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"
46 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
47 * represents memory used by driver (VRAM, system memory, etc.). The driver
48 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
49 * to create/destroy/set buffer object which are then managed by the kernel TTM
51 * The interfaces are also used internally by kernel clients, including gfx,
52 * uvd, etc. for kernel managed allocations used by the GPU.
56 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
58 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
62 if (bo->tbo.base.import_attach)
63 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
64 drm_gem_object_release(&bo->tbo.base);
65 amdgpu_bo_unref(&bo->parent);
69 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
71 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
72 struct amdgpu_bo_user *ubo;
74 ubo = to_amdgpu_bo_user(bo);
76 amdgpu_bo_destroy(tbo);
79 static void amdgpu_bo_vm_destroy(struct ttm_buffer_object *tbo)
81 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
82 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
83 struct amdgpu_bo_vm *vmbo;
85 vmbo = to_amdgpu_bo_vm(bo);
86 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
87 if (!list_empty(&vmbo->shadow_list)) {
88 mutex_lock(&adev->shadow_list_lock);
89 list_del_init(&vmbo->shadow_list);
90 mutex_unlock(&adev->shadow_list_lock);
93 amdgpu_bo_destroy(tbo);
97 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
98 * @bo: buffer object to be checked
100 * Uses destroy function associated with the object to determine if this is
104 * true if the object belongs to &amdgpu_bo, false if not.
106 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
108 if (bo->destroy == &amdgpu_bo_destroy ||
109 bo->destroy == &amdgpu_bo_user_destroy ||
110 bo->destroy == &amdgpu_bo_vm_destroy)
117 * amdgpu_bo_placement_from_domain - set buffer's placement
118 * @abo: &amdgpu_bo buffer object whose placement is to be set
119 * @domain: requested domain
121 * Sets buffer's placement according to requested domain and the buffer's
124 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
126 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
127 struct ttm_placement *placement = &abo->placement;
128 struct ttm_place *places = abo->placements;
129 u64 flags = abo->flags;
132 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
133 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
137 places[c].mem_type = TTM_PL_VRAM;
140 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
141 places[c].lpfn = visible_pfn;
143 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
145 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
146 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
150 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
154 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
155 AMDGPU_PL_PREEMPT : TTM_PL_TT;
160 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
163 places[c].mem_type = TTM_PL_SYSTEM;
168 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
171 places[c].mem_type = AMDGPU_PL_GDS;
176 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
179 places[c].mem_type = AMDGPU_PL_GWS;
184 if (domain & AMDGPU_GEM_DOMAIN_OA) {
187 places[c].mem_type = AMDGPU_PL_OA;
195 places[c].mem_type = TTM_PL_SYSTEM;
200 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
202 placement->num_placement = c;
203 placement->placement = places;
205 placement->num_busy_placement = c;
206 placement->busy_placement = places;
210 * amdgpu_bo_create_reserved - create reserved BO for kernel use
212 * @adev: amdgpu device object
213 * @size: size for the new BO
214 * @align: alignment for the new BO
215 * @domain: where to place it
216 * @bo_ptr: used to initialize BOs in structures
217 * @gpu_addr: GPU addr of the pinned BO
218 * @cpu_addr: optional CPU address mapping
220 * Allocates and pins a BO for kernel internal use, and returns it still
223 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
226 * 0 on success, negative error code otherwise.
228 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
229 unsigned long size, int align,
230 u32 domain, struct amdgpu_bo **bo_ptr,
231 u64 *gpu_addr, void **cpu_addr)
233 struct amdgpu_bo_param bp;
238 amdgpu_bo_unref(bo_ptr);
242 memset(&bp, 0, sizeof(bp));
244 bp.byte_align = align;
246 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
247 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
248 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
249 bp.type = ttm_bo_type_kernel;
251 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
254 r = amdgpu_bo_create(adev, &bp, bo_ptr);
256 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
263 r = amdgpu_bo_reserve(*bo_ptr, false);
265 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
269 r = amdgpu_bo_pin(*bo_ptr, domain);
271 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
272 goto error_unreserve;
275 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
277 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
282 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
285 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
287 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
295 amdgpu_bo_unpin(*bo_ptr);
297 amdgpu_bo_unreserve(*bo_ptr);
301 amdgpu_bo_unref(bo_ptr);
307 * amdgpu_bo_create_kernel - create BO for kernel use
309 * @adev: amdgpu device object
310 * @size: size for the new BO
311 * @align: alignment for the new BO
312 * @domain: where to place it
313 * @bo_ptr: used to initialize BOs in structures
314 * @gpu_addr: GPU addr of the pinned BO
315 * @cpu_addr: optional CPU address mapping
317 * Allocates and pins a BO for kernel internal use.
319 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
322 * 0 on success, negative error code otherwise.
324 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
325 unsigned long size, int align,
326 u32 domain, struct amdgpu_bo **bo_ptr,
327 u64 *gpu_addr, void **cpu_addr)
331 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
338 amdgpu_bo_unreserve(*bo_ptr);
344 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
346 * @adev: amdgpu device object
347 * @offset: offset of the BO
348 * @size: size of the BO
349 * @domain: where to place it
350 * @bo_ptr: used to initialize BOs in structures
351 * @cpu_addr: optional CPU address mapping
353 * Creates a kernel BO at a specific offset in the address space of the domain.
356 * 0 on success, negative error code otherwise.
358 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
359 uint64_t offset, uint64_t size, uint32_t domain,
360 struct amdgpu_bo **bo_ptr, void **cpu_addr)
362 struct ttm_operation_ctx ctx = { false, false };
367 size = ALIGN(size, PAGE_SIZE);
369 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
374 if ((*bo_ptr) == NULL)
378 * Remove the original mem node and create a new one at the request
382 amdgpu_bo_kunmap(*bo_ptr);
384 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
386 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
387 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
388 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
390 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
391 &(*bo_ptr)->tbo.resource, &ctx);
396 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
401 amdgpu_bo_unreserve(*bo_ptr);
405 amdgpu_bo_unreserve(*bo_ptr);
406 amdgpu_bo_unref(bo_ptr);
411 * amdgpu_bo_free_kernel - free BO for kernel use
413 * @bo: amdgpu BO to free
414 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
415 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
417 * unmaps and unpin a BO for kernel internal use.
419 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
425 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
427 amdgpu_bo_kunmap(*bo);
429 amdgpu_bo_unpin(*bo);
430 amdgpu_bo_unreserve(*bo);
441 /* Validate bo size is bit bigger then the request domain */
442 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
443 unsigned long size, u32 domain)
445 struct ttm_resource_manager *man = NULL;
448 * If GTT is part of requested domains the check must succeed to
449 * allow fall back to GTT
451 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
452 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
454 if (size < (man->size << PAGE_SHIFT))
460 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
461 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
463 if (size < (man->size << PAGE_SHIFT))
470 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
474 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
475 man->size << PAGE_SHIFT);
479 bool amdgpu_bo_support_uswc(u64 bo_flags)
483 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
484 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
487 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
488 /* Don't try to enable write-combining when it can't work, or things
490 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
493 #ifndef CONFIG_COMPILE_TEST
494 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
495 thanks to write-combining
498 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
499 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
500 "better performance thanks to write-combining\n");
503 /* For architectures that don't support WC memory,
504 * mask out the WC flag from the BO
506 if (!drm_arch_can_wc_memory())
514 * amdgpu_bo_create - create an &amdgpu_bo buffer object
515 * @adev: amdgpu device object
516 * @bp: parameters to be used for the buffer object
517 * @bo_ptr: pointer to the buffer object pointer
519 * Creates an &amdgpu_bo buffer object.
522 * 0 for success or a negative error code on failure.
524 int amdgpu_bo_create(struct amdgpu_device *adev,
525 struct amdgpu_bo_param *bp,
526 struct amdgpu_bo **bo_ptr)
528 struct ttm_operation_ctx ctx = {
529 .interruptible = (bp->type != ttm_bo_type_kernel),
530 .no_wait_gpu = bp->no_wait_gpu,
531 /* We opt to avoid OOM on system pages allocations */
532 .gfp_retry_mayfail = true,
533 .allow_res_evict = bp->type != ttm_bo_type_kernel,
536 struct amdgpu_bo *bo;
537 unsigned long page_align, size = bp->size;
540 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
541 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
542 /* GWS and OA don't need any alignment. */
543 page_align = bp->byte_align;
545 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
546 /* Both size and alignment must be a multiple of 4. */
547 page_align = ALIGN(bp->byte_align, 4);
548 size = ALIGN(size, 4) << PAGE_SHIFT;
550 /* Memory should be aligned at least to a page size. */
551 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
552 size = ALIGN(size, PAGE_SIZE);
555 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
558 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
561 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
564 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
566 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
568 bo->allowed_domains = bo->preferred_domains;
569 if (bp->type != ttm_bo_type_kernel &&
570 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
571 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
573 bo->flags = bp->flags;
575 if (!amdgpu_bo_support_uswc(bo->flags))
576 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
578 bo->tbo.bdev = &adev->mman.bdev;
579 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
580 AMDGPU_GEM_DOMAIN_GDS))
581 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
583 amdgpu_bo_placement_from_domain(bo, bp->domain);
584 if (bp->type == ttm_bo_type_kernel)
585 bo->tbo.priority = 1;
588 bp->destroy = &amdgpu_bo_destroy;
590 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
591 &bo->placement, page_align, &ctx, NULL,
592 bp->resv, bp->destroy);
593 if (unlikely(r != 0))
596 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
597 bo->tbo.resource->mem_type == TTM_PL_VRAM &&
598 bo->tbo.resource->start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
599 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
602 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
604 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
605 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
606 struct dma_fence *fence;
608 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
612 amdgpu_bo_fence(bo, fence, false);
613 dma_fence_put(bo->tbo.moving);
614 bo->tbo.moving = dma_fence_get(fence);
615 dma_fence_put(fence);
618 amdgpu_bo_unreserve(bo);
621 trace_amdgpu_bo_create(bo);
623 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
624 if (bp->type == ttm_bo_type_device)
625 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
631 dma_resv_unlock(bo->tbo.base.resv);
632 amdgpu_bo_unref(&bo);
637 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
638 * @adev: amdgpu device object
639 * @bp: parameters to be used for the buffer object
640 * @ubo_ptr: pointer to the buffer object pointer
642 * Create a BO to be used by user application;
645 * 0 for success or a negative error code on failure.
648 int amdgpu_bo_create_user(struct amdgpu_device *adev,
649 struct amdgpu_bo_param *bp,
650 struct amdgpu_bo_user **ubo_ptr)
652 struct amdgpu_bo *bo_ptr;
655 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
656 bp->destroy = &amdgpu_bo_user_destroy;
657 r = amdgpu_bo_create(adev, bp, &bo_ptr);
661 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
666 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
667 * @adev: amdgpu device object
668 * @bp: parameters to be used for the buffer object
669 * @vmbo_ptr: pointer to the buffer object pointer
671 * Create a BO to be for GPUVM.
674 * 0 for success or a negative error code on failure.
677 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
678 struct amdgpu_bo_param *bp,
679 struct amdgpu_bo_vm **vmbo_ptr)
681 struct amdgpu_bo *bo_ptr;
684 /* bo_ptr_size will be determined by the caller and it depends on
685 * num of amdgpu_vm_pt entries.
687 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
688 bp->destroy = &amdgpu_bo_vm_destroy;
689 r = amdgpu_bo_create(adev, bp, &bo_ptr);
693 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
694 INIT_LIST_HEAD(&(*vmbo_ptr)->shadow_list);
699 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list
701 * @vmbo: BO that will be inserted into the shadow list
703 * Insert a BO to the shadow list.
705 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo)
707 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev);
709 mutex_lock(&adev->shadow_list_lock);
710 list_add_tail(&vmbo->shadow_list, &adev->shadow_list);
711 mutex_unlock(&adev->shadow_list_lock);
715 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
717 * @shadow: &amdgpu_bo shadow to be restored
718 * @fence: dma_fence associated with the operation
720 * Copies a buffer object's shadow content back to the object.
721 * This is used for recovering a buffer from its shadow in case of a gpu
722 * reset where vram context may be lost.
725 * 0 for success or a negative error code on failure.
727 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
730 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
731 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
732 uint64_t shadow_addr, parent_addr;
734 shadow_addr = amdgpu_bo_gpu_offset(shadow);
735 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
737 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
738 amdgpu_bo_size(shadow), NULL, fence,
743 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
744 * @bo: &amdgpu_bo buffer object to be mapped
745 * @ptr: kernel virtual address to be returned
747 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
748 * amdgpu_bo_kptr() to get the kernel virtual address.
751 * 0 for success or a negative error code on failure.
753 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
758 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
761 kptr = amdgpu_bo_kptr(bo);
768 r = dma_resv_wait_timeout(bo->tbo.base.resv, false, false,
769 MAX_SCHEDULE_TIMEOUT);
773 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap);
778 *ptr = amdgpu_bo_kptr(bo);
784 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
785 * @bo: &amdgpu_bo buffer object
787 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
790 * the virtual address of a buffer object area.
792 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
796 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
800 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
801 * @bo: &amdgpu_bo buffer object to be unmapped
803 * Unmaps a kernel map set up by amdgpu_bo_kmap().
805 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
808 ttm_bo_kunmap(&bo->kmap);
812 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
813 * @bo: &amdgpu_bo buffer object
815 * References the contained &ttm_buffer_object.
818 * a refcounted pointer to the &amdgpu_bo buffer object.
820 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
825 ttm_bo_get(&bo->tbo);
830 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
831 * @bo: &amdgpu_bo buffer object
833 * Unreferences the contained &ttm_buffer_object and clear the pointer
835 void amdgpu_bo_unref(struct amdgpu_bo **bo)
837 struct ttm_buffer_object *tbo;
848 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
849 * @bo: &amdgpu_bo buffer object to be pinned
850 * @domain: domain to be pinned to
851 * @min_offset: the start of requested address range
852 * @max_offset: the end of requested address range
854 * Pins the buffer object according to requested domain and address range. If
855 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
856 * pin_count and pin_size accordingly.
858 * Pinning means to lock pages in memory along with keeping them at a fixed
859 * offset. It is required when a buffer can not be moved, for example, when
860 * a display buffer is being scanned out.
862 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
863 * where to pin a buffer if there are specific restrictions on where a buffer
867 * 0 for success or a negative error code on failure.
869 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
870 u64 min_offset, u64 max_offset)
872 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
873 struct ttm_operation_ctx ctx = { false, false };
876 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
879 if (WARN_ON_ONCE(min_offset > max_offset))
882 /* A shared bo cannot be migrated to VRAM */
883 if (bo->tbo.base.import_attach) {
884 if (domain & AMDGPU_GEM_DOMAIN_GTT)
885 domain = AMDGPU_GEM_DOMAIN_GTT;
890 if (bo->tbo.pin_count) {
891 uint32_t mem_type = bo->tbo.resource->mem_type;
892 uint32_t mem_flags = bo->tbo.resource->placement;
894 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
897 if ((mem_type == TTM_PL_VRAM) &&
898 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
899 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
902 ttm_bo_pin(&bo->tbo);
904 if (max_offset != 0) {
905 u64 domain_start = amdgpu_ttm_domain_start(adev,
907 WARN_ON_ONCE(max_offset <
908 (amdgpu_bo_gpu_offset(bo) - domain_start));
914 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
915 * See function amdgpu_display_supported_domains()
917 domain = amdgpu_bo_get_preferred_domain(adev, domain);
919 if (bo->tbo.base.import_attach)
920 dma_buf_pin(bo->tbo.base.import_attach);
922 /* force to pin into visible video ram */
923 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
924 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
925 amdgpu_bo_placement_from_domain(bo, domain);
926 for (i = 0; i < bo->placement.num_placement; i++) {
929 fpfn = min_offset >> PAGE_SHIFT;
930 lpfn = max_offset >> PAGE_SHIFT;
932 if (fpfn > bo->placements[i].fpfn)
933 bo->placements[i].fpfn = fpfn;
934 if (!bo->placements[i].lpfn ||
935 (lpfn && lpfn < bo->placements[i].lpfn))
936 bo->placements[i].lpfn = lpfn;
939 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
941 dev_err(adev->dev, "%p pin failed\n", bo);
945 ttm_bo_pin(&bo->tbo);
947 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
948 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
949 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
950 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
951 &adev->visible_pin_size);
952 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
953 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
961 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
962 * @bo: &amdgpu_bo buffer object to be pinned
963 * @domain: domain to be pinned to
965 * A simple wrapper to amdgpu_bo_pin_restricted().
966 * Provides a simpler API for buffers that do not have any strict restrictions
967 * on where a buffer must be located.
970 * 0 for success or a negative error code on failure.
972 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
974 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
975 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
979 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
980 * @bo: &amdgpu_bo buffer object to be unpinned
982 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
983 * Changes placement and pin size accordingly.
986 * 0 for success or a negative error code on failure.
988 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
990 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
992 ttm_bo_unpin(&bo->tbo);
993 if (bo->tbo.pin_count)
996 if (bo->tbo.base.import_attach)
997 dma_buf_unpin(bo->tbo.base.import_attach);
999 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1000 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1001 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1002 &adev->visible_pin_size);
1003 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1004 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1008 static const char *amdgpu_vram_names[] = {
1023 * amdgpu_bo_init - initialize memory manager
1024 * @adev: amdgpu device object
1026 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1029 * 0 for success or a negative error code on failure.
1031 int amdgpu_bo_init(struct amdgpu_device *adev)
1033 /* On A+A platform, VRAM can be mapped as WB */
1034 if (!adev->gmc.xgmi.connected_to_cpu) {
1035 /* reserve PAT memory space to WC for VRAM */
1036 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1037 adev->gmc.aper_size);
1040 DRM_ERROR("Unable to set WC memtype for the aperture base\n");
1044 /* Add an MTRR for the VRAM */
1045 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1046 adev->gmc.aper_size);
1049 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1050 adev->gmc.mc_vram_size >> 20,
1051 (unsigned long long)adev->gmc.aper_size >> 20);
1052 DRM_INFO("RAM width %dbits %s\n",
1053 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1054 return amdgpu_ttm_init(adev);
1058 * amdgpu_bo_fini - tear down memory manager
1059 * @adev: amdgpu device object
1061 * Reverses amdgpu_bo_init() to tear down memory manager.
1063 void amdgpu_bo_fini(struct amdgpu_device *adev)
1067 amdgpu_ttm_fini(adev);
1069 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
1071 if (!adev->gmc.xgmi.connected_to_cpu) {
1072 arch_phys_wc_del(adev->gmc.vram_mtrr);
1073 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1080 * amdgpu_bo_set_tiling_flags - set tiling flags
1081 * @bo: &amdgpu_bo buffer object
1082 * @tiling_flags: new flags
1084 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1085 * kernel driver to set the tiling flags on a buffer.
1088 * 0 for success or a negative error code on failure.
1090 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1092 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1093 struct amdgpu_bo_user *ubo;
1095 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1096 if (adev->family <= AMDGPU_FAMILY_CZ &&
1097 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1100 ubo = to_amdgpu_bo_user(bo);
1101 ubo->tiling_flags = tiling_flags;
1106 * amdgpu_bo_get_tiling_flags - get tiling flags
1107 * @bo: &amdgpu_bo buffer object
1108 * @tiling_flags: returned flags
1110 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1111 * set the tiling flags on a buffer.
1113 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1115 struct amdgpu_bo_user *ubo;
1117 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1118 dma_resv_assert_held(bo->tbo.base.resv);
1119 ubo = to_amdgpu_bo_user(bo);
1122 *tiling_flags = ubo->tiling_flags;
1126 * amdgpu_bo_set_metadata - set metadata
1127 * @bo: &amdgpu_bo buffer object
1128 * @metadata: new metadata
1129 * @metadata_size: size of the new metadata
1130 * @flags: flags of the new metadata
1132 * Sets buffer object's metadata, its size and flags.
1133 * Used via GEM ioctl.
1136 * 0 for success or a negative error code on failure.
1138 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1139 uint32_t metadata_size, uint64_t flags)
1141 struct amdgpu_bo_user *ubo;
1144 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1145 ubo = to_amdgpu_bo_user(bo);
1146 if (!metadata_size) {
1147 if (ubo->metadata_size) {
1148 kfree(ubo->metadata);
1149 ubo->metadata = NULL;
1150 ubo->metadata_size = 0;
1155 if (metadata == NULL)
1158 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1162 kfree(ubo->metadata);
1163 ubo->metadata_flags = flags;
1164 ubo->metadata = buffer;
1165 ubo->metadata_size = metadata_size;
1171 * amdgpu_bo_get_metadata - get metadata
1172 * @bo: &amdgpu_bo buffer object
1173 * @buffer: returned metadata
1174 * @buffer_size: size of the buffer
1175 * @metadata_size: size of the returned metadata
1176 * @flags: flags of the returned metadata
1178 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1179 * less than metadata_size.
1180 * Used via GEM ioctl.
1183 * 0 for success or a negative error code on failure.
1185 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1186 size_t buffer_size, uint32_t *metadata_size,
1189 struct amdgpu_bo_user *ubo;
1191 if (!buffer && !metadata_size)
1194 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1195 ubo = to_amdgpu_bo_user(bo);
1197 *metadata_size = ubo->metadata_size;
1200 if (buffer_size < ubo->metadata_size)
1203 if (ubo->metadata_size)
1204 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1208 *flags = ubo->metadata_flags;
1214 * amdgpu_bo_move_notify - notification about a memory move
1215 * @bo: pointer to a buffer object
1216 * @evict: if this move is evicting the buffer from the graphics address space
1217 * @new_mem: new information of the bufer object
1219 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1221 * TTM driver callback which is called when ttm moves a buffer.
1223 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1225 struct ttm_resource *new_mem)
1227 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1228 struct amdgpu_bo *abo;
1229 struct ttm_resource *old_mem = bo->resource;
1231 if (!amdgpu_bo_is_amdgpu_bo(bo))
1234 abo = ttm_to_amdgpu_bo(bo);
1235 amdgpu_vm_bo_invalidate(adev, abo, evict);
1237 amdgpu_bo_kunmap(abo);
1239 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1240 bo->resource->mem_type != TTM_PL_SYSTEM)
1241 dma_buf_move_notify(abo->tbo.base.dma_buf);
1243 /* remember the eviction */
1245 atomic64_inc(&adev->num_evictions);
1247 /* update statistics */
1251 /* move_notify is called before move happens */
1252 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1255 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1256 uint64_t *gtt_mem, uint64_t *cpu_mem)
1258 unsigned int domain;
1260 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1262 case AMDGPU_GEM_DOMAIN_VRAM:
1263 *vram_mem += amdgpu_bo_size(bo);
1265 case AMDGPU_GEM_DOMAIN_GTT:
1266 *gtt_mem += amdgpu_bo_size(bo);
1268 case AMDGPU_GEM_DOMAIN_CPU:
1270 *cpu_mem += amdgpu_bo_size(bo);
1276 * amdgpu_bo_release_notify - notification about a BO being released
1277 * @bo: pointer to a buffer object
1279 * Wipes VRAM buffers whose contents should not be leaked before the
1280 * memory is released.
1282 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1284 struct dma_fence *fence = NULL;
1285 struct amdgpu_bo *abo;
1288 if (!amdgpu_bo_is_amdgpu_bo(bo))
1291 abo = ttm_to_amdgpu_bo(bo);
1294 amdgpu_amdkfd_release_notify(abo);
1296 /* We only remove the fence if the resv has individualized. */
1297 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1298 && bo->base.resv != &bo->base._resv);
1299 if (bo->base.resv == &bo->base._resv)
1300 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1302 if (bo->resource->mem_type != TTM_PL_VRAM ||
1303 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1306 dma_resv_lock(bo->base.resv, NULL);
1308 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1310 amdgpu_bo_fence(abo, fence, false);
1311 dma_fence_put(fence);
1314 dma_resv_unlock(bo->base.resv);
1318 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1319 * @bo: pointer to a buffer object
1321 * Notifies the driver we are taking a fault on this BO and have reserved it,
1322 * also performs bookkeeping.
1323 * TTM driver callback for dealing with vm faults.
1326 * 0 for success or a negative error code on failure.
1328 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1330 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1331 struct ttm_operation_ctx ctx = { false, false };
1332 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1333 unsigned long offset;
1336 /* Remember that this BO was accessed by the CPU */
1337 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1339 if (bo->resource->mem_type != TTM_PL_VRAM)
1342 offset = bo->resource->start << PAGE_SHIFT;
1343 if ((offset + bo->base.size) <= adev->gmc.visible_vram_size)
1346 /* Can't move a pinned BO to visible VRAM */
1347 if (abo->tbo.pin_count > 0)
1348 return VM_FAULT_SIGBUS;
1350 /* hurrah the memory is not visible ! */
1351 atomic64_inc(&adev->num_vram_cpu_page_faults);
1352 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1353 AMDGPU_GEM_DOMAIN_GTT);
1355 /* Avoid costly evictions; only set GTT as a busy placement */
1356 abo->placement.num_busy_placement = 1;
1357 abo->placement.busy_placement = &abo->placements[1];
1359 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1360 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1361 return VM_FAULT_NOPAGE;
1362 else if (unlikely(r))
1363 return VM_FAULT_SIGBUS;
1365 offset = bo->resource->start << PAGE_SHIFT;
1366 /* this should never happen */
1367 if (bo->resource->mem_type == TTM_PL_VRAM &&
1368 (offset + bo->base.size) > adev->gmc.visible_vram_size)
1369 return VM_FAULT_SIGBUS;
1371 ttm_bo_move_to_lru_tail_unlocked(bo);
1376 * amdgpu_bo_fence - add fence to buffer object
1378 * @bo: buffer object in question
1379 * @fence: fence to add
1380 * @shared: true if fence should be added shared
1383 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1386 struct dma_resv *resv = bo->tbo.base.resv;
1389 dma_resv_add_shared_fence(resv, fence);
1391 dma_resv_add_excl_fence(resv, fence);
1395 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1397 * @adev: amdgpu device pointer
1398 * @resv: reservation object to sync to
1399 * @sync_mode: synchronization mode
1400 * @owner: fence owner
1401 * @intr: Whether the wait is interruptible
1403 * Extract the fences from the reservation object and waits for them to finish.
1406 * 0 on success, errno otherwise.
1408 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1409 enum amdgpu_sync_mode sync_mode, void *owner,
1412 struct amdgpu_sync sync;
1415 amdgpu_sync_create(&sync);
1416 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1417 r = amdgpu_sync_wait(&sync, intr);
1418 amdgpu_sync_free(&sync);
1423 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1424 * @bo: buffer object to wait for
1425 * @owner: fence owner
1426 * @intr: Whether the wait is interruptible
1428 * Wrapper to wait for fences in a BO.
1430 * 0 on success, errno otherwise.
1432 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1434 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1436 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1437 AMDGPU_SYNC_NE_OWNER, owner, intr);
1441 * amdgpu_bo_gpu_offset - return GPU offset of bo
1442 * @bo: amdgpu object for which we query the offset
1444 * Note: object should either be pinned or reserved when calling this
1445 * function, it might be useful to add check for this for debugging.
1448 * current GPU offset of the object.
1450 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1452 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1453 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1454 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1455 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1456 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1457 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1459 return amdgpu_bo_gpu_offset_no_check(bo);
1463 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1464 * @bo: amdgpu object for which we query the offset
1467 * current GPU offset of the object without raising warnings.
1469 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1471 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1474 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1475 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1477 return amdgpu_gmc_sign_extend(offset);
1481 * amdgpu_bo_get_preferred_domain - get preferred domain
1482 * @adev: amdgpu device object
1483 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1486 * Which of the allowed domains is preferred for allocating the BO.
1488 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1491 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1492 domain = AMDGPU_GEM_DOMAIN_VRAM;
1493 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1494 domain = AMDGPU_GEM_DOMAIN_GTT;
1499 #if defined(CONFIG_DEBUG_FS)
1500 #define amdgpu_bo_print_flag(m, bo, flag) \
1502 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1503 seq_printf((m), " " #flag); \
1508 * amdgpu_bo_print_info - print BO info in debugfs file
1510 * @id: Index or Id of the BO
1511 * @bo: Requested BO for printing info
1514 * Print BO information in debugfs file
1517 * Size of the BO in bytes.
1519 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1521 struct dma_buf_attachment *attachment;
1522 struct dma_buf *dma_buf;
1523 unsigned int domain;
1524 const char *placement;
1525 unsigned int pin_count;
1528 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1530 case AMDGPU_GEM_DOMAIN_VRAM:
1533 case AMDGPU_GEM_DOMAIN_GTT:
1536 case AMDGPU_GEM_DOMAIN_CPU:
1542 size = amdgpu_bo_size(bo);
1543 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1544 id, size, placement);
1546 pin_count = READ_ONCE(bo->tbo.pin_count);
1548 seq_printf(m, " pin count %d", pin_count);
1550 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1551 attachment = READ_ONCE(bo->tbo.base.import_attach);
1554 seq_printf(m, " imported from %p", dma_buf);
1556 seq_printf(m, " exported as %p", dma_buf);
1558 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1559 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1560 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1561 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1562 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1563 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1564 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);