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
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
32 #include <linux/list.h>
33 #include <linux/slab.h>
34 #include <linux/dma-buf.h>
36 #include <drm/drm_drv.h>
37 #include <drm/amdgpu_drm.h>
38 #include <drm/drm_cache.h>
40 #include "amdgpu_trace.h"
41 #include "amdgpu_amdkfd.h"
42 #include "amdgpu_vram_mgr.h"
43 #include "amdgpu_vm.h"
44 #include "amdgpu_dma_buf.h"
49 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
50 * represents memory used by driver (VRAM, system memory, etc.). The driver
51 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
52 * to create/destroy/set buffer object which are then managed by the kernel TTM
54 * The interfaces are also used internally by kernel clients, including gfx,
55 * uvd, etc. for kernel managed allocations used by the GPU.
59 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
61 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
65 if (bo->tbo.base.import_attach)
66 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
67 drm_gem_object_release(&bo->tbo.base);
68 amdgpu_bo_unref(&bo->parent);
72 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
74 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
75 struct amdgpu_bo_user *ubo;
77 ubo = to_amdgpu_bo_user(bo);
79 amdgpu_bo_destroy(tbo);
83 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
84 * @bo: buffer object to be checked
86 * Uses destroy function associated with the object to determine if this is
90 * true if the object belongs to &amdgpu_bo, false if not.
92 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
94 if (bo->destroy == &amdgpu_bo_destroy ||
95 bo->destroy == &amdgpu_bo_user_destroy)
102 * amdgpu_bo_placement_from_domain - set buffer's placement
103 * @abo: &amdgpu_bo buffer object whose placement is to be set
104 * @domain: requested domain
106 * Sets buffer's placement according to requested domain and the buffer's
109 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
111 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
112 struct ttm_placement *placement = &abo->placement;
113 struct ttm_place *places = abo->placements;
114 u64 flags = abo->flags;
117 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
118 unsigned int visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
119 int8_t mem_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
121 if (adev->gmc.mem_partitions && mem_id >= 0) {
122 places[c].fpfn = adev->gmc.mem_partitions[mem_id].range.fpfn;
124 * memory partition range lpfn is inclusive start + size - 1
125 * TTM place lpfn is exclusive start + size
127 places[c].lpfn = adev->gmc.mem_partitions[mem_id].range.lpfn + 1;
132 places[c].mem_type = TTM_PL_VRAM;
135 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
136 places[c].lpfn = min_not_zero(places[c].lpfn, visible_pfn);
138 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
140 if (abo->tbo.type == ttm_bo_type_kernel &&
141 flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
142 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
147 if (domain & AMDGPU_GEM_DOMAIN_DOORBELL) {
150 places[c].mem_type = AMDGPU_PL_DOORBELL;
155 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
159 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
160 AMDGPU_PL_PREEMPT : TTM_PL_TT;
163 * When GTT is just an alternative to VRAM make sure that we
164 * only use it as fallback and still try to fill up VRAM first.
166 if (domain & abo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
167 !(adev->flags & AMD_IS_APU))
168 places[c].flags |= TTM_PL_FLAG_FALLBACK;
172 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
175 places[c].mem_type = TTM_PL_SYSTEM;
180 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
183 places[c].mem_type = AMDGPU_PL_GDS;
188 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
191 places[c].mem_type = AMDGPU_PL_GWS;
196 if (domain & AMDGPU_GEM_DOMAIN_OA) {
199 places[c].mem_type = AMDGPU_PL_OA;
207 places[c].mem_type = TTM_PL_SYSTEM;
212 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
214 placement->num_placement = c;
215 placement->placement = places;
219 * amdgpu_bo_create_reserved - create reserved BO for kernel use
221 * @adev: amdgpu device object
222 * @size: size for the new BO
223 * @align: alignment for the new BO
224 * @domain: where to place it
225 * @bo_ptr: used to initialize BOs in structures
226 * @gpu_addr: GPU addr of the pinned BO
227 * @cpu_addr: optional CPU address mapping
229 * Allocates and pins a BO for kernel internal use, and returns it still
232 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
235 * 0 on success, negative error code otherwise.
237 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
238 unsigned long size, int align,
239 u32 domain, struct amdgpu_bo **bo_ptr,
240 u64 *gpu_addr, void **cpu_addr)
242 struct amdgpu_bo_param bp;
247 amdgpu_bo_unref(bo_ptr);
251 memset(&bp, 0, sizeof(bp));
253 bp.byte_align = align;
255 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
256 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
257 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
258 bp.type = ttm_bo_type_kernel;
260 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
263 r = amdgpu_bo_create(adev, &bp, bo_ptr);
265 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
272 r = amdgpu_bo_reserve(*bo_ptr, false);
274 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
278 r = amdgpu_bo_pin(*bo_ptr, domain);
280 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
281 goto error_unreserve;
284 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
286 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
291 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
294 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
296 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
304 amdgpu_bo_unpin(*bo_ptr);
306 amdgpu_bo_unreserve(*bo_ptr);
310 amdgpu_bo_unref(bo_ptr);
316 * amdgpu_bo_create_kernel - create BO for kernel use
318 * @adev: amdgpu device object
319 * @size: size for the new BO
320 * @align: alignment for the new BO
321 * @domain: where to place it
322 * @bo_ptr: used to initialize BOs in structures
323 * @gpu_addr: GPU addr of the pinned BO
324 * @cpu_addr: optional CPU address mapping
326 * Allocates and pins a BO for kernel internal use.
328 * This function is exported to allow the V4L2 isp device
329 * external to drm device to create and access the kernel BO.
331 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
334 * 0 on success, negative error code otherwise.
336 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
337 unsigned long size, int align,
338 u32 domain, struct amdgpu_bo **bo_ptr,
339 u64 *gpu_addr, void **cpu_addr)
343 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
350 amdgpu_bo_unreserve(*bo_ptr);
354 EXPORT_SYMBOL(amdgpu_bo_create_kernel);
357 * amdgpu_bo_create_isp_user - create user BO for isp
359 * @adev: amdgpu device object
360 * @dma_buf: DMABUF handle for isp buffer
361 * @domain: where to place it
362 * @bo: used to initialize BOs in structures
363 * @gpu_addr: GPU addr of the pinned BO
365 * Imports isp DMABUF to allocate and pin a user BO for isp internal use. It does
366 * GART alloc to generate gpu_addr for BO to make it accessible through the
367 * GART aperture for ISP HW.
369 * This function is exported to allow the V4L2 isp device external to drm device
370 * to create and access the isp user BO.
373 * 0 on success, negative error code otherwise.
375 int amdgpu_bo_create_isp_user(struct amdgpu_device *adev,
376 struct dma_buf *dma_buf, u32 domain, struct amdgpu_bo **bo,
380 struct drm_gem_object *gem_obj;
383 gem_obj = amdgpu_gem_prime_import(&adev->ddev, dma_buf);
384 *bo = gem_to_amdgpu_bo(gem_obj);
386 dev_err(adev->dev, "failed to get valid isp user bo\n");
390 r = amdgpu_bo_reserve(*bo, false);
392 dev_err(adev->dev, "(%d) failed to reserve isp user bo\n", r);
396 r = amdgpu_bo_pin(*bo, domain);
398 dev_err(adev->dev, "(%d) isp user bo pin failed\n", r);
399 goto error_unreserve;
402 r = amdgpu_ttm_alloc_gart(&(*bo)->tbo);
404 dev_err(adev->dev, "%p bind failed\n", *bo);
408 if (!WARN_ON(!gpu_addr))
409 *gpu_addr = amdgpu_bo_gpu_offset(*bo);
411 amdgpu_bo_unreserve(*bo);
416 amdgpu_bo_unpin(*bo);
418 amdgpu_bo_unreserve(*bo);
423 EXPORT_SYMBOL(amdgpu_bo_create_isp_user);
426 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
428 * @adev: amdgpu device object
429 * @offset: offset of the BO
430 * @size: size of the BO
431 * @bo_ptr: used to initialize BOs in structures
432 * @cpu_addr: optional CPU address mapping
434 * Creates a kernel BO at a specific offset in VRAM.
437 * 0 on success, negative error code otherwise.
439 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
440 uint64_t offset, uint64_t size,
441 struct amdgpu_bo **bo_ptr, void **cpu_addr)
443 struct ttm_operation_ctx ctx = { false, false };
448 size = ALIGN(size, PAGE_SIZE);
450 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
451 AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
456 if ((*bo_ptr) == NULL)
460 * Remove the original mem node and create a new one at the request
464 amdgpu_bo_kunmap(*bo_ptr);
466 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
468 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
469 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
470 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
472 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
473 &(*bo_ptr)->tbo.resource, &ctx);
478 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
483 amdgpu_bo_unreserve(*bo_ptr);
487 amdgpu_bo_unreserve(*bo_ptr);
488 amdgpu_bo_unref(bo_ptr);
493 * amdgpu_bo_free_kernel - free BO for kernel use
495 * @bo: amdgpu BO to free
496 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
497 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
499 * unmaps and unpin a BO for kernel internal use.
501 * This function is exported to allow the V4L2 isp device
502 * external to drm device to free the kernel BO.
504 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
510 WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend);
512 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
514 amdgpu_bo_kunmap(*bo);
516 amdgpu_bo_unpin(*bo);
517 amdgpu_bo_unreserve(*bo);
527 EXPORT_SYMBOL(amdgpu_bo_free_kernel);
530 * amdgpu_bo_free_isp_user - free BO for isp use
532 * @bo: amdgpu isp user BO to free
534 * unpin and unref BO for isp internal use.
536 * This function is exported to allow the V4L2 isp device
537 * external to drm device to free the isp user BO.
539 void amdgpu_bo_free_isp_user(struct amdgpu_bo *bo)
544 if (amdgpu_bo_reserve(bo, true) == 0) {
546 amdgpu_bo_unreserve(bo);
548 amdgpu_bo_unref(&bo);
550 EXPORT_SYMBOL(amdgpu_bo_free_isp_user);
552 /* Validate bo size is bit bigger than the request domain */
553 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
554 unsigned long size, u32 domain)
556 struct ttm_resource_manager *man = NULL;
559 * If GTT is part of requested domains the check must succeed to
560 * allow fall back to GTT.
562 if (domain & AMDGPU_GEM_DOMAIN_GTT)
563 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
564 else if (domain & AMDGPU_GEM_DOMAIN_VRAM)
565 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
570 if (domain & AMDGPU_GEM_DOMAIN_GTT)
571 WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
575 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU, _DOMAIN_DOORBELL */
576 if (size < man->size)
579 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, man->size);
583 bool amdgpu_bo_support_uswc(u64 bo_flags)
587 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
588 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
591 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
592 /* Don't try to enable write-combining when it can't work, or things
594 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
597 #ifndef CONFIG_COMPILE_TEST
598 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
599 thanks to write-combining
602 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
603 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
604 "better performance thanks to write-combining\n");
607 /* For architectures that don't support WC memory,
608 * mask out the WC flag from the BO
610 if (!drm_arch_can_wc_memory())
618 * amdgpu_bo_create - create an &amdgpu_bo buffer object
619 * @adev: amdgpu device object
620 * @bp: parameters to be used for the buffer object
621 * @bo_ptr: pointer to the buffer object pointer
623 * Creates an &amdgpu_bo buffer object.
626 * 0 for success or a negative error code on failure.
628 int amdgpu_bo_create(struct amdgpu_device *adev,
629 struct amdgpu_bo_param *bp,
630 struct amdgpu_bo **bo_ptr)
632 struct ttm_operation_ctx ctx = {
633 .interruptible = (bp->type != ttm_bo_type_kernel),
634 .no_wait_gpu = bp->no_wait_gpu,
635 /* We opt to avoid OOM on system pages allocations */
636 .gfp_retry_mayfail = true,
637 .allow_res_evict = bp->type != ttm_bo_type_kernel,
640 struct amdgpu_bo *bo;
641 unsigned long page_align, size = bp->size;
644 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
645 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
646 /* GWS and OA don't need any alignment. */
647 page_align = bp->byte_align;
650 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
651 /* Both size and alignment must be a multiple of 4. */
652 page_align = ALIGN(bp->byte_align, 4);
653 size = ALIGN(size, 4) << PAGE_SHIFT;
655 /* Memory should be aligned at least to a page size. */
656 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
657 size = ALIGN(size, PAGE_SIZE);
660 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
663 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
666 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
669 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
670 bo->tbo.base.funcs = &amdgpu_gem_object_funcs;
672 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
674 bo->allowed_domains = bo->preferred_domains;
675 if (bp->type != ttm_bo_type_kernel &&
676 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
677 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
678 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
680 bo->flags = bp->flags;
682 if (adev->gmc.mem_partitions)
683 /* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */
684 bo->xcp_id = bp->xcp_id_plus1 - 1;
686 /* For GPUs without spatial partitioning */
689 if (!amdgpu_bo_support_uswc(bo->flags))
690 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
692 bo->tbo.bdev = &adev->mman.bdev;
693 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
694 AMDGPU_GEM_DOMAIN_GDS))
695 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
697 amdgpu_bo_placement_from_domain(bo, bp->domain);
698 if (bp->type == ttm_bo_type_kernel)
699 bo->tbo.priority = 2;
700 else if (!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE))
701 bo->tbo.priority = 1;
704 bp->destroy = &amdgpu_bo_destroy;
706 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
707 &bo->placement, page_align, &ctx, NULL,
708 bp->resv, bp->destroy);
709 if (unlikely(r != 0))
712 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
713 amdgpu_res_cpu_visible(adev, bo->tbo.resource))
714 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
717 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
719 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
720 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
721 struct dma_fence *fence;
723 r = amdgpu_ttm_clear_buffer(bo, bo->tbo.base.resv, &fence);
727 dma_resv_add_fence(bo->tbo.base.resv, fence,
728 DMA_RESV_USAGE_KERNEL);
729 dma_fence_put(fence);
732 amdgpu_bo_unreserve(bo);
735 trace_amdgpu_bo_create(bo);
737 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
738 if (bp->type == ttm_bo_type_device)
739 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
745 dma_resv_unlock(bo->tbo.base.resv);
746 amdgpu_bo_unref(&bo);
751 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
752 * @adev: amdgpu device object
753 * @bp: parameters to be used for the buffer object
754 * @ubo_ptr: pointer to the buffer object pointer
756 * Create a BO to be used by user application;
759 * 0 for success or a negative error code on failure.
762 int amdgpu_bo_create_user(struct amdgpu_device *adev,
763 struct amdgpu_bo_param *bp,
764 struct amdgpu_bo_user **ubo_ptr)
766 struct amdgpu_bo *bo_ptr;
769 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
770 bp->destroy = &amdgpu_bo_user_destroy;
771 r = amdgpu_bo_create(adev, bp, &bo_ptr);
775 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
780 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
781 * @adev: amdgpu device object
782 * @bp: parameters to be used for the buffer object
783 * @vmbo_ptr: pointer to the buffer object pointer
785 * Create a BO to be for GPUVM.
788 * 0 for success or a negative error code on failure.
791 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
792 struct amdgpu_bo_param *bp,
793 struct amdgpu_bo_vm **vmbo_ptr)
795 struct amdgpu_bo *bo_ptr;
798 /* bo_ptr_size will be determined by the caller and it depends on
799 * num of amdgpu_vm_pt entries.
801 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
802 r = amdgpu_bo_create(adev, bp, &bo_ptr);
806 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
811 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
812 * @bo: &amdgpu_bo buffer object to be mapped
813 * @ptr: kernel virtual address to be returned
815 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
816 * amdgpu_bo_kptr() to get the kernel virtual address.
819 * 0 for success or a negative error code on failure.
821 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
826 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
829 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
830 false, MAX_SCHEDULE_TIMEOUT);
834 kptr = amdgpu_bo_kptr(bo);
841 r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
846 *ptr = amdgpu_bo_kptr(bo);
852 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
853 * @bo: &amdgpu_bo buffer object
855 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
858 * the virtual address of a buffer object area.
860 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
864 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
868 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
869 * @bo: &amdgpu_bo buffer object to be unmapped
871 * Unmaps a kernel map set up by amdgpu_bo_kmap().
873 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
876 ttm_bo_kunmap(&bo->kmap);
880 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
881 * @bo: &amdgpu_bo buffer object
883 * References the contained &ttm_buffer_object.
886 * a refcounted pointer to the &amdgpu_bo buffer object.
888 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
893 drm_gem_object_get(&bo->tbo.base);
898 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
899 * @bo: &amdgpu_bo buffer object
901 * Unreferences the contained &ttm_buffer_object and clear the pointer
903 void amdgpu_bo_unref(struct amdgpu_bo **bo)
908 drm_gem_object_put(&(*bo)->tbo.base);
913 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
914 * @bo: &amdgpu_bo buffer object to be pinned
915 * @domain: domain to be pinned to
917 * Pins the buffer object according to requested domain. If the memory is
918 * unbound gart memory, binds the pages into gart table. Adjusts pin_count and
919 * pin_size accordingly.
921 * Pinning means to lock pages in memory along with keeping them at a fixed
922 * offset. It is required when a buffer can not be moved, for example, when
923 * a display buffer is being scanned out.
926 * 0 for success or a negative error code on failure.
928 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
930 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
931 struct ttm_operation_ctx ctx = { false, false };
934 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
937 /* Check domain to be pinned to against preferred domains */
938 if (bo->preferred_domains & domain)
939 domain = bo->preferred_domains & domain;
941 /* A shared bo cannot be migrated to VRAM */
942 if (bo->tbo.base.import_attach) {
943 if (domain & AMDGPU_GEM_DOMAIN_GTT)
944 domain = AMDGPU_GEM_DOMAIN_GTT;
949 if (bo->tbo.pin_count) {
950 uint32_t mem_type = bo->tbo.resource->mem_type;
951 uint32_t mem_flags = bo->tbo.resource->placement;
953 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
956 if ((mem_type == TTM_PL_VRAM) &&
957 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
958 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
961 ttm_bo_pin(&bo->tbo);
965 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
966 * See function amdgpu_display_supported_domains()
968 domain = amdgpu_bo_get_preferred_domain(adev, domain);
970 if (bo->tbo.base.import_attach)
971 dma_buf_pin(bo->tbo.base.import_attach);
973 /* force to pin into visible video ram */
974 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
975 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
976 amdgpu_bo_placement_from_domain(bo, domain);
977 for (i = 0; i < bo->placement.num_placement; i++) {
978 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS &&
979 bo->placements[i].mem_type == TTM_PL_VRAM)
980 bo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
983 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
985 dev_err(adev->dev, "%p pin failed\n", bo);
989 ttm_bo_pin(&bo->tbo);
991 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
992 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
993 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
994 &adev->visible_pin_size);
995 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
996 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
1004 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
1005 * @bo: &amdgpu_bo buffer object to be unpinned
1007 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
1008 * Changes placement and pin size accordingly.
1011 * 0 for success or a negative error code on failure.
1013 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
1015 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1017 ttm_bo_unpin(&bo->tbo);
1018 if (bo->tbo.pin_count)
1021 if (bo->tbo.base.import_attach)
1022 dma_buf_unpin(bo->tbo.base.import_attach);
1024 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1025 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1026 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1027 &adev->visible_pin_size);
1028 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1029 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1034 static const char * const amdgpu_vram_names[] = {
1051 * amdgpu_bo_init - initialize memory manager
1052 * @adev: amdgpu device object
1054 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1057 * 0 for success or a negative error code on failure.
1059 int amdgpu_bo_init(struct amdgpu_device *adev)
1061 /* On A+A platform, VRAM can be mapped as WB */
1062 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1063 /* reserve PAT memory space to WC for VRAM */
1064 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1065 adev->gmc.aper_size);
1068 DRM_ERROR("Unable to set WC memtype for the aperture base\n");
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);
1077 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1078 adev->gmc.mc_vram_size >> 20,
1079 (unsigned long long)adev->gmc.aper_size >> 20);
1080 DRM_INFO("RAM width %dbits %s\n",
1081 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1082 return amdgpu_ttm_init(adev);
1086 * amdgpu_bo_fini - tear down memory manager
1087 * @adev: amdgpu device object
1089 * Reverses amdgpu_bo_init() to tear down memory manager.
1091 void amdgpu_bo_fini(struct amdgpu_device *adev)
1095 amdgpu_ttm_fini(adev);
1097 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
1098 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1099 arch_phys_wc_del(adev->gmc.vram_mtrr);
1100 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1107 * amdgpu_bo_set_tiling_flags - set tiling flags
1108 * @bo: &amdgpu_bo buffer object
1109 * @tiling_flags: new flags
1111 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1112 * kernel driver to set the tiling flags on a buffer.
1115 * 0 for success or a negative error code on failure.
1117 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1119 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1120 struct amdgpu_bo_user *ubo;
1122 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1123 if (adev->family <= AMDGPU_FAMILY_CZ &&
1124 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1127 ubo = to_amdgpu_bo_user(bo);
1128 ubo->tiling_flags = tiling_flags;
1133 * amdgpu_bo_get_tiling_flags - get tiling flags
1134 * @bo: &amdgpu_bo buffer object
1135 * @tiling_flags: returned flags
1137 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1138 * set the tiling flags on a buffer.
1140 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1142 struct amdgpu_bo_user *ubo;
1144 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1145 dma_resv_assert_held(bo->tbo.base.resv);
1146 ubo = to_amdgpu_bo_user(bo);
1149 *tiling_flags = ubo->tiling_flags;
1153 * amdgpu_bo_set_metadata - set metadata
1154 * @bo: &amdgpu_bo buffer object
1155 * @metadata: new metadata
1156 * @metadata_size: size of the new metadata
1157 * @flags: flags of the new metadata
1159 * Sets buffer object's metadata, its size and flags.
1160 * Used via GEM ioctl.
1163 * 0 for success or a negative error code on failure.
1165 int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata,
1166 u32 metadata_size, uint64_t flags)
1168 struct amdgpu_bo_user *ubo;
1171 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1172 ubo = to_amdgpu_bo_user(bo);
1173 if (!metadata_size) {
1174 if (ubo->metadata_size) {
1175 kfree(ubo->metadata);
1176 ubo->metadata = NULL;
1177 ubo->metadata_size = 0;
1182 if (metadata == NULL)
1185 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1189 kfree(ubo->metadata);
1190 ubo->metadata_flags = flags;
1191 ubo->metadata = buffer;
1192 ubo->metadata_size = metadata_size;
1198 * amdgpu_bo_get_metadata - get metadata
1199 * @bo: &amdgpu_bo buffer object
1200 * @buffer: returned metadata
1201 * @buffer_size: size of the buffer
1202 * @metadata_size: size of the returned metadata
1203 * @flags: flags of the returned metadata
1205 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1206 * less than metadata_size.
1207 * Used via GEM ioctl.
1210 * 0 for success or a negative error code on failure.
1212 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1213 size_t buffer_size, uint32_t *metadata_size,
1216 struct amdgpu_bo_user *ubo;
1218 if (!buffer && !metadata_size)
1221 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1222 ubo = to_amdgpu_bo_user(bo);
1224 *metadata_size = ubo->metadata_size;
1227 if (buffer_size < ubo->metadata_size)
1230 if (ubo->metadata_size)
1231 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1235 *flags = ubo->metadata_flags;
1241 * amdgpu_bo_move_notify - notification about a memory move
1242 * @bo: pointer to a buffer object
1243 * @evict: if this move is evicting the buffer from the graphics address space
1244 * @new_mem: new resource for backing the BO
1246 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1248 * TTM driver callback which is called when ttm moves a buffer.
1250 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1252 struct ttm_resource *new_mem)
1254 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1255 struct ttm_resource *old_mem = bo->resource;
1256 struct amdgpu_bo *abo;
1258 if (!amdgpu_bo_is_amdgpu_bo(bo))
1261 abo = ttm_to_amdgpu_bo(bo);
1262 amdgpu_vm_bo_invalidate(adev, abo, evict);
1264 amdgpu_bo_kunmap(abo);
1266 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1267 old_mem && old_mem->mem_type != TTM_PL_SYSTEM)
1268 dma_buf_move_notify(abo->tbo.base.dma_buf);
1270 /* move_notify is called before move happens */
1271 trace_amdgpu_bo_move(abo, new_mem ? new_mem->mem_type : -1,
1272 old_mem ? old_mem->mem_type : -1);
1275 void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
1276 struct amdgpu_mem_stats *stats,
1279 const unsigned int domain_to_pl[] = {
1280 [ilog2(AMDGPU_GEM_DOMAIN_CPU)] = TTM_PL_SYSTEM,
1281 [ilog2(AMDGPU_GEM_DOMAIN_GTT)] = TTM_PL_TT,
1282 [ilog2(AMDGPU_GEM_DOMAIN_VRAM)] = TTM_PL_VRAM,
1283 [ilog2(AMDGPU_GEM_DOMAIN_GDS)] = AMDGPU_PL_GDS,
1284 [ilog2(AMDGPU_GEM_DOMAIN_GWS)] = AMDGPU_PL_GWS,
1285 [ilog2(AMDGPU_GEM_DOMAIN_OA)] = AMDGPU_PL_OA,
1286 [ilog2(AMDGPU_GEM_DOMAIN_DOORBELL)] = AMDGPU_PL_DOORBELL,
1288 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1289 struct ttm_resource *res = bo->tbo.resource;
1290 struct drm_gem_object *obj = &bo->tbo.base;
1291 uint64_t size = amdgpu_bo_size(bo);
1296 * If no backing store use one of the preferred domain for basic
1297 * stats. We take the MSB since that should give a reasonable
1300 BUILD_BUG_ON(TTM_PL_VRAM < TTM_PL_TT ||
1301 TTM_PL_VRAM < TTM_PL_SYSTEM);
1302 type = fls(bo->preferred_domains & AMDGPU_GEM_DOMAIN_MASK);
1306 if (drm_WARN_ON_ONCE(&adev->ddev,
1307 type >= ARRAY_SIZE(domain_to_pl)))
1309 type = domain_to_pl[type];
1311 type = res->mem_type;
1314 if (drm_WARN_ON_ONCE(&adev->ddev, type >= sz))
1317 /* DRM stats common fields: */
1319 if (drm_gem_object_is_shared_for_memory_stats(obj))
1320 stats[type].drm.shared += size;
1322 stats[type].drm.private += size;
1325 stats[type].drm.resident += size;
1327 if (!dma_resv_test_signaled(obj->resv, DMA_RESV_USAGE_BOOKKEEP))
1328 stats[type].drm.active += size;
1329 else if (bo->flags & AMDGPU_GEM_CREATE_DISCARDABLE)
1330 stats[type].drm.purgeable += size;
1333 /* amdgpu specific stats: */
1335 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) {
1336 stats[TTM_PL_VRAM].requested += size;
1337 if (type != TTM_PL_VRAM)
1338 stats[TTM_PL_VRAM].evicted += size;
1339 } else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) {
1340 stats[TTM_PL_TT].requested += size;
1345 * amdgpu_bo_release_notify - notification about a BO being released
1346 * @bo: pointer to a buffer object
1348 * Wipes VRAM buffers whose contents should not be leaked before the
1349 * memory is released.
1351 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1353 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1354 struct dma_fence *fence = NULL;
1355 struct amdgpu_bo *abo;
1358 if (!amdgpu_bo_is_amdgpu_bo(bo))
1361 abo = ttm_to_amdgpu_bo(bo);
1363 WARN_ON(abo->vm_bo);
1366 amdgpu_amdkfd_release_notify(abo);
1368 /* We only remove the fence if the resv has individualized. */
1369 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1370 && bo->base.resv != &bo->base._resv);
1371 if (bo->base.resv == &bo->base._resv)
1372 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1374 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1375 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1376 adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev)))
1379 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1382 r = amdgpu_fill_buffer(abo, 0, bo->base.resv, &fence, true);
1384 amdgpu_vram_mgr_set_cleared(bo->resource);
1385 amdgpu_bo_fence(abo, fence, false);
1386 dma_fence_put(fence);
1389 dma_resv_unlock(bo->base.resv);
1393 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1394 * @bo: pointer to a buffer object
1396 * Notifies the driver we are taking a fault on this BO and have reserved it,
1397 * also performs bookkeeping.
1398 * TTM driver callback for dealing with vm faults.
1401 * 0 for success or a negative error code on failure.
1403 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1405 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1406 struct ttm_operation_ctx ctx = { false, false };
1407 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1410 /* Remember that this BO was accessed by the CPU */
1411 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1413 if (amdgpu_res_cpu_visible(adev, bo->resource))
1416 /* Can't move a pinned BO to visible VRAM */
1417 if (abo->tbo.pin_count > 0)
1418 return VM_FAULT_SIGBUS;
1420 /* hurrah the memory is not visible ! */
1421 atomic64_inc(&adev->num_vram_cpu_page_faults);
1422 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1423 AMDGPU_GEM_DOMAIN_GTT);
1425 /* Avoid costly evictions; only set GTT as a busy placement */
1426 abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
1428 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1429 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1430 return VM_FAULT_NOPAGE;
1431 else if (unlikely(r))
1432 return VM_FAULT_SIGBUS;
1434 /* this should never happen */
1435 if (bo->resource->mem_type == TTM_PL_VRAM &&
1436 !amdgpu_res_cpu_visible(adev, bo->resource))
1437 return VM_FAULT_SIGBUS;
1439 ttm_bo_move_to_lru_tail_unlocked(bo);
1444 * amdgpu_bo_fence - add fence to buffer object
1446 * @bo: buffer object in question
1447 * @fence: fence to add
1448 * @shared: true if fence should be added shared
1451 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1454 struct dma_resv *resv = bo->tbo.base.resv;
1457 r = dma_resv_reserve_fences(resv, 1);
1459 /* As last resort on OOM we block for the fence */
1460 dma_fence_wait(fence, false);
1464 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1465 DMA_RESV_USAGE_WRITE);
1469 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1471 * @adev: amdgpu device pointer
1472 * @resv: reservation object to sync to
1473 * @sync_mode: synchronization mode
1474 * @owner: fence owner
1475 * @intr: Whether the wait is interruptible
1477 * Extract the fences from the reservation object and waits for them to finish.
1480 * 0 on success, errno otherwise.
1482 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1483 enum amdgpu_sync_mode sync_mode, void *owner,
1486 struct amdgpu_sync sync;
1489 amdgpu_sync_create(&sync);
1490 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1491 r = amdgpu_sync_wait(&sync, intr);
1492 amdgpu_sync_free(&sync);
1497 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1498 * @bo: buffer object to wait for
1499 * @owner: fence owner
1500 * @intr: Whether the wait is interruptible
1502 * Wrapper to wait for fences in a BO.
1504 * 0 on success, errno otherwise.
1506 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1508 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1510 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1511 AMDGPU_SYNC_NE_OWNER, owner, intr);
1515 * amdgpu_bo_gpu_offset - return GPU offset of bo
1516 * @bo: amdgpu object for which we query the offset
1518 * Note: object should either be pinned or reserved when calling this
1519 * function, it might be useful to add check for this for debugging.
1522 * current GPU offset of the object.
1524 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1526 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1527 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1528 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1529 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1530 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1531 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1533 return amdgpu_bo_gpu_offset_no_check(bo);
1537 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1538 * @bo: amdgpu object for which we query the offset
1541 * current GPU offset of the object without raising warnings.
1543 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1545 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1546 uint64_t offset = AMDGPU_BO_INVALID_OFFSET;
1548 if (bo->tbo.resource->mem_type == TTM_PL_TT)
1549 offset = amdgpu_gmc_agp_addr(&bo->tbo);
1551 if (offset == AMDGPU_BO_INVALID_OFFSET)
1552 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1553 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1555 return amdgpu_gmc_sign_extend(offset);
1559 * amdgpu_bo_get_preferred_domain - get preferred domain
1560 * @adev: amdgpu device object
1561 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1564 * Which of the allowed domains is preferred for allocating the BO.
1566 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1569 if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1570 ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1571 domain = AMDGPU_GEM_DOMAIN_VRAM;
1572 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1573 domain = AMDGPU_GEM_DOMAIN_GTT;
1578 #if defined(CONFIG_DEBUG_FS)
1579 #define amdgpu_bo_print_flag(m, bo, flag) \
1581 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1582 seq_printf((m), " " #flag); \
1587 * amdgpu_bo_print_info - print BO info in debugfs file
1589 * @id: Index or Id of the BO
1590 * @bo: Requested BO for printing info
1593 * Print BO information in debugfs file
1596 * Size of the BO in bytes.
1598 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1600 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1601 struct dma_buf_attachment *attachment;
1602 struct dma_buf *dma_buf;
1603 const char *placement;
1604 unsigned int pin_count;
1607 if (dma_resv_trylock(bo->tbo.base.resv)) {
1608 if (!bo->tbo.resource) {
1611 switch (bo->tbo.resource->mem_type) {
1613 if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1614 placement = "VRAM VISIBLE";
1630 case AMDGPU_PL_PREEMPT:
1631 placement = "PREEMPTIBLE";
1633 case AMDGPU_PL_DOORBELL:
1634 placement = "DOORBELL";
1642 dma_resv_unlock(bo->tbo.base.resv);
1644 placement = "UNKNOWN";
1647 size = amdgpu_bo_size(bo);
1648 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1649 id, size, placement);
1651 pin_count = READ_ONCE(bo->tbo.pin_count);
1653 seq_printf(m, " pin count %d", pin_count);
1655 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1656 attachment = READ_ONCE(bo->tbo.base.import_attach);
1659 seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino);
1661 seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino);
1663 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1664 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1665 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1666 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1667 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1668 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1669 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);