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.
56 * amdgpu_bo_subtract_pin_size - Remove BO from pin_size accounting
58 * @bo: &amdgpu_bo buffer object
60 * This function is called when a BO stops being pinned, and updates the
61 * &amdgpu_device pin_size values accordingly.
63 static void amdgpu_bo_subtract_pin_size(struct amdgpu_bo *bo)
65 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
67 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
68 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
69 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
70 &adev->visible_pin_size);
71 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
72 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
76 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
78 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
79 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
81 if (bo->tbo.pin_count > 0)
82 amdgpu_bo_subtract_pin_size(bo);
86 if (bo->tbo.base.import_attach)
87 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
88 drm_gem_object_release(&bo->tbo.base);
89 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
90 if (!list_empty(&bo->shadow_list)) {
91 mutex_lock(&adev->shadow_list_lock);
92 list_del_init(&bo->shadow_list);
93 mutex_unlock(&adev->shadow_list_lock);
95 amdgpu_bo_unref(&bo->parent);
102 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
103 * @bo: buffer object to be checked
105 * Uses destroy function associated with the object to determine if this is
109 * true if the object belongs to &amdgpu_bo, false if not.
111 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
113 if (bo->destroy == &amdgpu_bo_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 visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
139 places[c].mem_type = TTM_PL_VRAM;
142 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
143 places[c].lpfn = visible_pfn;
145 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
147 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
148 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
152 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
155 places[c].mem_type = 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;
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.mem);
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.mem, &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())
512 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
513 struct amdgpu_bo_param *bp,
514 struct amdgpu_bo **bo_ptr)
516 struct ttm_operation_ctx ctx = {
517 .interruptible = (bp->type != ttm_bo_type_kernel),
518 .no_wait_gpu = bp->no_wait_gpu,
519 /* We opt to avoid OOM on system pages allocations */
520 .gfp_retry_mayfail = true,
521 .allow_res_evict = bp->type != ttm_bo_type_kernel,
524 struct amdgpu_bo *bo;
525 unsigned long page_align, size = bp->size;
528 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
529 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
530 /* GWS and OA don't need any alignment. */
531 page_align = bp->byte_align;
533 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
534 /* Both size and alignment must be a multiple of 4. */
535 page_align = ALIGN(bp->byte_align, 4);
536 size = ALIGN(size, 4) << PAGE_SHIFT;
538 /* Memory should be aligned at least to a page size. */
539 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
540 size = ALIGN(size, PAGE_SIZE);
543 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
548 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
551 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
552 INIT_LIST_HEAD(&bo->shadow_list);
554 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
556 bo->allowed_domains = bo->preferred_domains;
557 if (bp->type != ttm_bo_type_kernel &&
558 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
559 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
561 bo->flags = bp->flags;
563 if (!amdgpu_bo_support_uswc(bo->flags))
564 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
566 bo->tbo.bdev = &adev->mman.bdev;
567 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
568 AMDGPU_GEM_DOMAIN_GDS))
569 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
571 amdgpu_bo_placement_from_domain(bo, bp->domain);
572 if (bp->type == ttm_bo_type_kernel)
573 bo->tbo.priority = 1;
575 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
576 &bo->placement, page_align, &ctx, NULL,
577 bp->resv, &amdgpu_bo_destroy);
578 if (unlikely(r != 0))
581 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
582 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
583 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
584 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
587 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
589 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
590 bo->tbo.mem.mem_type == TTM_PL_VRAM) {
591 struct dma_fence *fence;
593 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
597 amdgpu_bo_fence(bo, fence, false);
598 dma_fence_put(bo->tbo.moving);
599 bo->tbo.moving = dma_fence_get(fence);
600 dma_fence_put(fence);
603 amdgpu_bo_unreserve(bo);
606 trace_amdgpu_bo_create(bo);
608 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
609 if (bp->type == ttm_bo_type_device)
610 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
616 dma_resv_unlock(bo->tbo.base.resv);
617 amdgpu_bo_unref(&bo);
621 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
623 struct amdgpu_bo *bo)
625 struct amdgpu_bo_param bp;
631 memset(&bp, 0, sizeof(bp));
633 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
634 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
635 AMDGPU_GEM_CREATE_SHADOW;
636 bp.type = ttm_bo_type_kernel;
637 bp.resv = bo->tbo.base.resv;
639 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
641 bo->shadow->parent = amdgpu_bo_ref(bo);
642 mutex_lock(&adev->shadow_list_lock);
643 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
644 mutex_unlock(&adev->shadow_list_lock);
651 * amdgpu_bo_create - create an &amdgpu_bo buffer object
652 * @adev: amdgpu device object
653 * @bp: parameters to be used for the buffer object
654 * @bo_ptr: pointer to the buffer object pointer
656 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
658 * Shadow object is used to backup the original buffer object, and is always
662 * 0 for success or a negative error code on failure.
664 int amdgpu_bo_create(struct amdgpu_device *adev,
665 struct amdgpu_bo_param *bp,
666 struct amdgpu_bo **bo_ptr)
668 u64 flags = bp->flags;
671 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
672 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
676 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
678 WARN_ON(dma_resv_lock((*bo_ptr)->tbo.base.resv,
681 r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
684 dma_resv_unlock((*bo_ptr)->tbo.base.resv);
687 amdgpu_bo_unref(bo_ptr);
694 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
695 * @bo: pointer to the buffer object
697 * Sets placement according to domain; and changes placement and caching
698 * policy of the buffer object according to the placement.
699 * This is used for validating shadow bos. It calls ttm_bo_validate() to
700 * make sure the buffer is resident where it needs to be.
703 * 0 for success or a negative error code on failure.
705 int amdgpu_bo_validate(struct amdgpu_bo *bo)
707 struct ttm_operation_ctx ctx = { false, false };
711 if (bo->tbo.pin_count)
714 domain = bo->preferred_domains;
717 amdgpu_bo_placement_from_domain(bo, domain);
718 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
719 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
720 domain = bo->allowed_domains;
728 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
730 * @shadow: &amdgpu_bo shadow to be restored
731 * @fence: dma_fence associated with the operation
733 * Copies a buffer object's shadow content back to the object.
734 * This is used for recovering a buffer from its shadow in case of a gpu
735 * reset where vram context may be lost.
738 * 0 for success or a negative error code on failure.
740 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
743 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
744 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
745 uint64_t shadow_addr, parent_addr;
747 shadow_addr = amdgpu_bo_gpu_offset(shadow);
748 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
750 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
751 amdgpu_bo_size(shadow), NULL, fence,
756 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
757 * @bo: &amdgpu_bo buffer object to be mapped
758 * @ptr: kernel virtual address to be returned
760 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
761 * amdgpu_bo_kptr() to get the kernel virtual address.
764 * 0 for success or a negative error code on failure.
766 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
771 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
774 kptr = amdgpu_bo_kptr(bo);
781 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false, false,
782 MAX_SCHEDULE_TIMEOUT);
786 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.mem.num_pages, &bo->kmap);
791 *ptr = amdgpu_bo_kptr(bo);
797 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
798 * @bo: &amdgpu_bo buffer object
800 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
803 * the virtual address of a buffer object area.
805 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
809 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
813 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
814 * @bo: &amdgpu_bo buffer object to be unmapped
816 * Unmaps a kernel map set up by amdgpu_bo_kmap().
818 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
821 ttm_bo_kunmap(&bo->kmap);
825 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
826 * @bo: &amdgpu_bo buffer object
828 * References the contained &ttm_buffer_object.
831 * a refcounted pointer to the &amdgpu_bo buffer object.
833 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
838 ttm_bo_get(&bo->tbo);
843 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
844 * @bo: &amdgpu_bo buffer object
846 * Unreferences the contained &ttm_buffer_object and clear the pointer
848 void amdgpu_bo_unref(struct amdgpu_bo **bo)
850 struct ttm_buffer_object *tbo;
861 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
862 * @bo: &amdgpu_bo buffer object to be pinned
863 * @domain: domain to be pinned to
864 * @min_offset: the start of requested address range
865 * @max_offset: the end of requested address range
867 * Pins the buffer object according to requested domain and address range. If
868 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
869 * pin_count and pin_size accordingly.
871 * Pinning means to lock pages in memory along with keeping them at a fixed
872 * offset. It is required when a buffer can not be moved, for example, when
873 * a display buffer is being scanned out.
875 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
876 * where to pin a buffer if there are specific restrictions on where a buffer
880 * 0 for success or a negative error code on failure.
882 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
883 u64 min_offset, u64 max_offset)
885 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
886 struct ttm_operation_ctx ctx = { false, false };
889 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
892 if (WARN_ON_ONCE(min_offset > max_offset))
895 /* A shared bo cannot be migrated to VRAM */
896 if (bo->prime_shared_count || bo->tbo.base.import_attach) {
897 if (domain & AMDGPU_GEM_DOMAIN_GTT)
898 domain = AMDGPU_GEM_DOMAIN_GTT;
903 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
904 * See function amdgpu_display_supported_domains()
906 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
908 if (bo->tbo.pin_count) {
909 uint32_t mem_type = bo->tbo.mem.mem_type;
910 uint32_t mem_flags = bo->tbo.mem.placement;
912 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
915 if ((mem_type == TTM_PL_VRAM) &&
916 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
917 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
920 ttm_bo_pin(&bo->tbo);
922 if (max_offset != 0) {
923 u64 domain_start = amdgpu_ttm_domain_start(adev,
925 WARN_ON_ONCE(max_offset <
926 (amdgpu_bo_gpu_offset(bo) - domain_start));
932 if (bo->tbo.base.import_attach)
933 dma_buf_pin(bo->tbo.base.import_attach);
935 /* force to pin into visible video ram */
936 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
937 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
938 amdgpu_bo_placement_from_domain(bo, domain);
939 for (i = 0; i < bo->placement.num_placement; i++) {
942 fpfn = min_offset >> PAGE_SHIFT;
943 lpfn = max_offset >> PAGE_SHIFT;
945 if (fpfn > bo->placements[i].fpfn)
946 bo->placements[i].fpfn = fpfn;
947 if (!bo->placements[i].lpfn ||
948 (lpfn && lpfn < bo->placements[i].lpfn))
949 bo->placements[i].lpfn = lpfn;
952 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
954 dev_err(adev->dev, "%p pin failed\n", bo);
958 ttm_bo_pin(&bo->tbo);
960 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
961 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
962 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
963 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
964 &adev->visible_pin_size);
965 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
966 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
974 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
975 * @bo: &amdgpu_bo buffer object to be pinned
976 * @domain: domain to be pinned to
978 * A simple wrapper to amdgpu_bo_pin_restricted().
979 * Provides a simpler API for buffers that do not have any strict restrictions
980 * on where a buffer must be located.
983 * 0 for success or a negative error code on failure.
985 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
987 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
988 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
992 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
993 * @bo: &amdgpu_bo buffer object to be unpinned
995 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
996 * Changes placement and pin size accordingly.
999 * 0 for success or a negative error code on failure.
1001 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
1003 ttm_bo_unpin(&bo->tbo);
1004 if (bo->tbo.pin_count)
1007 amdgpu_bo_subtract_pin_size(bo);
1009 if (bo->tbo.base.import_attach)
1010 dma_buf_unpin(bo->tbo.base.import_attach);
1014 * amdgpu_bo_evict_vram - evict VRAM buffers
1015 * @adev: amdgpu device object
1017 * Evicts all VRAM buffers on the lru list of the memory type.
1018 * Mainly used for evicting vram at suspend time.
1021 * 0 for success or a negative error code on failure.
1023 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1025 struct ttm_resource_manager *man;
1027 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
1028 #ifndef CONFIG_HIBERNATION
1029 if (adev->flags & AMD_IS_APU) {
1030 /* Useless to evict on IGP chips */
1035 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1036 return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
1039 static const char *amdgpu_vram_names[] = {
1054 * amdgpu_bo_init - initialize memory manager
1055 * @adev: amdgpu device object
1057 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1060 * 0 for success or a negative error code on failure.
1062 int amdgpu_bo_init(struct amdgpu_device *adev)
1064 /* On A+A platform, VRAM can be mapped as WB */
1065 if (!adev->gmc.xgmi.connected_to_cpu) {
1066 /* reserve PAT memory space to WC for VRAM */
1067 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1068 adev->gmc.aper_size);
1070 /* Add an MTRR for the VRAM */
1071 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1072 adev->gmc.aper_size);
1075 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1076 adev->gmc.mc_vram_size >> 20,
1077 (unsigned long long)adev->gmc.aper_size >> 20);
1078 DRM_INFO("RAM width %dbits %s\n",
1079 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1080 return amdgpu_ttm_init(adev);
1084 * amdgpu_bo_fini - tear down memory manager
1085 * @adev: amdgpu device object
1087 * Reverses amdgpu_bo_init() to tear down memory manager.
1089 void amdgpu_bo_fini(struct amdgpu_device *adev)
1091 amdgpu_ttm_fini(adev);
1092 if (!adev->gmc.xgmi.connected_to_cpu) {
1093 arch_phys_wc_del(adev->gmc.vram_mtrr);
1094 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1099 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1100 * @bo: &amdgpu_bo buffer object
1101 * @vma: vma as input from the fbdev mmap method
1103 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1106 * 0 for success or a negative error code on failure.
1108 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1109 struct vm_area_struct *vma)
1111 if (vma->vm_pgoff != 0)
1114 return ttm_bo_mmap_obj(vma, &bo->tbo);
1118 * amdgpu_bo_set_tiling_flags - set tiling flags
1119 * @bo: &amdgpu_bo buffer object
1120 * @tiling_flags: new flags
1122 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1123 * kernel driver to set the tiling flags on a buffer.
1126 * 0 for success or a negative error code on failure.
1128 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1130 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1132 if (adev->family <= AMDGPU_FAMILY_CZ &&
1133 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1136 bo->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 dma_resv_assert_held(bo->tbo.base.resv);
1153 *tiling_flags = bo->tiling_flags;
1157 * amdgpu_bo_set_metadata - set metadata
1158 * @bo: &amdgpu_bo buffer object
1159 * @metadata: new metadata
1160 * @metadata_size: size of the new metadata
1161 * @flags: flags of the new metadata
1163 * Sets buffer object's metadata, its size and flags.
1164 * Used via GEM ioctl.
1167 * 0 for success or a negative error code on failure.
1169 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1170 uint32_t metadata_size, uint64_t flags)
1174 if (!metadata_size) {
1175 if (bo->metadata_size) {
1176 kfree(bo->metadata);
1177 bo->metadata = NULL;
1178 bo->metadata_size = 0;
1183 if (metadata == NULL)
1186 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1190 kfree(bo->metadata);
1191 bo->metadata_flags = flags;
1192 bo->metadata = buffer;
1193 bo->metadata_size = metadata_size;
1199 * amdgpu_bo_get_metadata - get metadata
1200 * @bo: &amdgpu_bo buffer object
1201 * @buffer: returned metadata
1202 * @buffer_size: size of the buffer
1203 * @metadata_size: size of the returned metadata
1204 * @flags: flags of the returned metadata
1206 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1207 * less than metadata_size.
1208 * Used via GEM ioctl.
1211 * 0 for success or a negative error code on failure.
1213 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1214 size_t buffer_size, uint32_t *metadata_size,
1217 if (!buffer && !metadata_size)
1221 if (buffer_size < bo->metadata_size)
1224 if (bo->metadata_size)
1225 memcpy(buffer, bo->metadata, bo->metadata_size);
1229 *metadata_size = bo->metadata_size;
1231 *flags = bo->metadata_flags;
1237 * amdgpu_bo_move_notify - notification about a memory move
1238 * @bo: pointer to a buffer object
1239 * @evict: if this move is evicting the buffer from the graphics address space
1240 * @new_mem: new information of the bufer object
1242 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1244 * TTM driver callback which is called when ttm moves a buffer.
1246 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1248 struct ttm_resource *new_mem)
1250 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1251 struct amdgpu_bo *abo;
1252 struct ttm_resource *old_mem = &bo->mem;
1254 if (!amdgpu_bo_is_amdgpu_bo(bo))
1257 abo = ttm_to_amdgpu_bo(bo);
1258 amdgpu_vm_bo_invalidate(adev, abo, evict);
1260 amdgpu_bo_kunmap(abo);
1262 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1263 bo->mem.mem_type != TTM_PL_SYSTEM)
1264 dma_buf_move_notify(abo->tbo.base.dma_buf);
1266 /* remember the eviction */
1268 atomic64_inc(&adev->num_evictions);
1270 /* update statistics */
1274 /* move_notify is called before move happens */
1275 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1279 * amdgpu_bo_release_notify - notification about a BO being released
1280 * @bo: pointer to a buffer object
1282 * Wipes VRAM buffers whose contents should not be leaked before the
1283 * memory is released.
1285 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1287 struct dma_fence *fence = NULL;
1288 struct amdgpu_bo *abo;
1291 if (!amdgpu_bo_is_amdgpu_bo(bo))
1294 abo = ttm_to_amdgpu_bo(bo);
1297 amdgpu_amdkfd_unreserve_memory_limit(abo);
1299 /* We only remove the fence if the resv has individualized. */
1300 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1301 && bo->base.resv != &bo->base._resv);
1302 if (bo->base.resv == &bo->base._resv)
1303 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1305 if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
1306 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1309 dma_resv_lock(bo->base.resv, NULL);
1311 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1313 amdgpu_bo_fence(abo, fence, false);
1314 dma_fence_put(fence);
1317 dma_resv_unlock(bo->base.resv);
1321 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1322 * @bo: pointer to a buffer object
1324 * Notifies the driver we are taking a fault on this BO and have reserved it,
1325 * also performs bookkeeping.
1326 * TTM driver callback for dealing with vm faults.
1329 * 0 for success or a negative error code on failure.
1331 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1333 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1334 struct ttm_operation_ctx ctx = { false, false };
1335 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1336 unsigned long offset, size;
1339 /* Remember that this BO was accessed by the CPU */
1340 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1342 if (bo->mem.mem_type != TTM_PL_VRAM)
1345 size = bo->mem.num_pages << PAGE_SHIFT;
1346 offset = bo->mem.start << PAGE_SHIFT;
1347 if ((offset + size) <= adev->gmc.visible_vram_size)
1350 /* Can't move a pinned BO to visible VRAM */
1351 if (abo->tbo.pin_count > 0)
1352 return VM_FAULT_SIGBUS;
1354 /* hurrah the memory is not visible ! */
1355 atomic64_inc(&adev->num_vram_cpu_page_faults);
1356 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1357 AMDGPU_GEM_DOMAIN_GTT);
1359 /* Avoid costly evictions; only set GTT as a busy placement */
1360 abo->placement.num_busy_placement = 1;
1361 abo->placement.busy_placement = &abo->placements[1];
1363 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1364 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1365 return VM_FAULT_NOPAGE;
1366 else if (unlikely(r))
1367 return VM_FAULT_SIGBUS;
1369 offset = bo->mem.start << PAGE_SHIFT;
1370 /* this should never happen */
1371 if (bo->mem.mem_type == TTM_PL_VRAM &&
1372 (offset + size) > adev->gmc.visible_vram_size)
1373 return VM_FAULT_SIGBUS;
1375 ttm_bo_move_to_lru_tail_unlocked(bo);
1380 * amdgpu_bo_fence - add fence to buffer object
1382 * @bo: buffer object in question
1383 * @fence: fence to add
1384 * @shared: true if fence should be added shared
1387 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1390 struct dma_resv *resv = bo->tbo.base.resv;
1393 dma_resv_add_shared_fence(resv, fence);
1395 dma_resv_add_excl_fence(resv, fence);
1399 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1401 * @adev: amdgpu device pointer
1402 * @resv: reservation object to sync to
1403 * @sync_mode: synchronization mode
1404 * @owner: fence owner
1405 * @intr: Whether the wait is interruptible
1407 * Extract the fences from the reservation object and waits for them to finish.
1410 * 0 on success, errno otherwise.
1412 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1413 enum amdgpu_sync_mode sync_mode, void *owner,
1416 struct amdgpu_sync sync;
1419 amdgpu_sync_create(&sync);
1420 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1421 r = amdgpu_sync_wait(&sync, intr);
1422 amdgpu_sync_free(&sync);
1427 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1428 * @bo: buffer object to wait for
1429 * @owner: fence owner
1430 * @intr: Whether the wait is interruptible
1432 * Wrapper to wait for fences in a BO.
1434 * 0 on success, errno otherwise.
1436 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1438 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1440 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1441 AMDGPU_SYNC_NE_OWNER, owner, intr);
1445 * amdgpu_bo_gpu_offset - return GPU offset of bo
1446 * @bo: amdgpu object for which we query the offset
1448 * Note: object should either be pinned or reserved when calling this
1449 * function, it might be useful to add check for this for debugging.
1452 * current GPU offset of the object.
1454 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1456 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1457 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1458 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1459 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1460 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1461 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1463 return amdgpu_bo_gpu_offset_no_check(bo);
1467 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1468 * @bo: amdgpu object for which we query the offset
1471 * current GPU offset of the object without raising warnings.
1473 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1475 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1478 offset = (bo->tbo.mem.start << PAGE_SHIFT) +
1479 amdgpu_ttm_domain_start(adev, bo->tbo.mem.mem_type);
1481 return amdgpu_gmc_sign_extend(offset);
1485 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1486 * @adev: amdgpu device object
1487 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1490 * Which of the allowed domains is preferred for pinning the BO for scanout.
1492 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1495 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1496 domain = AMDGPU_GEM_DOMAIN_VRAM;
1497 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1498 domain = AMDGPU_GEM_DOMAIN_GTT;
1503 #if defined(CONFIG_DEBUG_FS)
1504 #define amdgpu_bo_print_flag(m, bo, flag) \
1506 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1507 seq_printf((m), " " #flag); \
1512 * amdgpu_bo_print_info - print BO info in debugfs file
1514 * @id: Index or Id of the BO
1515 * @bo: Requested BO for printing info
1518 * Print BO information in debugfs file
1521 * Size of the BO in bytes.
1523 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1525 struct dma_buf_attachment *attachment;
1526 struct dma_buf *dma_buf;
1527 unsigned int domain;
1528 const char *placement;
1529 unsigned int pin_count;
1532 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
1534 case AMDGPU_GEM_DOMAIN_VRAM:
1537 case AMDGPU_GEM_DOMAIN_GTT:
1540 case AMDGPU_GEM_DOMAIN_CPU:
1546 size = amdgpu_bo_size(bo);
1547 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1548 id, size, placement);
1550 pin_count = READ_ONCE(bo->tbo.pin_count);
1552 seq_printf(m, " pin count %d", pin_count);
1554 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1555 attachment = READ_ONCE(bo->tbo.base.import_attach);
1558 seq_printf(m, " imported from %p", dma_buf);
1560 seq_printf(m, " exported as %p", dma_buf);
1562 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1563 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1564 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1565 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1566 amdgpu_bo_print_flag(m, bo, SHADOW);
1567 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1568 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1569 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);