_FENCE
;
1902 exec2_list
[i
].flags
= 0;
1905 exec2
.buffers_ptr
= args
->buffers_ptr
;
1906 exec2
.buffer_count
= args
->buffer_count
;
1907 exec2
.batch_start_offset
= args
->batch_start_offset
;
1908 exec2
.batch_len
= args
->batch_len
;
1909 exec2
.DR1
= args
->DR1
;
1910 exec2
.DR4
= args
->DR4
;
1911 exec2
.num_cliprects
= args
->num_cliprects
;
1912 exec2
.cliprects_ptr
= args
->cliprects_ptr
;
1913 exec2
.flags
= I915_EXEC_RENDER
;
1914 i915_execbuffer2_set_context_id(exec2
, 0);
1916 ret
= i915_gem_do_execbuffer(dev
, data
, file
, &exec2
, exec2_list
);
1918 struct drm_i915_gem_exec_object __user
*user_exec_list
=
1919 u64_to_user_ptr(args
->buffers_ptr
);
1921 /* Copy the new buffer offsets back to the user's exec list. */
1922 for (i
= 0; i
< args
->buffer_count
; i
++) {
1923 exec2_list
[i
].offset
=
1924 gen8_canonical_addr(exec2_list
[i
].offset
);
1925 ret
= __copy_to_user(&user_exec_list
[i
].offset
,
1926 &exec2_list
[i
].offset
,
1927 sizeof(user_exec_list
[i
].offset
));
1930 DRM_DEBUG("failed to copy %d exec entries "
1931 "back to user (%d)\n",
1932 args
->buffer_count
, ret
);
1938 drm_free_large(exec_list
);
1939 drm_free_large(exec2_list
);
1944 i915_gem_execbuffer2(struct drm_device
*dev
, void *data
,
1945 struct drm_file
*file
)
1947 struct drm_i915_gem_execbuffer2
*args
= data
;
1948 struct drm_i915_gem_exec_object2
*exec2_list
= NULL
;
1951 if (args
->buffer_count
< 1 ||
1952 args
->buffer_count
> UINT_MAX
/ sizeof(*exec2_list
)) {
1953 DRM_DEBUG("execbuf2 with %d buffers\n", args
->buffer_count
);
1957 exec2_list
= drm_malloc_gfp(args
->buffer_count
,
1958 sizeof(*exec2_list
),
1960 if (exec2_list
== NULL
) {
1961 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1962 args
->buffer_count
);
1965 ret
= copy_from_user(exec2_list
,
1966 u64_to_user_ptr(args
->buffers_ptr
),
1967 sizeof(*exec2_list
) * args
->buffer_count
);
1969 DRM_DEBUG("copy %d exec entries failed %d\n",
1970 args
->buffer_count
, ret
);
1971 drm_free_large(exec2_list
);
1975 ret
= i915_gem_do_execbuffer(dev
, data
, file
, args
, exec2_list
);
1977 /* Copy the new buffer offsets back to the user's exec list. */
1978 struct drm_i915_gem_exec_object2 __user
*user_exec_list
=
1979 u64_to_user_ptr(args
->buffers_ptr
);
1982 for (i
= 0; i
< args
->buffer_count
; i
++) {
1983 exec2_list
[i
].offset
=
1984 gen8_canonical_addr(exec2_list
[i
].offset
);
1985 ret
= __copy_to_user(&user_exec_list
[i
].offset
,
1986 &exec2_list
[i
].offset
,
1987 sizeof(user_exec_list
[i
].offset
));
1990 DRM_DEBUG("failed to copy %d exec entries "
1992 args
->buffer_count
);
1998 drm_free_large(exec2_list
);