2 * Copyright 2009 Jerome Glisse.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
32 #include <linux/list.h>
33 #include <linux/slab.h>
35 #include <drm/amdgpu_drm.h>
36 #include <drm/drm_cache.h>
38 #include "amdgpu_trace.h"
39 #include "amdgpu_amdkfd.h"
44 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
45 * represents memory used by driver (VRAM, system memory, etc.). The driver
46 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
47 * to create/destroy/set buffer object which are then managed by the kernel TTM
49 * The interfaces are also used internally by kernel clients, including gfx,
50 * uvd, etc. for kernel managed allocations used by the GPU.
55 * amdgpu_bo_subtract_pin_size - Remove BO from pin_size accounting
57 * @bo: &amdgpu_bo buffer object
59 * This function is called when a BO stops being pinned, and updates the
60 * &amdgpu_device pin_size values accordingly.
62 static void amdgpu_bo_subtract_pin_size(struct amdgpu_bo
*bo
)
64 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
66 if (bo
->tbo
.mem
.mem_type
== TTM_PL_VRAM
) {
67 atomic64_sub(amdgpu_bo_size(bo
), &adev
->vram_pin_size
);
68 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo
),
69 &adev
->visible_pin_size
);
70 } else if (bo
->tbo
.mem
.mem_type
== TTM_PL_TT
) {
71 atomic64_sub(amdgpu_bo_size(bo
), &adev
->gart_pin_size
);
75 static void amdgpu_bo_destroy(struct ttm_buffer_object
*tbo
)
77 struct amdgpu_device
*adev
= amdgpu_ttm_adev(tbo
->bdev
);
78 struct amdgpu_bo
*bo
= ttm_to_amdgpu_bo(tbo
);
80 if (bo
->pin_count
> 0)
81 amdgpu_bo_subtract_pin_size(bo
);
85 if (bo
->tbo
.base
.import_attach
)
86 drm_prime_gem_destroy(&bo
->tbo
.base
, bo
->tbo
.sg
);
87 drm_gem_object_release(&bo
->tbo
.base
);
88 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
89 if (!list_empty(&bo
->shadow_list
)) {
90 mutex_lock(&adev
->shadow_list_lock
);
91 list_del_init(&bo
->shadow_list
);
92 mutex_unlock(&adev
->shadow_list_lock
);
94 amdgpu_bo_unref(&bo
->parent
);
101 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
102 * @bo: buffer object to be checked
104 * Uses destroy function associated with the object to determine if this is
108 * true if the object belongs to &amdgpu_bo, false if not.
110 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object
*bo
)
112 if (bo
->destroy
== &amdgpu_bo_destroy
)
118 * amdgpu_bo_placement_from_domain - set buffer's placement
119 * @abo: &amdgpu_bo buffer object whose placement is to be set
120 * @domain: requested domain
122 * Sets buffer's placement according to requested domain and the buffer's
125 void amdgpu_bo_placement_from_domain(struct amdgpu_bo
*abo
, u32 domain
)
127 struct amdgpu_device
*adev
= amdgpu_ttm_adev(abo
->tbo
.bdev
);
128 struct ttm_placement
*placement
= &abo
->placement
;
129 struct ttm_place
*places
= abo
->placements
;
130 u64 flags
= abo
->flags
;
133 if (domain
& AMDGPU_GEM_DOMAIN_VRAM
) {
134 unsigned visible_pfn
= adev
->gmc
.visible_vram_size
>> PAGE_SHIFT
;
138 places
[c
].flags
= TTM_PL_FLAG_WC
| TTM_PL_FLAG_UNCACHED
|
141 if (flags
& AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
)
142 places
[c
].lpfn
= visible_pfn
;
144 places
[c
].flags
|= TTM_PL_FLAG_TOPDOWN
;
146 if (flags
& AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS
)
147 places
[c
].flags
|= TTM_PL_FLAG_CONTIGUOUS
;
151 if (domain
& AMDGPU_GEM_DOMAIN_GTT
) {
154 places
[c
].flags
= TTM_PL_FLAG_TT
;
155 if (flags
& AMDGPU_GEM_CREATE_CPU_GTT_USWC
)
156 places
[c
].flags
|= TTM_PL_FLAG_WC
|
157 TTM_PL_FLAG_UNCACHED
;
159 places
[c
].flags
|= TTM_PL_FLAG_CACHED
;
163 if (domain
& AMDGPU_GEM_DOMAIN_CPU
) {
166 places
[c
].flags
= TTM_PL_FLAG_SYSTEM
;
167 if (flags
& AMDGPU_GEM_CREATE_CPU_GTT_USWC
)
168 places
[c
].flags
|= TTM_PL_FLAG_WC
|
169 TTM_PL_FLAG_UNCACHED
;
171 places
[c
].flags
|= TTM_PL_FLAG_CACHED
;
175 if (domain
& AMDGPU_GEM_DOMAIN_GDS
) {
178 places
[c
].flags
= TTM_PL_FLAG_UNCACHED
| AMDGPU_PL_FLAG_GDS
;
182 if (domain
& AMDGPU_GEM_DOMAIN_GWS
) {
185 places
[c
].flags
= TTM_PL_FLAG_UNCACHED
| AMDGPU_PL_FLAG_GWS
;
189 if (domain
& AMDGPU_GEM_DOMAIN_OA
) {
192 places
[c
].flags
= TTM_PL_FLAG_UNCACHED
| AMDGPU_PL_FLAG_OA
;
199 places
[c
].flags
= TTM_PL_MASK_CACHING
| TTM_PL_FLAG_SYSTEM
;
203 BUG_ON(c
>= AMDGPU_BO_MAX_PLACEMENTS
);
205 placement
->num_placement
= c
;
206 placement
->placement
= places
;
208 placement
->num_busy_placement
= c
;
209 placement
->busy_placement
= places
;
213 * amdgpu_bo_create_reserved - create reserved BO for kernel use
215 * @adev: amdgpu device object
216 * @size: size for the new BO
217 * @align: alignment for the new BO
218 * @domain: where to place it
219 * @bo_ptr: used to initialize BOs in structures
220 * @gpu_addr: GPU addr of the pinned BO
221 * @cpu_addr: optional CPU address mapping
223 * Allocates and pins a BO for kernel internal use, and returns it still
226 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
229 * 0 on success, negative error code otherwise.
231 int amdgpu_bo_create_reserved(struct amdgpu_device
*adev
,
232 unsigned long size
, int align
,
233 u32 domain
, struct amdgpu_bo
**bo_ptr
,
234 u64
*gpu_addr
, void **cpu_addr
)
236 struct amdgpu_bo_param bp
;
241 amdgpu_bo_unref(bo_ptr
);
245 memset(&bp
, 0, sizeof(bp
));
247 bp
.byte_align
= align
;
249 bp
.flags
= cpu_addr
? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
250 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS
;
251 bp
.flags
|= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS
;
252 bp
.type
= ttm_bo_type_kernel
;
256 r
= amdgpu_bo_create(adev
, &bp
, bo_ptr
);
258 dev_err(adev
->dev
, "(%d) failed to allocate kernel bo\n",
265 r
= amdgpu_bo_reserve(*bo_ptr
, false);
267 dev_err(adev
->dev
, "(%d) failed to reserve kernel bo\n", r
);
271 r
= amdgpu_bo_pin(*bo_ptr
, domain
);
273 dev_err(adev
->dev
, "(%d) kernel bo pin failed\n", r
);
274 goto error_unreserve
;
277 r
= amdgpu_ttm_alloc_gart(&(*bo_ptr
)->tbo
);
279 dev_err(adev
->dev
, "%p bind failed\n", *bo_ptr
);
284 *gpu_addr
= amdgpu_bo_gpu_offset(*bo_ptr
);
287 r
= amdgpu_bo_kmap(*bo_ptr
, cpu_addr
);
289 dev_err(adev
->dev
, "(%d) kernel bo map failed\n", r
);
297 amdgpu_bo_unpin(*bo_ptr
);
299 amdgpu_bo_unreserve(*bo_ptr
);
303 amdgpu_bo_unref(bo_ptr
);
309 * amdgpu_bo_create_kernel - create BO for kernel use
311 * @adev: amdgpu device object
312 * @size: size for the new BO
313 * @align: alignment for the new BO
314 * @domain: where to place it
315 * @bo_ptr: used to initialize BOs in structures
316 * @gpu_addr: GPU addr of the pinned BO
317 * @cpu_addr: optional CPU address mapping
319 * Allocates and pins a BO for kernel internal use.
321 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
324 * 0 on success, negative error code otherwise.
326 int amdgpu_bo_create_kernel(struct amdgpu_device
*adev
,
327 unsigned long size
, int align
,
328 u32 domain
, struct amdgpu_bo
**bo_ptr
,
329 u64
*gpu_addr
, void **cpu_addr
)
333 r
= amdgpu_bo_create_reserved(adev
, size
, align
, domain
, bo_ptr
,
340 amdgpu_bo_unreserve(*bo_ptr
);
346 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
348 * @adev: amdgpu device object
349 * @offset: offset of the BO
350 * @size: size of the BO
351 * @domain: where to place it
352 * @bo_ptr: used to initialize BOs in structures
353 * @cpu_addr: optional CPU address mapping
355 * Creates a kernel BO at a specific offset in the address space of the domain.
358 * 0 on success, negative error code otherwise.
360 int amdgpu_bo_create_kernel_at(struct amdgpu_device
*adev
,
361 uint64_t offset
, uint64_t size
, uint32_t domain
,
362 struct amdgpu_bo
**bo_ptr
, void **cpu_addr
)
364 struct ttm_operation_ctx ctx
= { false, false };
369 size
= ALIGN(size
, PAGE_SIZE
);
371 r
= amdgpu_bo_create_reserved(adev
, size
, PAGE_SIZE
, domain
, bo_ptr
,
377 * Remove the original mem node and create a new one at the request
381 amdgpu_bo_kunmap(*bo_ptr
);
383 ttm_bo_mem_put(&(*bo_ptr
)->tbo
, &(*bo_ptr
)->tbo
.mem
);
385 for (i
= 0; i
< (*bo_ptr
)->placement
.num_placement
; ++i
) {
386 (*bo_ptr
)->placements
[i
].fpfn
= offset
>> PAGE_SHIFT
;
387 (*bo_ptr
)->placements
[i
].lpfn
= (offset
+ size
) >> PAGE_SHIFT
;
389 r
= ttm_bo_mem_space(&(*bo_ptr
)->tbo
, &(*bo_ptr
)->placement
,
390 &(*bo_ptr
)->tbo
.mem
, &ctx
);
395 r
= amdgpu_bo_kmap(*bo_ptr
, cpu_addr
);
400 amdgpu_bo_unreserve(*bo_ptr
);
404 amdgpu_bo_unreserve(*bo_ptr
);
405 amdgpu_bo_unref(bo_ptr
);
410 * amdgpu_bo_free_kernel - free BO for kernel use
412 * @bo: amdgpu BO to free
413 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
414 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
416 * unmaps and unpin a BO for kernel internal use.
418 void amdgpu_bo_free_kernel(struct amdgpu_bo
**bo
, u64
*gpu_addr
,
424 if (likely(amdgpu_bo_reserve(*bo
, true) == 0)) {
426 amdgpu_bo_kunmap(*bo
);
428 amdgpu_bo_unpin(*bo
);
429 amdgpu_bo_unreserve(*bo
);
440 /* Validate bo size is bit bigger then the request domain */
441 static bool amdgpu_bo_validate_size(struct amdgpu_device
*adev
,
442 unsigned long size
, u32 domain
)
444 struct ttm_mem_type_manager
*man
= NULL
;
447 * If GTT is part of requested domains the check must succeed to
448 * allow fall back to GTT
450 if (domain
& AMDGPU_GEM_DOMAIN_GTT
) {
451 man
= &adev
->mman
.bdev
.man
[TTM_PL_TT
];
453 if (size
< (man
->size
<< PAGE_SHIFT
))
459 if (domain
& AMDGPU_GEM_DOMAIN_VRAM
) {
460 man
= &adev
->mman
.bdev
.man
[TTM_PL_VRAM
];
462 if (size
< (man
->size
<< PAGE_SHIFT
))
469 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
473 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size
,
474 man
->size
<< PAGE_SHIFT
);
478 bool amdgpu_bo_support_uswc(u64 bo_flags
)
482 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
483 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
486 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
487 /* Don't try to enable write-combining when it can't work, or things
489 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
492 #ifndef CONFIG_COMPILE_TEST
493 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
494 thanks to write-combining
497 if (bo_flags
& AMDGPU_GEM_CREATE_CPU_GTT_USWC
)
498 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
499 "better performance thanks to write-combining\n");
502 /* For architectures that don't support WC memory,
503 * mask out the WC flag from the BO
505 if (!drm_arch_can_wc_memory())
512 static int amdgpu_bo_do_create(struct amdgpu_device
*adev
,
513 struct amdgpu_bo_param
*bp
,
514 struct amdgpu_bo
**bo_ptr
)
516 struct ttm_operation_ctx ctx
= {
517 .interruptible
= (bp
->type
!= ttm_bo_type_kernel
),
518 .no_wait_gpu
= bp
->no_wait_gpu
,
520 .flags
= bp
->type
!= ttm_bo_type_kernel
?
521 TTM_OPT_FLAG_ALLOW_RES_EVICT
: 0
523 struct amdgpu_bo
*bo
;
524 unsigned long page_align
, size
= bp
->size
;
528 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
529 if (bp
->domain
& (AMDGPU_GEM_DOMAIN_GWS
| AMDGPU_GEM_DOMAIN_OA
)) {
530 /* GWS and OA don't need any alignment. */
531 page_align
= bp
->byte_align
;
533 } else if (bp
->domain
& AMDGPU_GEM_DOMAIN_GDS
) {
534 /* Both size and alignment must be a multiple of 4. */
535 page_align
= ALIGN(bp
->byte_align
, 4);
536 size
= ALIGN(size
, 4) << PAGE_SHIFT
;
538 /* Memory should be aligned at least to a page size. */
539 page_align
= ALIGN(bp
->byte_align
, PAGE_SIZE
) >> PAGE_SHIFT
;
540 size
= ALIGN(size
, PAGE_SIZE
);
543 if (!amdgpu_bo_validate_size(adev
, size
, bp
->domain
))
548 acc_size
= ttm_bo_dma_acc_size(&adev
->mman
.bdev
, size
,
549 sizeof(struct amdgpu_bo
));
551 bo
= kzalloc(sizeof(struct amdgpu_bo
), GFP_KERNEL
);
554 drm_gem_private_object_init(adev
->ddev
, &bo
->tbo
.base
, size
);
555 INIT_LIST_HEAD(&bo
->shadow_list
);
557 bo
->preferred_domains
= bp
->preferred_domain
? bp
->preferred_domain
:
559 bo
->allowed_domains
= bo
->preferred_domains
;
560 if (bp
->type
!= ttm_bo_type_kernel
&&
561 bo
->allowed_domains
== AMDGPU_GEM_DOMAIN_VRAM
)
562 bo
->allowed_domains
|= AMDGPU_GEM_DOMAIN_GTT
;
564 bo
->flags
= bp
->flags
;
566 if (!amdgpu_bo_support_uswc(bo
->flags
))
567 bo
->flags
&= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC
;
569 bo
->tbo
.bdev
= &adev
->mman
.bdev
;
570 if (bp
->domain
& (AMDGPU_GEM_DOMAIN_GWS
| AMDGPU_GEM_DOMAIN_OA
|
571 AMDGPU_GEM_DOMAIN_GDS
))
572 amdgpu_bo_placement_from_domain(bo
, AMDGPU_GEM_DOMAIN_CPU
);
574 amdgpu_bo_placement_from_domain(bo
, bp
->domain
);
575 if (bp
->type
== ttm_bo_type_kernel
)
576 bo
->tbo
.priority
= 1;
578 r
= ttm_bo_init_reserved(&adev
->mman
.bdev
, &bo
->tbo
, size
, bp
->type
,
579 &bo
->placement
, page_align
, &ctx
, acc_size
,
580 NULL
, bp
->resv
, &amdgpu_bo_destroy
);
581 if (unlikely(r
!= 0))
584 if (!amdgpu_gmc_vram_full_visible(&adev
->gmc
) &&
585 bo
->tbo
.mem
.mem_type
== TTM_PL_VRAM
&&
586 bo
->tbo
.mem
.start
< adev
->gmc
.visible_vram_size
>> PAGE_SHIFT
)
587 amdgpu_cs_report_moved_bytes(adev
, ctx
.bytes_moved
,
590 amdgpu_cs_report_moved_bytes(adev
, ctx
.bytes_moved
, 0);
592 if (bp
->flags
& AMDGPU_GEM_CREATE_VRAM_CLEARED
&&
593 bo
->tbo
.mem
.placement
& TTM_PL_FLAG_VRAM
) {
594 struct dma_fence
*fence
;
596 r
= amdgpu_fill_buffer(bo
, 0, bo
->tbo
.base
.resv
, &fence
);
600 amdgpu_bo_fence(bo
, fence
, false);
601 dma_fence_put(bo
->tbo
.moving
);
602 bo
->tbo
.moving
= dma_fence_get(fence
);
603 dma_fence_put(fence
);
606 amdgpu_bo_unreserve(bo
);
609 trace_amdgpu_bo_create(bo
);
611 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
612 if (bp
->type
== ttm_bo_type_device
)
613 bo
->flags
&= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
;
619 dma_resv_unlock(bo
->tbo
.base
.resv
);
620 amdgpu_bo_unref(&bo
);
624 static int amdgpu_bo_create_shadow(struct amdgpu_device
*adev
,
626 struct amdgpu_bo
*bo
)
628 struct amdgpu_bo_param bp
;
634 memset(&bp
, 0, sizeof(bp
));
636 bp
.domain
= AMDGPU_GEM_DOMAIN_GTT
;
637 bp
.flags
= AMDGPU_GEM_CREATE_CPU_GTT_USWC
|
638 AMDGPU_GEM_CREATE_SHADOW
;
639 bp
.type
= ttm_bo_type_kernel
;
640 bp
.resv
= bo
->tbo
.base
.resv
;
642 r
= amdgpu_bo_do_create(adev
, &bp
, &bo
->shadow
);
644 bo
->shadow
->parent
= amdgpu_bo_ref(bo
);
645 mutex_lock(&adev
->shadow_list_lock
);
646 list_add_tail(&bo
->shadow
->shadow_list
, &adev
->shadow_list
);
647 mutex_unlock(&adev
->shadow_list_lock
);
654 * amdgpu_bo_create - create an &amdgpu_bo buffer object
655 * @adev: amdgpu device object
656 * @bp: parameters to be used for the buffer object
657 * @bo_ptr: pointer to the buffer object pointer
659 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
661 * Shadow object is used to backup the original buffer object, and is always
665 * 0 for success or a negative error code on failure.
667 int amdgpu_bo_create(struct amdgpu_device
*adev
,
668 struct amdgpu_bo_param
*bp
,
669 struct amdgpu_bo
**bo_ptr
)
671 u64 flags
= bp
->flags
;
674 bp
->flags
= bp
->flags
& ~AMDGPU_GEM_CREATE_SHADOW
;
675 r
= amdgpu_bo_do_create(adev
, bp
, bo_ptr
);
679 if ((flags
& AMDGPU_GEM_CREATE_SHADOW
) && !(adev
->flags
& AMD_IS_APU
)) {
681 WARN_ON(dma_resv_lock((*bo_ptr
)->tbo
.base
.resv
,
684 r
= amdgpu_bo_create_shadow(adev
, bp
->size
, *bo_ptr
);
687 dma_resv_unlock((*bo_ptr
)->tbo
.base
.resv
);
690 amdgpu_bo_unref(bo_ptr
);
697 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
698 * @bo: pointer to the buffer object
700 * Sets placement according to domain; and changes placement and caching
701 * policy of the buffer object according to the placement.
702 * This is used for validating shadow bos. It calls ttm_bo_validate() to
703 * make sure the buffer is resident where it needs to be.
706 * 0 for success or a negative error code on failure.
708 int amdgpu_bo_validate(struct amdgpu_bo
*bo
)
710 struct ttm_operation_ctx ctx
= { false, false };
717 domain
= bo
->preferred_domains
;
720 amdgpu_bo_placement_from_domain(bo
, domain
);
721 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, &ctx
);
722 if (unlikely(r
== -ENOMEM
) && domain
!= bo
->allowed_domains
) {
723 domain
= bo
->allowed_domains
;
731 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
733 * @shadow: &amdgpu_bo shadow to be restored
734 * @fence: dma_fence associated with the operation
736 * Copies a buffer object's shadow content back to the object.
737 * This is used for recovering a buffer from its shadow in case of a gpu
738 * reset where vram context may be lost.
741 * 0 for success or a negative error code on failure.
743 int amdgpu_bo_restore_shadow(struct amdgpu_bo
*shadow
, struct dma_fence
**fence
)
746 struct amdgpu_device
*adev
= amdgpu_ttm_adev(shadow
->tbo
.bdev
);
747 struct amdgpu_ring
*ring
= adev
->mman
.buffer_funcs_ring
;
748 uint64_t shadow_addr
, parent_addr
;
750 shadow_addr
= amdgpu_bo_gpu_offset(shadow
);
751 parent_addr
= amdgpu_bo_gpu_offset(shadow
->parent
);
753 return amdgpu_copy_buffer(ring
, shadow_addr
, parent_addr
,
754 amdgpu_bo_size(shadow
), NULL
, fence
,
759 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
760 * @bo: &amdgpu_bo buffer object to be mapped
761 * @ptr: kernel virtual address to be returned
763 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
764 * amdgpu_bo_kptr() to get the kernel virtual address.
767 * 0 for success or a negative error code on failure.
769 int amdgpu_bo_kmap(struct amdgpu_bo
*bo
, void **ptr
)
774 if (bo
->flags
& AMDGPU_GEM_CREATE_NO_CPU_ACCESS
)
777 kptr
= amdgpu_bo_kptr(bo
);
784 r
= dma_resv_wait_timeout_rcu(bo
->tbo
.base
.resv
, false, false,
785 MAX_SCHEDULE_TIMEOUT
);
789 r
= ttm_bo_kmap(&bo
->tbo
, 0, bo
->tbo
.num_pages
, &bo
->kmap
);
794 *ptr
= amdgpu_bo_kptr(bo
);
800 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
801 * @bo: &amdgpu_bo buffer object
803 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
806 * the virtual address of a buffer object area.
808 void *amdgpu_bo_kptr(struct amdgpu_bo
*bo
)
812 return ttm_kmap_obj_virtual(&bo
->kmap
, &is_iomem
);
816 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
817 * @bo: &amdgpu_bo buffer object to be unmapped
819 * Unmaps a kernel map set up by amdgpu_bo_kmap().
821 void amdgpu_bo_kunmap(struct amdgpu_bo
*bo
)
824 ttm_bo_kunmap(&bo
->kmap
);
828 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
829 * @bo: &amdgpu_bo buffer object
831 * References the contained &ttm_buffer_object.
834 * a refcounted pointer to the &amdgpu_bo buffer object.
836 struct amdgpu_bo
*amdgpu_bo_ref(struct amdgpu_bo
*bo
)
841 ttm_bo_get(&bo
->tbo
);
846 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
847 * @bo: &amdgpu_bo buffer object
849 * Unreferences the contained &ttm_buffer_object and clear the pointer
851 void amdgpu_bo_unref(struct amdgpu_bo
**bo
)
853 struct ttm_buffer_object
*tbo
;
864 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
865 * @bo: &amdgpu_bo buffer object to be pinned
866 * @domain: domain to be pinned to
867 * @min_offset: the start of requested address range
868 * @max_offset: the end of requested address range
870 * Pins the buffer object according to requested domain and address range. If
871 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
872 * pin_count and pin_size accordingly.
874 * Pinning means to lock pages in memory along with keeping them at a fixed
875 * offset. It is required when a buffer can not be moved, for example, when
876 * a display buffer is being scanned out.
878 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
879 * where to pin a buffer if there are specific restrictions on where a buffer
883 * 0 for success or a negative error code on failure.
885 int amdgpu_bo_pin_restricted(struct amdgpu_bo
*bo
, u32 domain
,
886 u64 min_offset
, u64 max_offset
)
888 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
889 struct ttm_operation_ctx ctx
= { false, false };
892 if (amdgpu_ttm_tt_get_usermm(bo
->tbo
.ttm
))
895 if (WARN_ON_ONCE(min_offset
> max_offset
))
898 /* A shared bo cannot be migrated to VRAM */
899 if (bo
->prime_shared_count
) {
900 if (domain
& AMDGPU_GEM_DOMAIN_GTT
)
901 domain
= AMDGPU_GEM_DOMAIN_GTT
;
906 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
907 * See function amdgpu_display_supported_domains()
909 domain
= amdgpu_bo_get_preferred_pin_domain(adev
, domain
);
912 uint32_t mem_type
= bo
->tbo
.mem
.mem_type
;
914 if (!(domain
& amdgpu_mem_type_to_domain(mem_type
)))
919 if (max_offset
!= 0) {
920 u64 domain_start
= bo
->tbo
.bdev
->man
[mem_type
].gpu_offset
;
921 WARN_ON_ONCE(max_offset
<
922 (amdgpu_bo_gpu_offset(bo
) - domain_start
));
928 bo
->flags
|= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS
;
929 /* force to pin into visible video ram */
930 if (!(bo
->flags
& AMDGPU_GEM_CREATE_NO_CPU_ACCESS
))
931 bo
->flags
|= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
;
932 amdgpu_bo_placement_from_domain(bo
, domain
);
933 for (i
= 0; i
< bo
->placement
.num_placement
; i
++) {
936 fpfn
= min_offset
>> PAGE_SHIFT
;
937 lpfn
= max_offset
>> PAGE_SHIFT
;
939 if (fpfn
> bo
->placements
[i
].fpfn
)
940 bo
->placements
[i
].fpfn
= fpfn
;
941 if (!bo
->placements
[i
].lpfn
||
942 (lpfn
&& lpfn
< bo
->placements
[i
].lpfn
))
943 bo
->placements
[i
].lpfn
= lpfn
;
944 bo
->placements
[i
].flags
|= TTM_PL_FLAG_NO_EVICT
;
947 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, &ctx
);
949 dev_err(adev
->dev
, "%p pin failed\n", bo
);
955 domain
= amdgpu_mem_type_to_domain(bo
->tbo
.mem
.mem_type
);
956 if (domain
== AMDGPU_GEM_DOMAIN_VRAM
) {
957 atomic64_add(amdgpu_bo_size(bo
), &adev
->vram_pin_size
);
958 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo
),
959 &adev
->visible_pin_size
);
960 } else if (domain
== AMDGPU_GEM_DOMAIN_GTT
) {
961 atomic64_add(amdgpu_bo_size(bo
), &adev
->gart_pin_size
);
969 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
970 * @bo: &amdgpu_bo buffer object to be pinned
971 * @domain: domain to be pinned to
973 * A simple wrapper to amdgpu_bo_pin_restricted().
974 * Provides a simpler API for buffers that do not have any strict restrictions
975 * on where a buffer must be located.
978 * 0 for success or a negative error code on failure.
980 int amdgpu_bo_pin(struct amdgpu_bo
*bo
, u32 domain
)
982 return amdgpu_bo_pin_restricted(bo
, domain
, 0, 0);
986 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
987 * @bo: &amdgpu_bo buffer object to be unpinned
989 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
990 * Changes placement and pin size accordingly.
993 * 0 for success or a negative error code on failure.
995 int amdgpu_bo_unpin(struct amdgpu_bo
*bo
)
997 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
998 struct ttm_operation_ctx ctx
= { false, false };
1001 if (WARN_ON_ONCE(!bo
->pin_count
)) {
1002 dev_warn(adev
->dev
, "%p unpin not necessary\n", bo
);
1009 amdgpu_bo_subtract_pin_size(bo
);
1011 for (i
= 0; i
< bo
->placement
.num_placement
; i
++) {
1012 bo
->placements
[i
].lpfn
= 0;
1013 bo
->placements
[i
].flags
&= ~TTM_PL_FLAG_NO_EVICT
;
1015 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, &ctx
);
1017 dev_err(adev
->dev
, "%p validate failed for unpin\n", bo
);
1023 * amdgpu_bo_evict_vram - evict VRAM buffers
1024 * @adev: amdgpu device object
1026 * Evicts all VRAM buffers on the lru list of the memory type.
1027 * Mainly used for evicting vram at suspend time.
1030 * 0 for success or a negative error code on failure.
1032 int amdgpu_bo_evict_vram(struct amdgpu_device
*adev
)
1034 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
1035 #ifndef CONFIG_HIBERNATION
1036 if (adev
->flags
& AMD_IS_APU
) {
1037 /* Useless to evict on IGP chips */
1041 return ttm_bo_evict_mm(&adev
->mman
.bdev
, TTM_PL_VRAM
);
1044 static const char *amdgpu_vram_names
[] = {
1058 * amdgpu_bo_init - initialize memory manager
1059 * @adev: amdgpu device object
1061 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1064 * 0 for success or a negative error code on failure.
1066 int amdgpu_bo_init(struct amdgpu_device
*adev
)
1068 /* reserve PAT memory space to WC for VRAM */
1069 arch_io_reserve_memtype_wc(adev
->gmc
.aper_base
,
1070 adev
->gmc
.aper_size
);
1072 /* Add an MTRR for the VRAM */
1073 adev
->gmc
.vram_mtrr
= arch_phys_wc_add(adev
->gmc
.aper_base
,
1074 adev
->gmc
.aper_size
);
1075 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1076 adev
->gmc
.mc_vram_size
>> 20,
1077 (unsigned long long)adev
->gmc
.aper_size
>> 20);
1078 DRM_INFO("RAM width %dbits %s\n",
1079 adev
->gmc
.vram_width
, amdgpu_vram_names
[adev
->gmc
.vram_type
]);
1080 return amdgpu_ttm_init(adev
);
1084 * amdgpu_bo_late_init - late init
1085 * @adev: amdgpu device object
1087 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1091 * 0 for success or a negative error code on failure.
1093 int amdgpu_bo_late_init(struct amdgpu_device
*adev
)
1095 amdgpu_ttm_late_init(adev
);
1101 * amdgpu_bo_fini - tear down memory manager
1102 * @adev: amdgpu device object
1104 * Reverses amdgpu_bo_init() to tear down memory manager.
1106 void amdgpu_bo_fini(struct amdgpu_device
*adev
)
1108 amdgpu_ttm_fini(adev
);
1109 arch_phys_wc_del(adev
->gmc
.vram_mtrr
);
1110 arch_io_free_memtype_wc(adev
->gmc
.aper_base
, adev
->gmc
.aper_size
);
1114 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1115 * @bo: &amdgpu_bo buffer object
1116 * @vma: vma as input from the fbdev mmap method
1118 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1121 * 0 for success or a negative error code on failure.
1123 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo
*bo
,
1124 struct vm_area_struct
*vma
)
1126 if (vma
->vm_pgoff
!= 0)
1129 return ttm_bo_mmap_obj(vma
, &bo
->tbo
);
1133 * amdgpu_bo_set_tiling_flags - set tiling flags
1134 * @bo: &amdgpu_bo buffer object
1135 * @tiling_flags: new flags
1137 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1138 * kernel driver to set the tiling flags on a buffer.
1141 * 0 for success or a negative error code on failure.
1143 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo
*bo
, u64 tiling_flags
)
1145 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
1147 if (adev
->family
<= AMDGPU_FAMILY_CZ
&&
1148 AMDGPU_TILING_GET(tiling_flags
, TILE_SPLIT
) > 6)
1151 bo
->tiling_flags
= tiling_flags
;
1156 * amdgpu_bo_get_tiling_flags - get tiling flags
1157 * @bo: &amdgpu_bo buffer object
1158 * @tiling_flags: returned flags
1160 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1161 * set the tiling flags on a buffer.
1163 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo
*bo
, u64
*tiling_flags
)
1165 dma_resv_assert_held(bo
->tbo
.base
.resv
);
1168 *tiling_flags
= bo
->tiling_flags
;
1172 * amdgpu_bo_set_metadata - set metadata
1173 * @bo: &amdgpu_bo buffer object
1174 * @metadata: new metadata
1175 * @metadata_size: size of the new metadata
1176 * @flags: flags of the new metadata
1178 * Sets buffer object's metadata, its size and flags.
1179 * Used via GEM ioctl.
1182 * 0 for success or a negative error code on failure.
1184 int amdgpu_bo_set_metadata (struct amdgpu_bo
*bo
, void *metadata
,
1185 uint32_t metadata_size
, uint64_t flags
)
1189 if (!metadata_size
) {
1190 if (bo
->metadata_size
) {
1191 kfree(bo
->metadata
);
1192 bo
->metadata
= NULL
;
1193 bo
->metadata_size
= 0;
1198 if (metadata
== NULL
)
1201 buffer
= kmemdup(metadata
, metadata_size
, GFP_KERNEL
);
1205 kfree(bo
->metadata
);
1206 bo
->metadata_flags
= flags
;
1207 bo
->metadata
= buffer
;
1208 bo
->metadata_size
= metadata_size
;
1214 * amdgpu_bo_get_metadata - get metadata
1215 * @bo: &amdgpu_bo buffer object
1216 * @buffer: returned metadata
1217 * @buffer_size: size of the buffer
1218 * @metadata_size: size of the returned metadata
1219 * @flags: flags of the returned metadata
1221 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1222 * less than metadata_size.
1223 * Used via GEM ioctl.
1226 * 0 for success or a negative error code on failure.
1228 int amdgpu_bo_get_metadata(struct amdgpu_bo
*bo
, void *buffer
,
1229 size_t buffer_size
, uint32_t *metadata_size
,
1232 if (!buffer
&& !metadata_size
)
1236 if (buffer_size
< bo
->metadata_size
)
1239 if (bo
->metadata_size
)
1240 memcpy(buffer
, bo
->metadata
, bo
->metadata_size
);
1244 *metadata_size
= bo
->metadata_size
;
1246 *flags
= bo
->metadata_flags
;
1252 * amdgpu_bo_move_notify - notification about a memory move
1253 * @bo: pointer to a buffer object
1254 * @evict: if this move is evicting the buffer from the graphics address space
1255 * @new_mem: new information of the bufer object
1257 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1259 * TTM driver callback which is called when ttm moves a buffer.
1261 void amdgpu_bo_move_notify(struct ttm_buffer_object
*bo
,
1263 struct ttm_mem_reg
*new_mem
)
1265 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->bdev
);
1266 struct amdgpu_bo
*abo
;
1267 struct ttm_mem_reg
*old_mem
= &bo
->mem
;
1269 if (!amdgpu_bo_is_amdgpu_bo(bo
))
1272 abo
= ttm_to_amdgpu_bo(bo
);
1273 amdgpu_vm_bo_invalidate(adev
, abo
, evict
);
1275 amdgpu_bo_kunmap(abo
);
1277 /* remember the eviction */
1279 atomic64_inc(&adev
->num_evictions
);
1281 /* update statistics */
1285 /* move_notify is called before move happens */
1286 trace_amdgpu_bo_move(abo
, new_mem
->mem_type
, old_mem
->mem_type
);
1290 * amdgpu_bo_move_notify - notification about a BO being released
1291 * @bo: pointer to a buffer object
1293 * Wipes VRAM buffers whose contents should not be leaked before the
1294 * memory is released.
1296 void amdgpu_bo_release_notify(struct ttm_buffer_object
*bo
)
1298 struct dma_fence
*fence
= NULL
;
1299 struct amdgpu_bo
*abo
;
1302 if (!amdgpu_bo_is_amdgpu_bo(bo
))
1305 abo
= ttm_to_amdgpu_bo(bo
);
1308 amdgpu_amdkfd_unreserve_memory_limit(abo
);
1310 if (bo
->mem
.mem_type
!= TTM_PL_VRAM
|| !bo
->mem
.mm_node
||
1311 !(abo
->flags
& AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE
))
1314 dma_resv_lock(bo
->base
.resv
, NULL
);
1316 r
= amdgpu_fill_buffer(abo
, AMDGPU_POISON
, bo
->base
.resv
, &fence
);
1318 amdgpu_bo_fence(abo
, fence
, false);
1319 dma_fence_put(fence
);
1322 dma_resv_unlock(bo
->base
.resv
);
1326 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1327 * @bo: pointer to a buffer object
1329 * Notifies the driver we are taking a fault on this BO and have reserved it,
1330 * also performs bookkeeping.
1331 * TTM driver callback for dealing with vm faults.
1334 * 0 for success or a negative error code on failure.
1336 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object
*bo
)
1338 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->bdev
);
1339 struct ttm_operation_ctx ctx
= { false, false };
1340 struct amdgpu_bo
*abo
;
1341 unsigned long offset
, size
;
1344 if (!amdgpu_bo_is_amdgpu_bo(bo
))
1347 abo
= ttm_to_amdgpu_bo(bo
);
1349 /* Remember that this BO was accessed by the CPU */
1350 abo
->flags
|= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
;
1352 if (bo
->mem
.mem_type
!= TTM_PL_VRAM
)
1355 size
= bo
->mem
.num_pages
<< PAGE_SHIFT
;
1356 offset
= bo
->mem
.start
<< PAGE_SHIFT
;
1357 if ((offset
+ size
) <= adev
->gmc
.visible_vram_size
)
1360 /* Can't move a pinned BO to visible VRAM */
1361 if (abo
->pin_count
> 0)
1364 /* hurrah the memory is not visible ! */
1365 atomic64_inc(&adev
->num_vram_cpu_page_faults
);
1366 amdgpu_bo_placement_from_domain(abo
, AMDGPU_GEM_DOMAIN_VRAM
|
1367 AMDGPU_GEM_DOMAIN_GTT
);
1369 /* Avoid costly evictions; only set GTT as a busy placement */
1370 abo
->placement
.num_busy_placement
= 1;
1371 abo
->placement
.busy_placement
= &abo
->placements
[1];
1373 r
= ttm_bo_validate(bo
, &abo
->placement
, &ctx
);
1374 if (unlikely(r
!= 0))
1377 offset
= bo
->mem
.start
<< PAGE_SHIFT
;
1378 /* this should never happen */
1379 if (bo
->mem
.mem_type
== TTM_PL_VRAM
&&
1380 (offset
+ size
) > adev
->gmc
.visible_vram_size
)
1387 * amdgpu_bo_fence - add fence to buffer object
1389 * @bo: buffer object in question
1390 * @fence: fence to add
1391 * @shared: true if fence should be added shared
1394 void amdgpu_bo_fence(struct amdgpu_bo
*bo
, struct dma_fence
*fence
,
1397 struct dma_resv
*resv
= bo
->tbo
.base
.resv
;
1400 dma_resv_add_shared_fence(resv
, fence
);
1402 dma_resv_add_excl_fence(resv
, fence
);
1406 * amdgpu_sync_wait_resv - Wait for BO reservation fences
1408 * @bo: buffer object
1409 * @owner: fence owner
1410 * @intr: Whether the wait is interruptible
1413 * 0 on success, errno otherwise.
1415 int amdgpu_bo_sync_wait(struct amdgpu_bo
*bo
, void *owner
, bool intr
)
1417 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
1418 struct amdgpu_sync sync
;
1421 amdgpu_sync_create(&sync
);
1422 amdgpu_sync_resv(adev
, &sync
, bo
->tbo
.base
.resv
, owner
, false);
1423 r
= amdgpu_sync_wait(&sync
, intr
);
1424 amdgpu_sync_free(&sync
);
1430 * amdgpu_bo_gpu_offset - return GPU offset of bo
1431 * @bo: amdgpu object for which we query the offset
1433 * Note: object should either be pinned or reserved when calling this
1434 * function, it might be useful to add check for this for debugging.
1437 * current GPU offset of the object.
1439 u64
amdgpu_bo_gpu_offset(struct amdgpu_bo
*bo
)
1441 WARN_ON_ONCE(bo
->tbo
.mem
.mem_type
== TTM_PL_SYSTEM
);
1442 WARN_ON_ONCE(!dma_resv_is_locked(bo
->tbo
.base
.resv
) &&
1443 !bo
->pin_count
&& bo
->tbo
.type
!= ttm_bo_type_kernel
);
1444 WARN_ON_ONCE(bo
->tbo
.mem
.start
== AMDGPU_BO_INVALID_OFFSET
);
1445 WARN_ON_ONCE(bo
->tbo
.mem
.mem_type
== TTM_PL_VRAM
&&
1446 !(bo
->flags
& AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS
));
1448 return amdgpu_gmc_sign_extend(bo
->tbo
.offset
);
1452 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1453 * @adev: amdgpu device object
1454 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1457 * Which of the allowed domains is preferred for pinning the BO for scanout.
1459 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device
*adev
,
1462 if (domain
== (AMDGPU_GEM_DOMAIN_VRAM
| AMDGPU_GEM_DOMAIN_GTT
)) {
1463 domain
= AMDGPU_GEM_DOMAIN_VRAM
;
1464 if (adev
->gmc
.real_vram_size
<= AMDGPU_SG_THRESHOLD
)
1465 domain
= AMDGPU_GEM_DOMAIN_GTT
;