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 (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
158 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
162 if (domain & AMDGPU_GEM_DOMAIN_DOORBELL) {
165 places[c].mem_type = AMDGPU_PL_DOORBELL;
170 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
174 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
175 AMDGPU_PL_PREEMPT : TTM_PL_TT;
178 * When GTT is just an alternative to VRAM make sure that we
179 * only use it as fallback and still try to fill up VRAM first.
181 if (domain & abo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM)
182 places[c].flags |= TTM_PL_FLAG_FALLBACK;
186 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
189 places[c].mem_type = TTM_PL_SYSTEM;
194 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
197 places[c].mem_type = AMDGPU_PL_GDS;
202 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
205 places[c].mem_type = AMDGPU_PL_GWS;
210 if (domain & AMDGPU_GEM_DOMAIN_OA) {
213 places[c].mem_type = AMDGPU_PL_OA;
221 places[c].mem_type = TTM_PL_SYSTEM;
226 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
228 placement->num_placement = c;
229 placement->placement = places;
233 * amdgpu_bo_create_reserved - create reserved BO for kernel use
235 * @adev: amdgpu device object
236 * @size: size for the new BO
237 * @align: alignment for the new BO
238 * @domain: where to place it
239 * @bo_ptr: used to initialize BOs in structures
240 * @gpu_addr: GPU addr of the pinned BO
241 * @cpu_addr: optional CPU address mapping
243 * Allocates and pins a BO for kernel internal use, and returns it still
246 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
249 * 0 on success, negative error code otherwise.
251 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
252 unsigned long size, int align,
253 u32 domain, struct amdgpu_bo **bo_ptr,
254 u64 *gpu_addr, void **cpu_addr)
256 struct amdgpu_bo_param bp;
261 amdgpu_bo_unref(bo_ptr);
265 memset(&bp, 0, sizeof(bp));
267 bp.byte_align = align;
269 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
270 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
271 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
272 bp.type = ttm_bo_type_kernel;
274 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
277 r = amdgpu_bo_create(adev, &bp, bo_ptr);
279 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
286 r = amdgpu_bo_reserve(*bo_ptr, false);
288 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
292 r = amdgpu_bo_pin(*bo_ptr, domain);
294 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
295 goto error_unreserve;
298 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
300 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
305 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
308 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
310 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
318 amdgpu_bo_unpin(*bo_ptr);
320 amdgpu_bo_unreserve(*bo_ptr);
324 amdgpu_bo_unref(bo_ptr);
330 * amdgpu_bo_create_kernel - create BO for kernel use
332 * @adev: amdgpu device object
333 * @size: size for the new BO
334 * @align: alignment for the new BO
335 * @domain: where to place it
336 * @bo_ptr: used to initialize BOs in structures
337 * @gpu_addr: GPU addr of the pinned BO
338 * @cpu_addr: optional CPU address mapping
340 * Allocates and pins a BO for kernel internal use.
342 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
345 * 0 on success, negative error code otherwise.
347 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
348 unsigned long size, int align,
349 u32 domain, struct amdgpu_bo **bo_ptr,
350 u64 *gpu_addr, void **cpu_addr)
354 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
361 amdgpu_bo_unreserve(*bo_ptr);
367 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
369 * @adev: amdgpu device object
370 * @offset: offset of the BO
371 * @size: size of the BO
372 * @bo_ptr: used to initialize BOs in structures
373 * @cpu_addr: optional CPU address mapping
375 * Creates a kernel BO at a specific offset in VRAM.
378 * 0 on success, negative error code otherwise.
380 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
381 uint64_t offset, uint64_t size,
382 struct amdgpu_bo **bo_ptr, void **cpu_addr)
384 struct ttm_operation_ctx ctx = { false, false };
389 size = ALIGN(size, PAGE_SIZE);
391 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
392 AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
397 if ((*bo_ptr) == NULL)
401 * Remove the original mem node and create a new one at the request
405 amdgpu_bo_kunmap(*bo_ptr);
407 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
409 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
410 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
411 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
413 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
414 &(*bo_ptr)->tbo.resource, &ctx);
419 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
424 amdgpu_bo_unreserve(*bo_ptr);
428 amdgpu_bo_unreserve(*bo_ptr);
429 amdgpu_bo_unref(bo_ptr);
434 * amdgpu_bo_free_kernel - free BO for kernel use
436 * @bo: amdgpu BO to free
437 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
438 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
440 * unmaps and unpin a BO for kernel internal use.
442 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
448 WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend);
450 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
452 amdgpu_bo_kunmap(*bo);
454 amdgpu_bo_unpin(*bo);
455 amdgpu_bo_unreserve(*bo);
466 /* Validate bo size is bit bigger than the request domain */
467 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
468 unsigned long size, u32 domain)
470 struct ttm_resource_manager *man = NULL;
473 * If GTT is part of requested domains the check must succeed to
474 * allow fall back to GTT.
476 if (domain & AMDGPU_GEM_DOMAIN_GTT)
477 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
478 else if (domain & AMDGPU_GEM_DOMAIN_VRAM)
479 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
484 if (domain & AMDGPU_GEM_DOMAIN_GTT)
485 WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
489 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU, _DOMAIN_DOORBELL */
490 if (size < man->size)
493 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, man->size);
497 bool amdgpu_bo_support_uswc(u64 bo_flags)
501 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
502 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
505 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
506 /* Don't try to enable write-combining when it can't work, or things
508 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
511 #ifndef CONFIG_COMPILE_TEST
512 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
513 thanks to write-combining
516 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
517 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
518 "better performance thanks to write-combining\n");
521 /* For architectures that don't support WC memory,
522 * mask out the WC flag from the BO
524 if (!drm_arch_can_wc_memory())
532 * amdgpu_bo_create - create an &amdgpu_bo buffer object
533 * @adev: amdgpu device object
534 * @bp: parameters to be used for the buffer object
535 * @bo_ptr: pointer to the buffer object pointer
537 * Creates an &amdgpu_bo buffer object.
540 * 0 for success or a negative error code on failure.
542 int amdgpu_bo_create(struct amdgpu_device *adev,
543 struct amdgpu_bo_param *bp,
544 struct amdgpu_bo **bo_ptr)
546 struct ttm_operation_ctx ctx = {
547 .interruptible = (bp->type != ttm_bo_type_kernel),
548 .no_wait_gpu = bp->no_wait_gpu,
549 /* We opt to avoid OOM on system pages allocations */
550 .gfp_retry_mayfail = true,
551 .allow_res_evict = bp->type != ttm_bo_type_kernel,
554 struct amdgpu_bo *bo;
555 unsigned long page_align, size = bp->size;
558 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
559 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
560 /* GWS and OA don't need any alignment. */
561 page_align = bp->byte_align;
564 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
565 /* Both size and alignment must be a multiple of 4. */
566 page_align = ALIGN(bp->byte_align, 4);
567 size = ALIGN(size, 4) << PAGE_SHIFT;
569 /* Memory should be aligned at least to a page size. */
570 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
571 size = ALIGN(size, PAGE_SIZE);
574 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
577 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
580 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
583 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
585 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
587 bo->allowed_domains = bo->preferred_domains;
588 if (bp->type != ttm_bo_type_kernel &&
589 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
590 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
591 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
593 bo->flags = bp->flags;
595 if (adev->gmc.mem_partitions)
596 /* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */
597 bo->xcp_id = bp->xcp_id_plus1 - 1;
599 /* For GPUs without spatial partitioning */
602 if (!amdgpu_bo_support_uswc(bo->flags))
603 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
605 bo->flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
607 bo->tbo.bdev = &adev->mman.bdev;
608 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
609 AMDGPU_GEM_DOMAIN_GDS))
610 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
612 amdgpu_bo_placement_from_domain(bo, bp->domain);
613 if (bp->type == ttm_bo_type_kernel)
614 bo->tbo.priority = 2;
615 else if (!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE))
616 bo->tbo.priority = 1;
619 bp->destroy = &amdgpu_bo_destroy;
621 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
622 &bo->placement, page_align, &ctx, NULL,
623 bp->resv, bp->destroy);
624 if (unlikely(r != 0))
627 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
628 amdgpu_res_cpu_visible(adev, bo->tbo.resource))
629 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
632 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
634 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
635 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
636 struct dma_fence *fence;
638 r = amdgpu_ttm_clear_buffer(bo, bo->tbo.base.resv, &fence);
642 dma_resv_add_fence(bo->tbo.base.resv, fence,
643 DMA_RESV_USAGE_KERNEL);
644 dma_fence_put(fence);
647 amdgpu_bo_unreserve(bo);
650 trace_amdgpu_bo_create(bo);
652 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
653 if (bp->type == ttm_bo_type_device)
654 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
660 dma_resv_unlock(bo->tbo.base.resv);
661 amdgpu_bo_unref(&bo);
666 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
667 * @adev: amdgpu device object
668 * @bp: parameters to be used for the buffer object
669 * @ubo_ptr: pointer to the buffer object pointer
671 * Create a BO to be used by user application;
674 * 0 for success or a negative error code on failure.
677 int amdgpu_bo_create_user(struct amdgpu_device *adev,
678 struct amdgpu_bo_param *bp,
679 struct amdgpu_bo_user **ubo_ptr)
681 struct amdgpu_bo *bo_ptr;
684 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
685 bp->destroy = &amdgpu_bo_user_destroy;
686 r = amdgpu_bo_create(adev, bp, &bo_ptr);
690 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
695 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
696 * @adev: amdgpu device object
697 * @bp: parameters to be used for the buffer object
698 * @vmbo_ptr: pointer to the buffer object pointer
700 * Create a BO to be for GPUVM.
703 * 0 for success or a negative error code on failure.
706 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
707 struct amdgpu_bo_param *bp,
708 struct amdgpu_bo_vm **vmbo_ptr)
710 struct amdgpu_bo *bo_ptr;
713 /* bo_ptr_size will be determined by the caller and it depends on
714 * num of amdgpu_vm_pt entries.
716 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
717 r = amdgpu_bo_create(adev, bp, &bo_ptr);
721 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
726 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list
728 * @vmbo: BO that will be inserted into the shadow list
730 * Insert a BO to the shadow list.
732 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo)
734 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev);
736 mutex_lock(&adev->shadow_list_lock);
737 list_add_tail(&vmbo->shadow_list, &adev->shadow_list);
738 vmbo->shadow->parent = amdgpu_bo_ref(&vmbo->bo);
739 vmbo->shadow->tbo.destroy = &amdgpu_bo_vm_destroy;
740 mutex_unlock(&adev->shadow_list_lock);
744 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
746 * @shadow: &amdgpu_bo shadow to be restored
747 * @fence: dma_fence associated with the operation
749 * Copies a buffer object's shadow content back to the object.
750 * This is used for recovering a buffer from its shadow in case of a gpu
751 * reset where vram context may be lost.
754 * 0 for success or a negative error code on failure.
756 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
759 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
760 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
761 uint64_t shadow_addr, parent_addr;
763 shadow_addr = amdgpu_bo_gpu_offset(shadow);
764 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
766 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
767 amdgpu_bo_size(shadow), NULL, fence,
772 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
773 * @bo: &amdgpu_bo buffer object to be mapped
774 * @ptr: kernel virtual address to be returned
776 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
777 * amdgpu_bo_kptr() to get the kernel virtual address.
780 * 0 for success or a negative error code on failure.
782 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
787 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
790 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
791 false, MAX_SCHEDULE_TIMEOUT);
795 kptr = amdgpu_bo_kptr(bo);
802 r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
807 *ptr = amdgpu_bo_kptr(bo);
813 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
814 * @bo: &amdgpu_bo buffer object
816 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
819 * the virtual address of a buffer object area.
821 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
825 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
829 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
830 * @bo: &amdgpu_bo buffer object to be unmapped
832 * Unmaps a kernel map set up by amdgpu_bo_kmap().
834 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
837 ttm_bo_kunmap(&bo->kmap);
841 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
842 * @bo: &amdgpu_bo buffer object
844 * References the contained &ttm_buffer_object.
847 * a refcounted pointer to the &amdgpu_bo buffer object.
849 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
854 ttm_bo_get(&bo->tbo);
859 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
860 * @bo: &amdgpu_bo buffer object
862 * Unreferences the contained &ttm_buffer_object and clear the pointer
864 void amdgpu_bo_unref(struct amdgpu_bo **bo)
866 struct ttm_buffer_object *tbo;
877 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
878 * @bo: &amdgpu_bo buffer object to be pinned
879 * @domain: domain to be pinned to
880 * @min_offset: the start of requested address range
881 * @max_offset: the end of requested address range
883 * Pins the buffer object according to requested domain and address range. If
884 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
885 * pin_count and pin_size accordingly.
887 * Pinning means to lock pages in memory along with keeping them at a fixed
888 * offset. It is required when a buffer can not be moved, for example, when
889 * a display buffer is being scanned out.
891 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
892 * where to pin a buffer if there are specific restrictions on where a buffer
896 * 0 for success or a negative error code on failure.
898 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
899 u64 min_offset, u64 max_offset)
901 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
902 struct ttm_operation_ctx ctx = { false, false };
905 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
908 if (WARN_ON_ONCE(min_offset > max_offset))
911 /* Check domain to be pinned to against preferred domains */
912 if (bo->preferred_domains & domain)
913 domain = bo->preferred_domains & domain;
915 /* A shared bo cannot be migrated to VRAM */
916 if (bo->tbo.base.import_attach) {
917 if (domain & AMDGPU_GEM_DOMAIN_GTT)
918 domain = AMDGPU_GEM_DOMAIN_GTT;
923 if (bo->tbo.pin_count) {
924 uint32_t mem_type = bo->tbo.resource->mem_type;
925 uint32_t mem_flags = bo->tbo.resource->placement;
927 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
930 if ((mem_type == TTM_PL_VRAM) &&
931 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
932 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
935 ttm_bo_pin(&bo->tbo);
937 if (max_offset != 0) {
938 u64 domain_start = amdgpu_ttm_domain_start(adev,
940 WARN_ON_ONCE(max_offset <
941 (amdgpu_bo_gpu_offset(bo) - domain_start));
947 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
948 * See function amdgpu_display_supported_domains()
950 domain = amdgpu_bo_get_preferred_domain(adev, domain);
952 if (bo->tbo.base.import_attach)
953 dma_buf_pin(bo->tbo.base.import_attach);
955 /* force to pin into visible video ram */
956 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
957 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
958 amdgpu_bo_placement_from_domain(bo, domain);
959 for (i = 0; i < bo->placement.num_placement; i++) {
960 unsigned int fpfn, lpfn;
962 fpfn = min_offset >> PAGE_SHIFT;
963 lpfn = max_offset >> PAGE_SHIFT;
965 if (fpfn > bo->placements[i].fpfn)
966 bo->placements[i].fpfn = fpfn;
967 if (!bo->placements[i].lpfn ||
968 (lpfn && lpfn < bo->placements[i].lpfn))
969 bo->placements[i].lpfn = lpfn;
972 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
974 dev_err(adev->dev, "%p pin failed\n", bo);
978 ttm_bo_pin(&bo->tbo);
980 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
981 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
982 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
983 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
984 &adev->visible_pin_size);
985 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
986 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
994 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
995 * @bo: &amdgpu_bo buffer object to be pinned
996 * @domain: domain to be pinned to
998 * A simple wrapper to amdgpu_bo_pin_restricted().
999 * Provides a simpler API for buffers that do not have any strict restrictions
1000 * on where a buffer must be located.
1003 * 0 for success or a negative error code on failure.
1005 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
1007 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1008 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1012 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
1013 * @bo: &amdgpu_bo buffer object to be unpinned
1015 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
1016 * Changes placement and pin size accordingly.
1019 * 0 for success or a negative error code on failure.
1021 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
1023 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1025 ttm_bo_unpin(&bo->tbo);
1026 if (bo->tbo.pin_count)
1029 if (bo->tbo.base.import_attach)
1030 dma_buf_unpin(bo->tbo.base.import_attach);
1032 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1033 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1034 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1035 &adev->visible_pin_size);
1036 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1037 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1042 static const char * const amdgpu_vram_names[] = {
1059 * amdgpu_bo_init - initialize memory manager
1060 * @adev: amdgpu device object
1062 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1065 * 0 for success or a negative error code on failure.
1067 int amdgpu_bo_init(struct amdgpu_device *adev)
1069 /* On A+A platform, VRAM can be mapped as WB */
1070 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1071 /* reserve PAT memory space to WC for VRAM */
1072 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1073 adev->gmc.aper_size);
1076 DRM_ERROR("Unable to set WC memtype for the aperture base\n");
1080 /* Add an MTRR for the VRAM */
1081 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1082 adev->gmc.aper_size);
1085 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1086 adev->gmc.mc_vram_size >> 20,
1087 (unsigned long long)adev->gmc.aper_size >> 20);
1088 DRM_INFO("RAM width %dbits %s\n",
1089 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1090 return amdgpu_ttm_init(adev);
1094 * amdgpu_bo_fini - tear down memory manager
1095 * @adev: amdgpu device object
1097 * Reverses amdgpu_bo_init() to tear down memory manager.
1099 void amdgpu_bo_fini(struct amdgpu_device *adev)
1103 amdgpu_ttm_fini(adev);
1105 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
1106 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1107 arch_phys_wc_del(adev->gmc.vram_mtrr);
1108 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1115 * amdgpu_bo_set_tiling_flags - set tiling flags
1116 * @bo: &amdgpu_bo buffer object
1117 * @tiling_flags: new flags
1119 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1120 * kernel driver to set the tiling flags on a buffer.
1123 * 0 for success or a negative error code on failure.
1125 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1127 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1128 struct amdgpu_bo_user *ubo;
1130 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1131 if (adev->family <= AMDGPU_FAMILY_CZ &&
1132 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1135 ubo = to_amdgpu_bo_user(bo);
1136 ubo->tiling_flags = tiling_flags;
1141 * amdgpu_bo_get_tiling_flags - get tiling flags
1142 * @bo: &amdgpu_bo buffer object
1143 * @tiling_flags: returned flags
1145 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1146 * set the tiling flags on a buffer.
1148 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1150 struct amdgpu_bo_user *ubo;
1152 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1153 dma_resv_assert_held(bo->tbo.base.resv);
1154 ubo = to_amdgpu_bo_user(bo);
1157 *tiling_flags = ubo->tiling_flags;
1161 * amdgpu_bo_set_metadata - set metadata
1162 * @bo: &amdgpu_bo buffer object
1163 * @metadata: new metadata
1164 * @metadata_size: size of the new metadata
1165 * @flags: flags of the new metadata
1167 * Sets buffer object's metadata, its size and flags.
1168 * Used via GEM ioctl.
1171 * 0 for success or a negative error code on failure.
1173 int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata,
1174 u32 metadata_size, uint64_t flags)
1176 struct amdgpu_bo_user *ubo;
1179 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1180 ubo = to_amdgpu_bo_user(bo);
1181 if (!metadata_size) {
1182 if (ubo->metadata_size) {
1183 kfree(ubo->metadata);
1184 ubo->metadata = NULL;
1185 ubo->metadata_size = 0;
1190 if (metadata == NULL)
1193 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1197 kfree(ubo->metadata);
1198 ubo->metadata_flags = flags;
1199 ubo->metadata = buffer;
1200 ubo->metadata_size = metadata_size;
1206 * amdgpu_bo_get_metadata - get metadata
1207 * @bo: &amdgpu_bo buffer object
1208 * @buffer: returned metadata
1209 * @buffer_size: size of the buffer
1210 * @metadata_size: size of the returned metadata
1211 * @flags: flags of the returned metadata
1213 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1214 * less than metadata_size.
1215 * Used via GEM ioctl.
1218 * 0 for success or a negative error code on failure.
1220 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1221 size_t buffer_size, uint32_t *metadata_size,
1224 struct amdgpu_bo_user *ubo;
1226 if (!buffer && !metadata_size)
1229 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1230 ubo = to_amdgpu_bo_user(bo);
1232 *metadata_size = ubo->metadata_size;
1235 if (buffer_size < ubo->metadata_size)
1238 if (ubo->metadata_size)
1239 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1243 *flags = ubo->metadata_flags;
1249 * amdgpu_bo_move_notify - notification about a memory move
1250 * @bo: pointer to a buffer object
1251 * @evict: if this move is evicting the buffer from the graphics address space
1253 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1255 * TTM driver callback which is called when ttm moves a buffer.
1257 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo, bool evict)
1259 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1260 struct amdgpu_bo *abo;
1262 if (!amdgpu_bo_is_amdgpu_bo(bo))
1265 abo = ttm_to_amdgpu_bo(bo);
1266 amdgpu_vm_bo_invalidate(adev, abo, evict);
1268 amdgpu_bo_kunmap(abo);
1270 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1271 bo->resource->mem_type != TTM_PL_SYSTEM)
1272 dma_buf_move_notify(abo->tbo.base.dma_buf);
1274 /* remember the eviction */
1276 atomic64_inc(&adev->num_evictions);
1279 void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
1280 struct amdgpu_mem_stats *stats)
1282 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1283 struct ttm_resource *res = bo->tbo.resource;
1284 uint64_t size = amdgpu_bo_size(bo);
1285 struct drm_gem_object *obj;
1286 unsigned int domain;
1289 /* Abort if the BO doesn't currently have a backing store */
1293 obj = &bo->tbo.base;
1294 shared = drm_gem_object_is_shared_for_memory_stats(obj);
1296 domain = amdgpu_mem_type_to_domain(res->mem_type);
1298 case AMDGPU_GEM_DOMAIN_VRAM:
1299 stats->vram += size;
1300 if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1301 stats->visible_vram += size;
1303 stats->vram_shared += size;
1305 case AMDGPU_GEM_DOMAIN_GTT:
1308 stats->gtt_shared += size;
1310 case AMDGPU_GEM_DOMAIN_CPU:
1314 stats->cpu_shared += size;
1318 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) {
1319 stats->requested_vram += size;
1320 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1321 stats->requested_visible_vram += size;
1323 if (domain != AMDGPU_GEM_DOMAIN_VRAM) {
1324 stats->evicted_vram += size;
1325 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1326 stats->evicted_visible_vram += size;
1328 } else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) {
1329 stats->requested_gtt += size;
1334 * amdgpu_bo_release_notify - notification about a BO being released
1335 * @bo: pointer to a buffer object
1337 * Wipes VRAM buffers whose contents should not be leaked before the
1338 * memory is released.
1340 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1342 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1343 struct dma_fence *fence = NULL;
1344 struct amdgpu_bo *abo;
1347 if (!amdgpu_bo_is_amdgpu_bo(bo))
1350 abo = ttm_to_amdgpu_bo(bo);
1352 WARN_ON(abo->vm_bo);
1355 amdgpu_amdkfd_release_notify(abo);
1357 /* We only remove the fence if the resv has individualized. */
1358 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1359 && bo->base.resv != &bo->base._resv);
1360 if (bo->base.resv == &bo->base._resv)
1361 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1363 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1364 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1365 adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev)))
1368 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1371 r = amdgpu_fill_buffer(abo, 0, bo->base.resv, &fence, true);
1373 amdgpu_vram_mgr_set_cleared(bo->resource);
1374 amdgpu_bo_fence(abo, fence, false);
1375 dma_fence_put(fence);
1378 dma_resv_unlock(bo->base.resv);
1382 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1383 * @bo: pointer to a buffer object
1385 * Notifies the driver we are taking a fault on this BO and have reserved it,
1386 * also performs bookkeeping.
1387 * TTM driver callback for dealing with vm faults.
1390 * 0 for success or a negative error code on failure.
1392 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1394 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1395 struct ttm_operation_ctx ctx = { false, false };
1396 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1399 /* Remember that this BO was accessed by the CPU */
1400 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1402 if (amdgpu_res_cpu_visible(adev, bo->resource))
1405 /* Can't move a pinned BO to visible VRAM */
1406 if (abo->tbo.pin_count > 0)
1407 return VM_FAULT_SIGBUS;
1409 /* hurrah the memory is not visible ! */
1410 atomic64_inc(&adev->num_vram_cpu_page_faults);
1411 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1412 AMDGPU_GEM_DOMAIN_GTT);
1414 /* Avoid costly evictions; only set GTT as a busy placement */
1415 abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
1417 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1418 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1419 return VM_FAULT_NOPAGE;
1420 else if (unlikely(r))
1421 return VM_FAULT_SIGBUS;
1423 /* this should never happen */
1424 if (bo->resource->mem_type == TTM_PL_VRAM &&
1425 !amdgpu_res_cpu_visible(adev, bo->resource))
1426 return VM_FAULT_SIGBUS;
1428 ttm_bo_move_to_lru_tail_unlocked(bo);
1433 * amdgpu_bo_fence - add fence to buffer object
1435 * @bo: buffer object in question
1436 * @fence: fence to add
1437 * @shared: true if fence should be added shared
1440 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1443 struct dma_resv *resv = bo->tbo.base.resv;
1446 r = dma_resv_reserve_fences(resv, 1);
1448 /* As last resort on OOM we block for the fence */
1449 dma_fence_wait(fence, false);
1453 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1454 DMA_RESV_USAGE_WRITE);
1458 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1460 * @adev: amdgpu device pointer
1461 * @resv: reservation object to sync to
1462 * @sync_mode: synchronization mode
1463 * @owner: fence owner
1464 * @intr: Whether the wait is interruptible
1466 * Extract the fences from the reservation object and waits for them to finish.
1469 * 0 on success, errno otherwise.
1471 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1472 enum amdgpu_sync_mode sync_mode, void *owner,
1475 struct amdgpu_sync sync;
1478 amdgpu_sync_create(&sync);
1479 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1480 r = amdgpu_sync_wait(&sync, intr);
1481 amdgpu_sync_free(&sync);
1486 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1487 * @bo: buffer object to wait for
1488 * @owner: fence owner
1489 * @intr: Whether the wait is interruptible
1491 * Wrapper to wait for fences in a BO.
1493 * 0 on success, errno otherwise.
1495 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1497 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1499 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1500 AMDGPU_SYNC_NE_OWNER, owner, intr);
1504 * amdgpu_bo_gpu_offset - return GPU offset of bo
1505 * @bo: amdgpu object for which we query the offset
1507 * Note: object should either be pinned or reserved when calling this
1508 * function, it might be useful to add check for this for debugging.
1511 * current GPU offset of the object.
1513 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1515 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1516 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1517 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1518 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1519 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1520 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1522 return amdgpu_bo_gpu_offset_no_check(bo);
1526 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1527 * @bo: amdgpu object for which we query the offset
1530 * current GPU offset of the object without raising warnings.
1532 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1534 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1535 uint64_t offset = AMDGPU_BO_INVALID_OFFSET;
1537 if (bo->tbo.resource->mem_type == TTM_PL_TT)
1538 offset = amdgpu_gmc_agp_addr(&bo->tbo);
1540 if (offset == AMDGPU_BO_INVALID_OFFSET)
1541 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1542 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1544 return amdgpu_gmc_sign_extend(offset);
1548 * amdgpu_bo_get_preferred_domain - get preferred domain
1549 * @adev: amdgpu device object
1550 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1553 * Which of the allowed domains is preferred for allocating the BO.
1555 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1558 if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1559 ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1560 domain = AMDGPU_GEM_DOMAIN_VRAM;
1561 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1562 domain = AMDGPU_GEM_DOMAIN_GTT;
1567 #if defined(CONFIG_DEBUG_FS)
1568 #define amdgpu_bo_print_flag(m, bo, flag) \
1570 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1571 seq_printf((m), " " #flag); \
1576 * amdgpu_bo_print_info - print BO info in debugfs file
1578 * @id: Index or Id of the BO
1579 * @bo: Requested BO for printing info
1582 * Print BO information in debugfs file
1585 * Size of the BO in bytes.
1587 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1589 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1590 struct dma_buf_attachment *attachment;
1591 struct dma_buf *dma_buf;
1592 const char *placement;
1593 unsigned int pin_count;
1596 if (dma_resv_trylock(bo->tbo.base.resv)) {
1597 unsigned int domain;
1599 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1601 case AMDGPU_GEM_DOMAIN_VRAM:
1602 if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1603 placement = "VRAM VISIBLE";
1607 case AMDGPU_GEM_DOMAIN_GTT:
1610 case AMDGPU_GEM_DOMAIN_CPU:
1615 dma_resv_unlock(bo->tbo.base.resv);
1617 placement = "UNKNOWN";
1620 size = amdgpu_bo_size(bo);
1621 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1622 id, size, placement);
1624 pin_count = READ_ONCE(bo->tbo.pin_count);
1626 seq_printf(m, " pin count %d", pin_count);
1628 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1629 attachment = READ_ONCE(bo->tbo.base.import_attach);
1632 seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino);
1634 seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino);
1636 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1637 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1638 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1639 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1640 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1641 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1642 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);