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"
47 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
48 * represents memory used by driver (VRAM, system memory, etc.). The driver
49 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
50 * to create/destroy/set buffer object which are then managed by the kernel TTM
52 * The interfaces are also used internally by kernel clients, including gfx,
53 * uvd, etc. for kernel managed allocations used by the GPU.
57 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
59 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
63 if (bo->tbo.base.import_attach)
64 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
65 drm_gem_object_release(&bo->tbo.base);
66 amdgpu_bo_unref(&bo->parent);
70 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
72 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
73 struct amdgpu_bo_user *ubo;
75 ubo = to_amdgpu_bo_user(bo);
77 amdgpu_bo_destroy(tbo);
80 static void amdgpu_bo_vm_destroy(struct ttm_buffer_object *tbo)
82 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
83 struct amdgpu_bo *shadow_bo = ttm_to_amdgpu_bo(tbo), *bo;
84 struct amdgpu_bo_vm *vmbo;
86 bo = shadow_bo->parent;
87 vmbo = to_amdgpu_bo_vm(bo);
88 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
89 if (!list_empty(&vmbo->shadow_list)) {
90 mutex_lock(&adev->shadow_list_lock);
91 list_del_init(&vmbo->shadow_list);
92 mutex_unlock(&adev->shadow_list_lock);
95 amdgpu_bo_destroy(tbo);
99 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
100 * @bo: buffer object to be checked
102 * Uses destroy function associated with the object to determine if this is
106 * true if the object belongs to &amdgpu_bo, false if not.
108 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
110 if (bo->destroy == &amdgpu_bo_destroy ||
111 bo->destroy == &amdgpu_bo_user_destroy ||
112 bo->destroy == &amdgpu_bo_vm_destroy)
119 * amdgpu_bo_placement_from_domain - set buffer's placement
120 * @abo: &amdgpu_bo buffer object whose placement is to be set
121 * @domain: requested domain
123 * Sets buffer's placement according to requested domain and the buffer's
126 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
128 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
129 struct ttm_placement *placement = &abo->placement;
130 struct ttm_place *places = abo->placements;
131 u64 flags = abo->flags;
134 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
135 unsigned int visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
136 int8_t mem_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
138 if (adev->gmc.mem_partitions && mem_id >= 0) {
139 places[c].fpfn = adev->gmc.mem_partitions[mem_id].range.fpfn;
141 * memory partition range lpfn is inclusive start + size - 1
142 * TTM place lpfn is exclusive start + size
144 places[c].lpfn = adev->gmc.mem_partitions[mem_id].range.lpfn + 1;
149 places[c].mem_type = TTM_PL_VRAM;
152 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
153 places[c].lpfn = min_not_zero(places[c].lpfn, visible_pfn);
155 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
157 if (abo->tbo.type == ttm_bo_type_kernel &&
158 flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
159 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
164 if (domain & AMDGPU_GEM_DOMAIN_DOORBELL) {
167 places[c].mem_type = AMDGPU_PL_DOORBELL;
172 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
176 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
177 AMDGPU_PL_PREEMPT : TTM_PL_TT;
180 * When GTT is just an alternative to VRAM make sure that we
181 * only use it as fallback and still try to fill up VRAM first.
183 if (domain & abo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM)
184 places[c].flags |= TTM_PL_FLAG_FALLBACK;
188 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
191 places[c].mem_type = TTM_PL_SYSTEM;
196 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
199 places[c].mem_type = AMDGPU_PL_GDS;
204 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
207 places[c].mem_type = AMDGPU_PL_GWS;
212 if (domain & AMDGPU_GEM_DOMAIN_OA) {
215 places[c].mem_type = AMDGPU_PL_OA;
223 places[c].mem_type = TTM_PL_SYSTEM;
228 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
230 placement->num_placement = c;
231 placement->placement = places;
235 * amdgpu_bo_create_reserved - create reserved BO for kernel use
237 * @adev: amdgpu device object
238 * @size: size for the new BO
239 * @align: alignment for the new BO
240 * @domain: where to place it
241 * @bo_ptr: used to initialize BOs in structures
242 * @gpu_addr: GPU addr of the pinned BO
243 * @cpu_addr: optional CPU address mapping
245 * Allocates and pins a BO for kernel internal use, and returns it still
248 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
251 * 0 on success, negative error code otherwise.
253 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
254 unsigned long size, int align,
255 u32 domain, struct amdgpu_bo **bo_ptr,
256 u64 *gpu_addr, void **cpu_addr)
258 struct amdgpu_bo_param bp;
263 amdgpu_bo_unref(bo_ptr);
267 memset(&bp, 0, sizeof(bp));
269 bp.byte_align = align;
271 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
272 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
273 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
274 bp.type = ttm_bo_type_kernel;
276 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
279 r = amdgpu_bo_create(adev, &bp, bo_ptr);
281 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
288 r = amdgpu_bo_reserve(*bo_ptr, false);
290 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
294 r = amdgpu_bo_pin(*bo_ptr, domain);
296 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
297 goto error_unreserve;
300 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
302 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
307 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
310 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
312 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
320 amdgpu_bo_unpin(*bo_ptr);
322 amdgpu_bo_unreserve(*bo_ptr);
326 amdgpu_bo_unref(bo_ptr);
332 * amdgpu_bo_create_kernel - create BO for kernel use
334 * @adev: amdgpu device object
335 * @size: size for the new BO
336 * @align: alignment for the new BO
337 * @domain: where to place it
338 * @bo_ptr: used to initialize BOs in structures
339 * @gpu_addr: GPU addr of the pinned BO
340 * @cpu_addr: optional CPU address mapping
342 * Allocates and pins a BO for kernel internal use.
344 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
347 * 0 on success, negative error code otherwise.
349 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
350 unsigned long size, int align,
351 u32 domain, struct amdgpu_bo **bo_ptr,
352 u64 *gpu_addr, void **cpu_addr)
356 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
363 amdgpu_bo_unreserve(*bo_ptr);
369 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
371 * @adev: amdgpu device object
372 * @offset: offset of the BO
373 * @size: size of the BO
374 * @bo_ptr: used to initialize BOs in structures
375 * @cpu_addr: optional CPU address mapping
377 * Creates a kernel BO at a specific offset in VRAM.
380 * 0 on success, negative error code otherwise.
382 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
383 uint64_t offset, uint64_t size,
384 struct amdgpu_bo **bo_ptr, void **cpu_addr)
386 struct ttm_operation_ctx ctx = { false, false };
391 size = ALIGN(size, PAGE_SIZE);
393 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
394 AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
399 if ((*bo_ptr) == NULL)
403 * Remove the original mem node and create a new one at the request
407 amdgpu_bo_kunmap(*bo_ptr);
409 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
411 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
412 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
413 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
415 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
416 &(*bo_ptr)->tbo.resource, &ctx);
421 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
426 amdgpu_bo_unreserve(*bo_ptr);
430 amdgpu_bo_unreserve(*bo_ptr);
431 amdgpu_bo_unref(bo_ptr);
436 * amdgpu_bo_free_kernel - free BO for kernel use
438 * @bo: amdgpu BO to free
439 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
440 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
442 * unmaps and unpin a BO for kernel internal use.
444 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
450 WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend);
452 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
454 amdgpu_bo_kunmap(*bo);
456 amdgpu_bo_unpin(*bo);
457 amdgpu_bo_unreserve(*bo);
468 /* Validate bo size is bit bigger than the request domain */
469 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
470 unsigned long size, u32 domain)
472 struct ttm_resource_manager *man = NULL;
475 * If GTT is part of requested domains the check must succeed to
476 * allow fall back to GTT.
478 if (domain & AMDGPU_GEM_DOMAIN_GTT)
479 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
480 else if (domain & AMDGPU_GEM_DOMAIN_VRAM)
481 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
486 if (domain & AMDGPU_GEM_DOMAIN_GTT)
487 WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
491 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU, _DOMAIN_DOORBELL */
492 if (size < man->size)
495 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, man->size);
499 bool amdgpu_bo_support_uswc(u64 bo_flags)
503 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
504 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
507 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
508 /* Don't try to enable write-combining when it can't work, or things
510 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
513 #ifndef CONFIG_COMPILE_TEST
514 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
515 thanks to write-combining
518 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
519 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
520 "better performance thanks to write-combining\n");
523 /* For architectures that don't support WC memory,
524 * mask out the WC flag from the BO
526 if (!drm_arch_can_wc_memory())
534 * amdgpu_bo_create - create an &amdgpu_bo buffer object
535 * @adev: amdgpu device object
536 * @bp: parameters to be used for the buffer object
537 * @bo_ptr: pointer to the buffer object pointer
539 * Creates an &amdgpu_bo buffer object.
542 * 0 for success or a negative error code on failure.
544 int amdgpu_bo_create(struct amdgpu_device *adev,
545 struct amdgpu_bo_param *bp,
546 struct amdgpu_bo **bo_ptr)
548 struct ttm_operation_ctx ctx = {
549 .interruptible = (bp->type != ttm_bo_type_kernel),
550 .no_wait_gpu = bp->no_wait_gpu,
551 /* We opt to avoid OOM on system pages allocations */
552 .gfp_retry_mayfail = true,
553 .allow_res_evict = bp->type != ttm_bo_type_kernel,
556 struct amdgpu_bo *bo;
557 unsigned long page_align, size = bp->size;
560 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
561 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
562 /* GWS and OA don't need any alignment. */
563 page_align = bp->byte_align;
566 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
567 /* Both size and alignment must be a multiple of 4. */
568 page_align = ALIGN(bp->byte_align, 4);
569 size = ALIGN(size, 4) << PAGE_SHIFT;
571 /* Memory should be aligned at least to a page size. */
572 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
573 size = ALIGN(size, PAGE_SIZE);
576 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
579 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
582 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
585 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
587 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
589 bo->allowed_domains = bo->preferred_domains;
590 if (bp->type != ttm_bo_type_kernel &&
591 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
592 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
593 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
595 bo->flags = bp->flags;
597 if (adev->gmc.mem_partitions)
598 /* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */
599 bo->xcp_id = bp->xcp_id_plus1 - 1;
601 /* For GPUs without spatial partitioning */
604 if (!amdgpu_bo_support_uswc(bo->flags))
605 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
607 bo->flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
609 bo->tbo.bdev = &adev->mman.bdev;
610 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
611 AMDGPU_GEM_DOMAIN_GDS))
612 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
614 amdgpu_bo_placement_from_domain(bo, bp->domain);
615 if (bp->type == ttm_bo_type_kernel)
616 bo->tbo.priority = 2;
617 else if (!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE))
618 bo->tbo.priority = 1;
621 bp->destroy = &amdgpu_bo_destroy;
623 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
624 &bo->placement, page_align, &ctx, NULL,
625 bp->resv, bp->destroy);
626 if (unlikely(r != 0))
629 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
630 amdgpu_res_cpu_visible(adev, bo->tbo.resource))
631 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
634 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
636 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
637 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
638 struct dma_fence *fence;
640 r = amdgpu_ttm_clear_buffer(bo, bo->tbo.base.resv, &fence);
644 dma_resv_add_fence(bo->tbo.base.resv, fence,
645 DMA_RESV_USAGE_KERNEL);
646 dma_fence_put(fence);
649 amdgpu_bo_unreserve(bo);
652 trace_amdgpu_bo_create(bo);
654 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
655 if (bp->type == ttm_bo_type_device)
656 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
662 dma_resv_unlock(bo->tbo.base.resv);
663 amdgpu_bo_unref(&bo);
668 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
669 * @adev: amdgpu device object
670 * @bp: parameters to be used for the buffer object
671 * @ubo_ptr: pointer to the buffer object pointer
673 * Create a BO to be used by user application;
676 * 0 for success or a negative error code on failure.
679 int amdgpu_bo_create_user(struct amdgpu_device *adev,
680 struct amdgpu_bo_param *bp,
681 struct amdgpu_bo_user **ubo_ptr)
683 struct amdgpu_bo *bo_ptr;
686 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
687 bp->destroy = &amdgpu_bo_user_destroy;
688 r = amdgpu_bo_create(adev, bp, &bo_ptr);
692 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
697 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
698 * @adev: amdgpu device object
699 * @bp: parameters to be used for the buffer object
700 * @vmbo_ptr: pointer to the buffer object pointer
702 * Create a BO to be for GPUVM.
705 * 0 for success or a negative error code on failure.
708 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
709 struct amdgpu_bo_param *bp,
710 struct amdgpu_bo_vm **vmbo_ptr)
712 struct amdgpu_bo *bo_ptr;
715 /* bo_ptr_size will be determined by the caller and it depends on
716 * num of amdgpu_vm_pt entries.
718 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
719 r = amdgpu_bo_create(adev, bp, &bo_ptr);
723 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
728 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list
730 * @vmbo: BO that will be inserted into the shadow list
732 * Insert a BO to the shadow list.
734 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo)
736 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev);
738 mutex_lock(&adev->shadow_list_lock);
739 list_add_tail(&vmbo->shadow_list, &adev->shadow_list);
740 vmbo->shadow->parent = amdgpu_bo_ref(&vmbo->bo);
741 vmbo->shadow->tbo.destroy = &amdgpu_bo_vm_destroy;
742 mutex_unlock(&adev->shadow_list_lock);
746 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
748 * @shadow: &amdgpu_bo shadow to be restored
749 * @fence: dma_fence associated with the operation
751 * Copies a buffer object's shadow content back to the object.
752 * This is used for recovering a buffer from its shadow in case of a gpu
753 * reset where vram context may be lost.
756 * 0 for success or a negative error code on failure.
758 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
761 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
762 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
763 uint64_t shadow_addr, parent_addr;
765 shadow_addr = amdgpu_bo_gpu_offset(shadow);
766 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
768 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
769 amdgpu_bo_size(shadow), NULL, fence,
774 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
775 * @bo: &amdgpu_bo buffer object to be mapped
776 * @ptr: kernel virtual address to be returned
778 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
779 * amdgpu_bo_kptr() to get the kernel virtual address.
782 * 0 for success or a negative error code on failure.
784 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
789 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
792 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
793 false, MAX_SCHEDULE_TIMEOUT);
797 kptr = amdgpu_bo_kptr(bo);
804 r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
809 *ptr = amdgpu_bo_kptr(bo);
815 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
816 * @bo: &amdgpu_bo buffer object
818 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
821 * the virtual address of a buffer object area.
823 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
827 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
831 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
832 * @bo: &amdgpu_bo buffer object to be unmapped
834 * Unmaps a kernel map set up by amdgpu_bo_kmap().
836 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
839 ttm_bo_kunmap(&bo->kmap);
843 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
844 * @bo: &amdgpu_bo buffer object
846 * References the contained &ttm_buffer_object.
849 * a refcounted pointer to the &amdgpu_bo buffer object.
851 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
856 ttm_bo_get(&bo->tbo);
861 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
862 * @bo: &amdgpu_bo buffer object
864 * Unreferences the contained &ttm_buffer_object and clear the pointer
866 void amdgpu_bo_unref(struct amdgpu_bo **bo)
868 struct ttm_buffer_object *tbo;
879 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
880 * @bo: &amdgpu_bo buffer object to be pinned
881 * @domain: domain to be pinned to
882 * @min_offset: the start of requested address range
883 * @max_offset: the end of requested address range
885 * Pins the buffer object according to requested domain and address range. If
886 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
887 * pin_count and pin_size accordingly.
889 * Pinning means to lock pages in memory along with keeping them at a fixed
890 * offset. It is required when a buffer can not be moved, for example, when
891 * a display buffer is being scanned out.
893 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
894 * where to pin a buffer if there are specific restrictions on where a buffer
898 * 0 for success or a negative error code on failure.
900 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
901 u64 min_offset, u64 max_offset)
903 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
904 struct ttm_operation_ctx ctx = { false, false };
907 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
910 if (WARN_ON_ONCE(min_offset > max_offset))
913 /* Check domain to be pinned to against preferred domains */
914 if (bo->preferred_domains & domain)
915 domain = bo->preferred_domains & domain;
917 /* A shared bo cannot be migrated to VRAM */
918 if (bo->tbo.base.import_attach) {
919 if (domain & AMDGPU_GEM_DOMAIN_GTT)
920 domain = AMDGPU_GEM_DOMAIN_GTT;
925 if (bo->tbo.pin_count) {
926 uint32_t mem_type = bo->tbo.resource->mem_type;
927 uint32_t mem_flags = bo->tbo.resource->placement;
929 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
932 if ((mem_type == TTM_PL_VRAM) &&
933 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
934 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
937 ttm_bo_pin(&bo->tbo);
939 if (max_offset != 0) {
940 u64 domain_start = amdgpu_ttm_domain_start(adev,
942 WARN_ON_ONCE(max_offset <
943 (amdgpu_bo_gpu_offset(bo) - domain_start));
949 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
950 * See function amdgpu_display_supported_domains()
952 domain = amdgpu_bo_get_preferred_domain(adev, domain);
954 if (bo->tbo.base.import_attach)
955 dma_buf_pin(bo->tbo.base.import_attach);
957 /* force to pin into visible video ram */
958 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
959 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
960 amdgpu_bo_placement_from_domain(bo, domain);
961 for (i = 0; i < bo->placement.num_placement; i++) {
962 unsigned int fpfn, lpfn;
964 fpfn = min_offset >> PAGE_SHIFT;
965 lpfn = max_offset >> PAGE_SHIFT;
967 if (fpfn > bo->placements[i].fpfn)
968 bo->placements[i].fpfn = fpfn;
969 if (!bo->placements[i].lpfn ||
970 (lpfn && lpfn < bo->placements[i].lpfn))
971 bo->placements[i].lpfn = lpfn;
973 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS &&
974 bo->placements[i].mem_type == TTM_PL_VRAM)
975 bo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
978 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
980 dev_err(adev->dev, "%p pin failed\n", bo);
984 ttm_bo_pin(&bo->tbo);
986 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
987 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
988 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
989 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
990 &adev->visible_pin_size);
991 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
992 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
1000 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
1001 * @bo: &amdgpu_bo buffer object to be pinned
1002 * @domain: domain to be pinned to
1004 * A simple wrapper to amdgpu_bo_pin_restricted().
1005 * Provides a simpler API for buffers that do not have any strict restrictions
1006 * on where a buffer must be located.
1009 * 0 for success or a negative error code on failure.
1011 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
1013 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1014 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1018 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
1019 * @bo: &amdgpu_bo buffer object to be unpinned
1021 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
1022 * Changes placement and pin size accordingly.
1025 * 0 for success or a negative error code on failure.
1027 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
1029 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1031 ttm_bo_unpin(&bo->tbo);
1032 if (bo->tbo.pin_count)
1035 if (bo->tbo.base.import_attach)
1036 dma_buf_unpin(bo->tbo.base.import_attach);
1038 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1039 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1040 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1041 &adev->visible_pin_size);
1042 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1043 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1048 static const char * const amdgpu_vram_names[] = {
1065 * amdgpu_bo_init - initialize memory manager
1066 * @adev: amdgpu device object
1068 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1071 * 0 for success or a negative error code on failure.
1073 int amdgpu_bo_init(struct amdgpu_device *adev)
1075 /* On A+A platform, VRAM can be mapped as WB */
1076 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1077 /* reserve PAT memory space to WC for VRAM */
1078 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1079 adev->gmc.aper_size);
1082 DRM_ERROR("Unable to set WC memtype for the aperture base\n");
1086 /* Add an MTRR for the VRAM */
1087 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1088 adev->gmc.aper_size);
1091 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1092 adev->gmc.mc_vram_size >> 20,
1093 (unsigned long long)adev->gmc.aper_size >> 20);
1094 DRM_INFO("RAM width %dbits %s\n",
1095 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1096 return amdgpu_ttm_init(adev);
1100 * amdgpu_bo_fini - tear down memory manager
1101 * @adev: amdgpu device object
1103 * Reverses amdgpu_bo_init() to tear down memory manager.
1105 void amdgpu_bo_fini(struct amdgpu_device *adev)
1109 amdgpu_ttm_fini(adev);
1111 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
1112 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1113 arch_phys_wc_del(adev->gmc.vram_mtrr);
1114 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1121 * amdgpu_bo_set_tiling_flags - set tiling flags
1122 * @bo: &amdgpu_bo buffer object
1123 * @tiling_flags: new flags
1125 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1126 * kernel driver to set the tiling flags on a buffer.
1129 * 0 for success or a negative error code on failure.
1131 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1133 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1134 struct amdgpu_bo_user *ubo;
1136 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1137 if (adev->family <= AMDGPU_FAMILY_CZ &&
1138 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1141 ubo = to_amdgpu_bo_user(bo);
1142 ubo->tiling_flags = tiling_flags;
1147 * amdgpu_bo_get_tiling_flags - get tiling flags
1148 * @bo: &amdgpu_bo buffer object
1149 * @tiling_flags: returned flags
1151 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1152 * set the tiling flags on a buffer.
1154 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1156 struct amdgpu_bo_user *ubo;
1158 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1159 dma_resv_assert_held(bo->tbo.base.resv);
1160 ubo = to_amdgpu_bo_user(bo);
1163 *tiling_flags = ubo->tiling_flags;
1167 * amdgpu_bo_set_metadata - set metadata
1168 * @bo: &amdgpu_bo buffer object
1169 * @metadata: new metadata
1170 * @metadata_size: size of the new metadata
1171 * @flags: flags of the new metadata
1173 * Sets buffer object's metadata, its size and flags.
1174 * Used via GEM ioctl.
1177 * 0 for success or a negative error code on failure.
1179 int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata,
1180 u32 metadata_size, uint64_t flags)
1182 struct amdgpu_bo_user *ubo;
1185 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1186 ubo = to_amdgpu_bo_user(bo);
1187 if (!metadata_size) {
1188 if (ubo->metadata_size) {
1189 kfree(ubo->metadata);
1190 ubo->metadata = NULL;
1191 ubo->metadata_size = 0;
1196 if (metadata == NULL)
1199 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1203 kfree(ubo->metadata);
1204 ubo->metadata_flags = flags;
1205 ubo->metadata = buffer;
1206 ubo->metadata_size = metadata_size;
1212 * amdgpu_bo_get_metadata - get metadata
1213 * @bo: &amdgpu_bo buffer object
1214 * @buffer: returned metadata
1215 * @buffer_size: size of the buffer
1216 * @metadata_size: size of the returned metadata
1217 * @flags: flags of the returned metadata
1219 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1220 * less than metadata_size.
1221 * Used via GEM ioctl.
1224 * 0 for success or a negative error code on failure.
1226 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1227 size_t buffer_size, uint32_t *metadata_size,
1230 struct amdgpu_bo_user *ubo;
1232 if (!buffer && !metadata_size)
1235 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1236 ubo = to_amdgpu_bo_user(bo);
1238 *metadata_size = ubo->metadata_size;
1241 if (buffer_size < ubo->metadata_size)
1244 if (ubo->metadata_size)
1245 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1249 *flags = ubo->metadata_flags;
1255 * amdgpu_bo_move_notify - notification about a memory move
1256 * @bo: pointer to a buffer object
1257 * @evict: if this move is evicting the buffer from the graphics address space
1259 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1261 * TTM driver callback which is called when ttm moves a buffer.
1263 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo, bool evict)
1265 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1266 struct amdgpu_bo *abo;
1268 if (!amdgpu_bo_is_amdgpu_bo(bo))
1271 abo = ttm_to_amdgpu_bo(bo);
1272 amdgpu_vm_bo_invalidate(adev, abo, evict);
1274 amdgpu_bo_kunmap(abo);
1276 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1277 bo->resource->mem_type != TTM_PL_SYSTEM)
1278 dma_buf_move_notify(abo->tbo.base.dma_buf);
1280 /* remember the eviction */
1282 atomic64_inc(&adev->num_evictions);
1285 void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
1286 struct amdgpu_mem_stats *stats)
1288 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1289 struct ttm_resource *res = bo->tbo.resource;
1290 uint64_t size = amdgpu_bo_size(bo);
1291 struct drm_gem_object *obj;
1292 unsigned int domain;
1295 /* Abort if the BO doesn't currently have a backing store */
1299 obj = &bo->tbo.base;
1300 shared = drm_gem_object_is_shared_for_memory_stats(obj);
1302 domain = amdgpu_mem_type_to_domain(res->mem_type);
1304 case AMDGPU_GEM_DOMAIN_VRAM:
1305 stats->vram += size;
1306 if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1307 stats->visible_vram += size;
1309 stats->vram_shared += size;
1311 case AMDGPU_GEM_DOMAIN_GTT:
1314 stats->gtt_shared += size;
1316 case AMDGPU_GEM_DOMAIN_CPU:
1320 stats->cpu_shared += size;
1324 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) {
1325 stats->requested_vram += size;
1326 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1327 stats->requested_visible_vram += size;
1329 if (domain != AMDGPU_GEM_DOMAIN_VRAM) {
1330 stats->evicted_vram += size;
1331 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1332 stats->evicted_visible_vram += size;
1334 } else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) {
1335 stats->requested_gtt += size;
1340 * amdgpu_bo_release_notify - notification about a BO being released
1341 * @bo: pointer to a buffer object
1343 * Wipes VRAM buffers whose contents should not be leaked before the
1344 * memory is released.
1346 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1348 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1349 struct dma_fence *fence = NULL;
1350 struct amdgpu_bo *abo;
1353 if (!amdgpu_bo_is_amdgpu_bo(bo))
1356 abo = ttm_to_amdgpu_bo(bo);
1358 WARN_ON(abo->vm_bo);
1361 amdgpu_amdkfd_release_notify(abo);
1363 /* We only remove the fence if the resv has individualized. */
1364 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1365 && bo->base.resv != &bo->base._resv);
1366 if (bo->base.resv == &bo->base._resv)
1367 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1369 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1370 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1371 adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev)))
1374 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1377 r = amdgpu_fill_buffer(abo, 0, bo->base.resv, &fence, true);
1379 amdgpu_vram_mgr_set_cleared(bo->resource);
1380 amdgpu_bo_fence(abo, fence, false);
1381 dma_fence_put(fence);
1384 dma_resv_unlock(bo->base.resv);
1388 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1389 * @bo: pointer to a buffer object
1391 * Notifies the driver we are taking a fault on this BO and have reserved it,
1392 * also performs bookkeeping.
1393 * TTM driver callback for dealing with vm faults.
1396 * 0 for success or a negative error code on failure.
1398 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1400 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1401 struct ttm_operation_ctx ctx = { false, false };
1402 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1405 /* Remember that this BO was accessed by the CPU */
1406 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1408 if (amdgpu_res_cpu_visible(adev, bo->resource))
1411 /* Can't move a pinned BO to visible VRAM */
1412 if (abo->tbo.pin_count > 0)
1413 return VM_FAULT_SIGBUS;
1415 /* hurrah the memory is not visible ! */
1416 atomic64_inc(&adev->num_vram_cpu_page_faults);
1417 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1418 AMDGPU_GEM_DOMAIN_GTT);
1420 /* Avoid costly evictions; only set GTT as a busy placement */
1421 abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
1423 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1424 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1425 return VM_FAULT_NOPAGE;
1426 else if (unlikely(r))
1427 return VM_FAULT_SIGBUS;
1429 /* this should never happen */
1430 if (bo->resource->mem_type == TTM_PL_VRAM &&
1431 !amdgpu_res_cpu_visible(adev, bo->resource))
1432 return VM_FAULT_SIGBUS;
1434 ttm_bo_move_to_lru_tail_unlocked(bo);
1439 * amdgpu_bo_fence - add fence to buffer object
1441 * @bo: buffer object in question
1442 * @fence: fence to add
1443 * @shared: true if fence should be added shared
1446 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1449 struct dma_resv *resv = bo->tbo.base.resv;
1452 r = dma_resv_reserve_fences(resv, 1);
1454 /* As last resort on OOM we block for the fence */
1455 dma_fence_wait(fence, false);
1459 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1460 DMA_RESV_USAGE_WRITE);
1464 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1466 * @adev: amdgpu device pointer
1467 * @resv: reservation object to sync to
1468 * @sync_mode: synchronization mode
1469 * @owner: fence owner
1470 * @intr: Whether the wait is interruptible
1472 * Extract the fences from the reservation object and waits for them to finish.
1475 * 0 on success, errno otherwise.
1477 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1478 enum amdgpu_sync_mode sync_mode, void *owner,
1481 struct amdgpu_sync sync;
1484 amdgpu_sync_create(&sync);
1485 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1486 r = amdgpu_sync_wait(&sync, intr);
1487 amdgpu_sync_free(&sync);
1492 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1493 * @bo: buffer object to wait for
1494 * @owner: fence owner
1495 * @intr: Whether the wait is interruptible
1497 * Wrapper to wait for fences in a BO.
1499 * 0 on success, errno otherwise.
1501 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1503 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1505 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1506 AMDGPU_SYNC_NE_OWNER, owner, intr);
1510 * amdgpu_bo_gpu_offset - return GPU offset of bo
1511 * @bo: amdgpu object for which we query the offset
1513 * Note: object should either be pinned or reserved when calling this
1514 * function, it might be useful to add check for this for debugging.
1517 * current GPU offset of the object.
1519 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1521 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1522 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1523 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1524 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1525 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1526 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1528 return amdgpu_bo_gpu_offset_no_check(bo);
1532 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1533 * @bo: amdgpu object for which we query the offset
1536 * current GPU offset of the object without raising warnings.
1538 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1540 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1541 uint64_t offset = AMDGPU_BO_INVALID_OFFSET;
1543 if (bo->tbo.resource->mem_type == TTM_PL_TT)
1544 offset = amdgpu_gmc_agp_addr(&bo->tbo);
1546 if (offset == AMDGPU_BO_INVALID_OFFSET)
1547 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1548 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1550 return amdgpu_gmc_sign_extend(offset);
1554 * amdgpu_bo_get_preferred_domain - get preferred domain
1555 * @adev: amdgpu device object
1556 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1559 * Which of the allowed domains is preferred for allocating the BO.
1561 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1564 if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1565 ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1566 domain = AMDGPU_GEM_DOMAIN_VRAM;
1567 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1568 domain = AMDGPU_GEM_DOMAIN_GTT;
1573 #if defined(CONFIG_DEBUG_FS)
1574 #define amdgpu_bo_print_flag(m, bo, flag) \
1576 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1577 seq_printf((m), " " #flag); \
1582 * amdgpu_bo_print_info - print BO info in debugfs file
1584 * @id: Index or Id of the BO
1585 * @bo: Requested BO for printing info
1588 * Print BO information in debugfs file
1591 * Size of the BO in bytes.
1593 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1595 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1596 struct dma_buf_attachment *attachment;
1597 struct dma_buf *dma_buf;
1598 const char *placement;
1599 unsigned int pin_count;
1602 if (dma_resv_trylock(bo->tbo.base.resv)) {
1603 unsigned int domain;
1605 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1607 case AMDGPU_GEM_DOMAIN_VRAM:
1608 if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1609 placement = "VRAM VISIBLE";
1613 case AMDGPU_GEM_DOMAIN_GTT:
1616 case AMDGPU_GEM_DOMAIN_CPU:
1621 dma_resv_unlock(bo->tbo.base.resv);
1623 placement = "UNKNOWN";
1626 size = amdgpu_bo_size(bo);
1627 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1628 id, size, placement);
1630 pin_count = READ_ONCE(bo->tbo.pin_count);
1632 seq_printf(m, " pin count %d", pin_count);
1634 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1635 attachment = READ_ONCE(bo->tbo.base.import_attach);
1638 seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino);
1640 seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino);
1642 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1643 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1644 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1645 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1646 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1647 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1648 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);