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/amdgpu_drm.h>
37 #include <drm/drm_cache.h>
39 #include "amdgpu_trace.h"
40 #include "amdgpu_amdkfd.h"
45 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
46 * represents memory used by driver (VRAM, system memory, etc.). The driver
47 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
48 * to create/destroy/set buffer object which are then managed by the kernel TTM
50 * The interfaces are also used internally by kernel clients, including gfx,
51 * uvd, etc. for kernel managed allocations used by the GPU.
55 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
57 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
61 if (bo->tbo.base.import_attach)
62 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
63 drm_gem_object_release(&bo->tbo.base);
64 amdgpu_bo_unref(&bo->parent);
68 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
70 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
71 struct amdgpu_bo_user *ubo;
73 ubo = to_amdgpu_bo_user(bo);
75 amdgpu_bo_destroy(tbo);
78 static void amdgpu_bo_vm_destroy(struct ttm_buffer_object *tbo)
80 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
81 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
82 struct amdgpu_bo_vm *vmbo;
84 vmbo = to_amdgpu_bo_vm(bo);
85 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
86 if (!list_empty(&vmbo->shadow_list)) {
87 mutex_lock(&adev->shadow_list_lock);
88 list_del_init(&vmbo->shadow_list);
89 mutex_unlock(&adev->shadow_list_lock);
92 amdgpu_bo_destroy(tbo);
96 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
97 * @bo: buffer object to be checked
99 * Uses destroy function associated with the object to determine if this is
103 * true if the object belongs to &amdgpu_bo, false if not.
105 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
107 if (bo->destroy == &amdgpu_bo_destroy ||
108 bo->destroy == &amdgpu_bo_user_destroy ||
109 bo->destroy == &amdgpu_bo_vm_destroy)
116 * amdgpu_bo_placement_from_domain - set buffer's placement
117 * @abo: &amdgpu_bo buffer object whose placement is to be set
118 * @domain: requested domain
120 * Sets buffer's placement according to requested domain and the buffer's
123 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
125 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
126 struct ttm_placement *placement = &abo->placement;
127 struct ttm_place *places = abo->placements;
128 u64 flags = abo->flags;
131 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
132 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
136 places[c].mem_type = TTM_PL_VRAM;
139 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
140 places[c].lpfn = visible_pfn;
142 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
144 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
145 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
149 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
153 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
154 AMDGPU_PL_PREEMPT : TTM_PL_TT;
159 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
162 places[c].mem_type = TTM_PL_SYSTEM;
167 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
170 places[c].mem_type = AMDGPU_PL_GDS;
175 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
178 places[c].mem_type = AMDGPU_PL_GWS;
183 if (domain & AMDGPU_GEM_DOMAIN_OA) {
186 places[c].mem_type = AMDGPU_PL_OA;
194 places[c].mem_type = TTM_PL_SYSTEM;
199 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
201 placement->num_placement = c;
202 placement->placement = places;
204 placement->num_busy_placement = c;
205 placement->busy_placement = places;
209 * amdgpu_bo_create_reserved - create reserved BO for kernel use
211 * @adev: amdgpu device object
212 * @size: size for the new BO
213 * @align: alignment for the new BO
214 * @domain: where to place it
215 * @bo_ptr: used to initialize BOs in structures
216 * @gpu_addr: GPU addr of the pinned BO
217 * @cpu_addr: optional CPU address mapping
219 * Allocates and pins a BO for kernel internal use, and returns it still
222 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
225 * 0 on success, negative error code otherwise.
227 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
228 unsigned long size, int align,
229 u32 domain, struct amdgpu_bo **bo_ptr,
230 u64 *gpu_addr, void **cpu_addr)
232 struct amdgpu_bo_param bp;
237 amdgpu_bo_unref(bo_ptr);
241 memset(&bp, 0, sizeof(bp));
243 bp.byte_align = align;
245 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
246 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
247 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
248 bp.type = ttm_bo_type_kernel;
250 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
253 r = amdgpu_bo_create(adev, &bp, bo_ptr);
255 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
262 r = amdgpu_bo_reserve(*bo_ptr, false);
264 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
268 r = amdgpu_bo_pin(*bo_ptr, domain);
270 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
271 goto error_unreserve;
274 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
276 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
281 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
284 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
286 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
294 amdgpu_bo_unpin(*bo_ptr);
296 amdgpu_bo_unreserve(*bo_ptr);
300 amdgpu_bo_unref(bo_ptr);
306 * amdgpu_bo_create_kernel - create BO for kernel use
308 * @adev: amdgpu device object
309 * @size: size for the new BO
310 * @align: alignment for the new BO
311 * @domain: where to place it
312 * @bo_ptr: used to initialize BOs in structures
313 * @gpu_addr: GPU addr of the pinned BO
314 * @cpu_addr: optional CPU address mapping
316 * Allocates and pins a BO for kernel internal use.
318 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
321 * 0 on success, negative error code otherwise.
323 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
324 unsigned long size, int align,
325 u32 domain, struct amdgpu_bo **bo_ptr,
326 u64 *gpu_addr, void **cpu_addr)
330 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
337 amdgpu_bo_unreserve(*bo_ptr);
343 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
345 * @adev: amdgpu device object
346 * @offset: offset of the BO
347 * @size: size of the BO
348 * @domain: where to place it
349 * @bo_ptr: used to initialize BOs in structures
350 * @cpu_addr: optional CPU address mapping
352 * Creates a kernel BO at a specific offset in the address space of the domain.
355 * 0 on success, negative error code otherwise.
357 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
358 uint64_t offset, uint64_t size, uint32_t domain,
359 struct amdgpu_bo **bo_ptr, void **cpu_addr)
361 struct ttm_operation_ctx ctx = { false, false };
366 size = ALIGN(size, PAGE_SIZE);
368 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
373 if ((*bo_ptr) == NULL)
377 * Remove the original mem node and create a new one at the request
381 amdgpu_bo_kunmap(*bo_ptr);
383 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
385 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
386 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
387 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
389 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
390 &(*bo_ptr)->tbo.resource, &ctx);
395 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
400 amdgpu_bo_unreserve(*bo_ptr);
404 amdgpu_bo_unreserve(*bo_ptr);
405 amdgpu_bo_unref(bo_ptr);
410 * amdgpu_bo_free_kernel - free BO for kernel use
412 * @bo: amdgpu BO to free
413 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
414 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
416 * unmaps and unpin a BO for kernel internal use.
418 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
424 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
426 amdgpu_bo_kunmap(*bo);
428 amdgpu_bo_unpin(*bo);
429 amdgpu_bo_unreserve(*bo);
440 /* Validate bo size is bit bigger then the request domain */
441 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
442 unsigned long size, u32 domain)
444 struct ttm_resource_manager *man = NULL;
447 * If GTT is part of requested domains the check must succeed to
448 * allow fall back to GTT
450 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
451 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
453 if (size < (man->size << PAGE_SHIFT))
459 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
460 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
462 if (size < (man->size << PAGE_SHIFT))
469 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
473 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
474 man->size << PAGE_SHIFT);
478 bool amdgpu_bo_support_uswc(u64 bo_flags)
482 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
483 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
486 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
487 /* Don't try to enable write-combining when it can't work, or things
489 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
492 #ifndef CONFIG_COMPILE_TEST
493 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
494 thanks to write-combining
497 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
498 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
499 "better performance thanks to write-combining\n");
502 /* For architectures that don't support WC memory,
503 * mask out the WC flag from the BO
505 if (!drm_arch_can_wc_memory())
513 * amdgpu_bo_create - create an &amdgpu_bo buffer object
514 * @adev: amdgpu device object
515 * @bp: parameters to be used for the buffer object
516 * @bo_ptr: pointer to the buffer object pointer
518 * Creates an &amdgpu_bo buffer object.
521 * 0 for success or a negative error code on failure.
523 int amdgpu_bo_create(struct amdgpu_device *adev,
524 struct amdgpu_bo_param *bp,
525 struct amdgpu_bo **bo_ptr)
527 struct ttm_operation_ctx ctx = {
528 .interruptible = (bp->type != ttm_bo_type_kernel),
529 .no_wait_gpu = bp->no_wait_gpu,
530 /* We opt to avoid OOM on system pages allocations */
531 .gfp_retry_mayfail = true,
532 .allow_res_evict = bp->type != ttm_bo_type_kernel,
535 struct amdgpu_bo *bo;
536 unsigned long page_align, size = bp->size;
539 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
540 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
541 /* GWS and OA don't need any alignment. */
542 page_align = bp->byte_align;
544 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
545 /* Both size and alignment must be a multiple of 4. */
546 page_align = ALIGN(bp->byte_align, 4);
547 size = ALIGN(size, 4) << PAGE_SHIFT;
549 /* Memory should be aligned at least to a page size. */
550 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
551 size = ALIGN(size, PAGE_SIZE);
554 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
557 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
560 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
563 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
565 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
567 bo->allowed_domains = bo->preferred_domains;
568 if (bp->type != ttm_bo_type_kernel &&
569 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
570 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
572 bo->flags = bp->flags;
574 if (!amdgpu_bo_support_uswc(bo->flags))
575 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
577 bo->tbo.bdev = &adev->mman.bdev;
578 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
579 AMDGPU_GEM_DOMAIN_GDS))
580 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
582 amdgpu_bo_placement_from_domain(bo, bp->domain);
583 if (bp->type == ttm_bo_type_kernel)
584 bo->tbo.priority = 1;
587 bp->destroy = &amdgpu_bo_destroy;
589 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
590 &bo->placement, page_align, &ctx, NULL,
591 bp->resv, bp->destroy);
592 if (unlikely(r != 0))
595 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
596 bo->tbo.resource->mem_type == TTM_PL_VRAM &&
597 bo->tbo.resource->start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
598 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
601 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
603 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
604 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
605 struct dma_fence *fence;
607 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
611 amdgpu_bo_fence(bo, fence, false);
612 dma_fence_put(bo->tbo.moving);
613 bo->tbo.moving = dma_fence_get(fence);
614 dma_fence_put(fence);
617 amdgpu_bo_unreserve(bo);
620 trace_amdgpu_bo_create(bo);
622 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
623 if (bp->type == ttm_bo_type_device)
624 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
630 dma_resv_unlock(bo->tbo.base.resv);
631 amdgpu_bo_unref(&bo);
636 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
637 * @adev: amdgpu device object
638 * @bp: parameters to be used for the buffer object
639 * @ubo_ptr: pointer to the buffer object pointer
641 * Create a BO to be used by user application;
644 * 0 for success or a negative error code on failure.
647 int amdgpu_bo_create_user(struct amdgpu_device *adev,
648 struct amdgpu_bo_param *bp,
649 struct amdgpu_bo_user **ubo_ptr)
651 struct amdgpu_bo *bo_ptr;
654 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
655 bp->destroy = &amdgpu_bo_user_destroy;
656 r = amdgpu_bo_create(adev, bp, &bo_ptr);
660 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
665 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
666 * @adev: amdgpu device object
667 * @bp: parameters to be used for the buffer object
668 * @vmbo_ptr: pointer to the buffer object pointer
670 * Create a BO to be for GPUVM.
673 * 0 for success or a negative error code on failure.
676 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
677 struct amdgpu_bo_param *bp,
678 struct amdgpu_bo_vm **vmbo_ptr)
680 struct amdgpu_bo *bo_ptr;
683 /* bo_ptr_size will be determined by the caller and it depends on
684 * num of amdgpu_vm_pt entries.
686 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
687 bp->destroy = &amdgpu_bo_vm_destroy;
688 r = amdgpu_bo_create(adev, bp, &bo_ptr);
692 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
693 INIT_LIST_HEAD(&(*vmbo_ptr)->shadow_list);
698 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list
700 * @vmbo: BO that will be inserted into the shadow list
702 * Insert a BO to the shadow list.
704 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo)
706 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev);
708 mutex_lock(&adev->shadow_list_lock);
709 list_add_tail(&vmbo->shadow_list, &adev->shadow_list);
710 mutex_unlock(&adev->shadow_list_lock);
714 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
716 * @shadow: &amdgpu_bo shadow to be restored
717 * @fence: dma_fence associated with the operation
719 * Copies a buffer object's shadow content back to the object.
720 * This is used for recovering a buffer from its shadow in case of a gpu
721 * reset where vram context may be lost.
724 * 0 for success or a negative error code on failure.
726 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
729 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
730 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
731 uint64_t shadow_addr, parent_addr;
733 shadow_addr = amdgpu_bo_gpu_offset(shadow);
734 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
736 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
737 amdgpu_bo_size(shadow), NULL, fence,
742 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
743 * @bo: &amdgpu_bo buffer object to be mapped
744 * @ptr: kernel virtual address to be returned
746 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
747 * amdgpu_bo_kptr() to get the kernel virtual address.
750 * 0 for success or a negative error code on failure.
752 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
757 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
760 kptr = amdgpu_bo_kptr(bo);
767 r = dma_resv_wait_timeout(bo->tbo.base.resv, false, false,
768 MAX_SCHEDULE_TIMEOUT);
772 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap);
777 *ptr = amdgpu_bo_kptr(bo);
783 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
784 * @bo: &amdgpu_bo buffer object
786 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
789 * the virtual address of a buffer object area.
791 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
795 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
799 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
800 * @bo: &amdgpu_bo buffer object to be unmapped
802 * Unmaps a kernel map set up by amdgpu_bo_kmap().
804 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
807 ttm_bo_kunmap(&bo->kmap);
811 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
812 * @bo: &amdgpu_bo buffer object
814 * References the contained &ttm_buffer_object.
817 * a refcounted pointer to the &amdgpu_bo buffer object.
819 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
824 ttm_bo_get(&bo->tbo);
829 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
830 * @bo: &amdgpu_bo buffer object
832 * Unreferences the contained &ttm_buffer_object and clear the pointer
834 void amdgpu_bo_unref(struct amdgpu_bo **bo)
836 struct ttm_buffer_object *tbo;
847 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
848 * @bo: &amdgpu_bo buffer object to be pinned
849 * @domain: domain to be pinned to
850 * @min_offset: the start of requested address range
851 * @max_offset: the end of requested address range
853 * Pins the buffer object according to requested domain and address range. If
854 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
855 * pin_count and pin_size accordingly.
857 * Pinning means to lock pages in memory along with keeping them at a fixed
858 * offset. It is required when a buffer can not be moved, for example, when
859 * a display buffer is being scanned out.
861 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
862 * where to pin a buffer if there are specific restrictions on where a buffer
866 * 0 for success or a negative error code on failure.
868 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
869 u64 min_offset, u64 max_offset)
871 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
872 struct ttm_operation_ctx ctx = { false, false };
875 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
878 if (WARN_ON_ONCE(min_offset > max_offset))
881 /* A shared bo cannot be migrated to VRAM */
882 if (bo->tbo.base.import_attach) {
883 if (domain & AMDGPU_GEM_DOMAIN_GTT)
884 domain = AMDGPU_GEM_DOMAIN_GTT;
889 if (bo->tbo.pin_count) {
890 uint32_t mem_type = bo->tbo.resource->mem_type;
891 uint32_t mem_flags = bo->tbo.resource->placement;
893 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
896 if ((mem_type == TTM_PL_VRAM) &&
897 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
898 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
901 ttm_bo_pin(&bo->tbo);
903 if (max_offset != 0) {
904 u64 domain_start = amdgpu_ttm_domain_start(adev,
906 WARN_ON_ONCE(max_offset <
907 (amdgpu_bo_gpu_offset(bo) - domain_start));
913 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
914 * See function amdgpu_display_supported_domains()
916 domain = amdgpu_bo_get_preferred_domain(adev, domain);
918 if (bo->tbo.base.import_attach)
919 dma_buf_pin(bo->tbo.base.import_attach);
921 /* force to pin into visible video ram */
922 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
923 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
924 amdgpu_bo_placement_from_domain(bo, domain);
925 for (i = 0; i < bo->placement.num_placement; i++) {
928 fpfn = min_offset >> PAGE_SHIFT;
929 lpfn = max_offset >> PAGE_SHIFT;
931 if (fpfn > bo->placements[i].fpfn)
932 bo->placements[i].fpfn = fpfn;
933 if (!bo->placements[i].lpfn ||
934 (lpfn && lpfn < bo->placements[i].lpfn))
935 bo->placements[i].lpfn = lpfn;
938 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
940 dev_err(adev->dev, "%p pin failed\n", bo);
944 ttm_bo_pin(&bo->tbo);
946 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
947 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
948 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
949 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
950 &adev->visible_pin_size);
951 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
952 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
960 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
961 * @bo: &amdgpu_bo buffer object to be pinned
962 * @domain: domain to be pinned to
964 * A simple wrapper to amdgpu_bo_pin_restricted().
965 * Provides a simpler API for buffers that do not have any strict restrictions
966 * on where a buffer must be located.
969 * 0 for success or a negative error code on failure.
971 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
973 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
974 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
978 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
979 * @bo: &amdgpu_bo buffer object to be unpinned
981 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
982 * Changes placement and pin size accordingly.
985 * 0 for success or a negative error code on failure.
987 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
989 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
991 ttm_bo_unpin(&bo->tbo);
992 if (bo->tbo.pin_count)
995 if (bo->tbo.base.import_attach)
996 dma_buf_unpin(bo->tbo.base.import_attach);
998 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
999 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1000 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1001 &adev->visible_pin_size);
1002 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1003 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1007 static const char *amdgpu_vram_names[] = {
1022 * amdgpu_bo_init - initialize memory manager
1023 * @adev: amdgpu device object
1025 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1028 * 0 for success or a negative error code on failure.
1030 int amdgpu_bo_init(struct amdgpu_device *adev)
1032 /* On A+A platform, VRAM can be mapped as WB */
1033 if (!adev->gmc.xgmi.connected_to_cpu) {
1034 /* reserve PAT memory space to WC for VRAM */
1035 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1036 adev->gmc.aper_size);
1038 /* Add an MTRR for the VRAM */
1039 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1040 adev->gmc.aper_size);
1043 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1044 adev->gmc.mc_vram_size >> 20,
1045 (unsigned long long)adev->gmc.aper_size >> 20);
1046 DRM_INFO("RAM width %dbits %s\n",
1047 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1048 return amdgpu_ttm_init(adev);
1052 * amdgpu_bo_fini - tear down memory manager
1053 * @adev: amdgpu device object
1055 * Reverses amdgpu_bo_init() to tear down memory manager.
1057 void amdgpu_bo_fini(struct amdgpu_device *adev)
1059 amdgpu_ttm_fini(adev);
1063 * amdgpu_bo_set_tiling_flags - set tiling flags
1064 * @bo: &amdgpu_bo buffer object
1065 * @tiling_flags: new flags
1067 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1068 * kernel driver to set the tiling flags on a buffer.
1071 * 0 for success or a negative error code on failure.
1073 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1075 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1076 struct amdgpu_bo_user *ubo;
1078 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1079 if (adev->family <= AMDGPU_FAMILY_CZ &&
1080 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1083 ubo = to_amdgpu_bo_user(bo);
1084 ubo->tiling_flags = tiling_flags;
1089 * amdgpu_bo_get_tiling_flags - get tiling flags
1090 * @bo: &amdgpu_bo buffer object
1091 * @tiling_flags: returned flags
1093 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1094 * set the tiling flags on a buffer.
1096 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1098 struct amdgpu_bo_user *ubo;
1100 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1101 dma_resv_assert_held(bo->tbo.base.resv);
1102 ubo = to_amdgpu_bo_user(bo);
1105 *tiling_flags = ubo->tiling_flags;
1109 * amdgpu_bo_set_metadata - set metadata
1110 * @bo: &amdgpu_bo buffer object
1111 * @metadata: new metadata
1112 * @metadata_size: size of the new metadata
1113 * @flags: flags of the new metadata
1115 * Sets buffer object's metadata, its size and flags.
1116 * Used via GEM ioctl.
1119 * 0 for success or a negative error code on failure.
1121 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1122 uint32_t metadata_size, uint64_t flags)
1124 struct amdgpu_bo_user *ubo;
1127 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1128 ubo = to_amdgpu_bo_user(bo);
1129 if (!metadata_size) {
1130 if (ubo->metadata_size) {
1131 kfree(ubo->metadata);
1132 ubo->metadata = NULL;
1133 ubo->metadata_size = 0;
1138 if (metadata == NULL)
1141 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1145 kfree(ubo->metadata);
1146 ubo->metadata_flags = flags;
1147 ubo->metadata = buffer;
1148 ubo->metadata_size = metadata_size;
1154 * amdgpu_bo_get_metadata - get metadata
1155 * @bo: &amdgpu_bo buffer object
1156 * @buffer: returned metadata
1157 * @buffer_size: size of the buffer
1158 * @metadata_size: size of the returned metadata
1159 * @flags: flags of the returned metadata
1161 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1162 * less than metadata_size.
1163 * Used via GEM ioctl.
1166 * 0 for success or a negative error code on failure.
1168 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1169 size_t buffer_size, uint32_t *metadata_size,
1172 struct amdgpu_bo_user *ubo;
1174 if (!buffer && !metadata_size)
1177 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1178 ubo = to_amdgpu_bo_user(bo);
1180 *metadata_size = ubo->metadata_size;
1183 if (buffer_size < ubo->metadata_size)
1186 if (ubo->metadata_size)
1187 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1191 *flags = ubo->metadata_flags;
1197 * amdgpu_bo_move_notify - notification about a memory move
1198 * @bo: pointer to a buffer object
1199 * @evict: if this move is evicting the buffer from the graphics address space
1200 * @new_mem: new information of the bufer object
1202 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1204 * TTM driver callback which is called when ttm moves a buffer.
1206 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1208 struct ttm_resource *new_mem)
1210 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1211 struct amdgpu_bo *abo;
1212 struct ttm_resource *old_mem = bo->resource;
1214 if (!amdgpu_bo_is_amdgpu_bo(bo))
1217 abo = ttm_to_amdgpu_bo(bo);
1218 amdgpu_vm_bo_invalidate(adev, abo, evict);
1220 amdgpu_bo_kunmap(abo);
1222 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1223 bo->resource->mem_type != TTM_PL_SYSTEM)
1224 dma_buf_move_notify(abo->tbo.base.dma_buf);
1226 /* remember the eviction */
1228 atomic64_inc(&adev->num_evictions);
1230 /* update statistics */
1234 /* move_notify is called before move happens */
1235 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1238 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1239 uint64_t *gtt_mem, uint64_t *cpu_mem)
1241 unsigned int domain;
1243 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1245 case AMDGPU_GEM_DOMAIN_VRAM:
1246 *vram_mem += amdgpu_bo_size(bo);
1248 case AMDGPU_GEM_DOMAIN_GTT:
1249 *gtt_mem += amdgpu_bo_size(bo);
1251 case AMDGPU_GEM_DOMAIN_CPU:
1253 *cpu_mem += amdgpu_bo_size(bo);
1259 * amdgpu_bo_release_notify - notification about a BO being released
1260 * @bo: pointer to a buffer object
1262 * Wipes VRAM buffers whose contents should not be leaked before the
1263 * memory is released.
1265 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1267 struct dma_fence *fence = NULL;
1268 struct amdgpu_bo *abo;
1271 if (!amdgpu_bo_is_amdgpu_bo(bo))
1274 abo = ttm_to_amdgpu_bo(bo);
1277 amdgpu_amdkfd_unreserve_memory_limit(abo);
1279 /* We only remove the fence if the resv has individualized. */
1280 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1281 && bo->base.resv != &bo->base._resv);
1282 if (bo->base.resv == &bo->base._resv)
1283 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1285 if (bo->resource->mem_type != TTM_PL_VRAM ||
1286 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1289 dma_resv_lock(bo->base.resv, NULL);
1291 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1293 amdgpu_bo_fence(abo, fence, false);
1294 dma_fence_put(fence);
1297 dma_resv_unlock(bo->base.resv);
1301 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1302 * @bo: pointer to a buffer object
1304 * Notifies the driver we are taking a fault on this BO and have reserved it,
1305 * also performs bookkeeping.
1306 * TTM driver callback for dealing with vm faults.
1309 * 0 for success or a negative error code on failure.
1311 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1313 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1314 struct ttm_operation_ctx ctx = { false, false };
1315 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1316 unsigned long offset;
1319 /* Remember that this BO was accessed by the CPU */
1320 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1322 if (bo->resource->mem_type != TTM_PL_VRAM)
1325 offset = bo->resource->start << PAGE_SHIFT;
1326 if ((offset + bo->base.size) <= adev->gmc.visible_vram_size)
1329 /* Can't move a pinned BO to visible VRAM */
1330 if (abo->tbo.pin_count > 0)
1331 return VM_FAULT_SIGBUS;
1333 /* hurrah the memory is not visible ! */
1334 atomic64_inc(&adev->num_vram_cpu_page_faults);
1335 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1336 AMDGPU_GEM_DOMAIN_GTT);
1338 /* Avoid costly evictions; only set GTT as a busy placement */
1339 abo->placement.num_busy_placement = 1;
1340 abo->placement.busy_placement = &abo->placements[1];
1342 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1343 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1344 return VM_FAULT_NOPAGE;
1345 else if (unlikely(r))
1346 return VM_FAULT_SIGBUS;
1348 offset = bo->resource->start << PAGE_SHIFT;
1349 /* this should never happen */
1350 if (bo->resource->mem_type == TTM_PL_VRAM &&
1351 (offset + bo->base.size) > adev->gmc.visible_vram_size)
1352 return VM_FAULT_SIGBUS;
1354 ttm_bo_move_to_lru_tail_unlocked(bo);
1359 * amdgpu_bo_fence - add fence to buffer object
1361 * @bo: buffer object in question
1362 * @fence: fence to add
1363 * @shared: true if fence should be added shared
1366 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1369 struct dma_resv *resv = bo->tbo.base.resv;
1372 dma_resv_add_shared_fence(resv, fence);
1374 dma_resv_add_excl_fence(resv, fence);
1378 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1380 * @adev: amdgpu device pointer
1381 * @resv: reservation object to sync to
1382 * @sync_mode: synchronization mode
1383 * @owner: fence owner
1384 * @intr: Whether the wait is interruptible
1386 * Extract the fences from the reservation object and waits for them to finish.
1389 * 0 on success, errno otherwise.
1391 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1392 enum amdgpu_sync_mode sync_mode, void *owner,
1395 struct amdgpu_sync sync;
1398 amdgpu_sync_create(&sync);
1399 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1400 r = amdgpu_sync_wait(&sync, intr);
1401 amdgpu_sync_free(&sync);
1406 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1407 * @bo: buffer object to wait for
1408 * @owner: fence owner
1409 * @intr: Whether the wait is interruptible
1411 * Wrapper to wait for fences in a BO.
1413 * 0 on success, errno otherwise.
1415 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1417 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1419 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1420 AMDGPU_SYNC_NE_OWNER, owner, intr);
1424 * amdgpu_bo_gpu_offset - return GPU offset of bo
1425 * @bo: amdgpu object for which we query the offset
1427 * Note: object should either be pinned or reserved when calling this
1428 * function, it might be useful to add check for this for debugging.
1431 * current GPU offset of the object.
1433 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1435 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1436 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1437 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1438 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1439 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1440 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1442 return amdgpu_bo_gpu_offset_no_check(bo);
1446 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1447 * @bo: amdgpu object for which we query the offset
1450 * current GPU offset of the object without raising warnings.
1452 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1454 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1457 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1458 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1460 return amdgpu_gmc_sign_extend(offset);
1464 * amdgpu_bo_get_preferred_domain - get preferred domain
1465 * @adev: amdgpu device object
1466 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1469 * Which of the allowed domains is preferred for allocating the BO.
1471 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1474 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1475 domain = AMDGPU_GEM_DOMAIN_VRAM;
1476 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1477 domain = AMDGPU_GEM_DOMAIN_GTT;
1482 #if defined(CONFIG_DEBUG_FS)
1483 #define amdgpu_bo_print_flag(m, bo, flag) \
1485 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1486 seq_printf((m), " " #flag); \
1491 * amdgpu_bo_print_info - print BO info in debugfs file
1493 * @id: Index or Id of the BO
1494 * @bo: Requested BO for printing info
1497 * Print BO information in debugfs file
1500 * Size of the BO in bytes.
1502 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1504 struct dma_buf_attachment *attachment;
1505 struct dma_buf *dma_buf;
1506 unsigned int domain;
1507 const char *placement;
1508 unsigned int pin_count;
1511 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1513 case AMDGPU_GEM_DOMAIN_VRAM:
1516 case AMDGPU_GEM_DOMAIN_GTT:
1519 case AMDGPU_GEM_DOMAIN_CPU:
1525 size = amdgpu_bo_size(bo);
1526 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1527 id, size, placement);
1529 pin_count = READ_ONCE(bo->tbo.pin_count);
1531 seq_printf(m, " pin count %d", pin_count);
1533 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1534 attachment = READ_ONCE(bo->tbo.base.import_attach);
1537 seq_printf(m, " imported from %p", dma_buf);
1539 seq_printf(m, " exported as %p", dma_buf);
1541 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1542 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1543 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1544 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1545 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1546 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1547 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);