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_device *adev = amdgpu_ttm_adev(tbo->bdev);
58 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
59 struct amdgpu_bo_user *ubo;
63 if (bo->tbo.base.import_attach)
64 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
65 drm_gem_object_release(&bo->tbo.base);
66 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
67 if (!list_empty(&bo->shadow_list)) {
68 mutex_lock(&adev->shadow_list_lock);
69 list_del_init(&bo->shadow_list);
70 mutex_unlock(&adev->shadow_list_lock);
72 amdgpu_bo_unref(&bo->parent);
74 if (bo->tbo.type != ttm_bo_type_kernel) {
75 ubo = to_amdgpu_bo_user(bo);
83 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
84 * @bo: buffer object to be checked
86 * Uses destroy function associated with the object to determine if this is
90 * true if the object belongs to &amdgpu_bo, false if not.
92 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
94 if (bo->destroy == &amdgpu_bo_destroy)
100 * amdgpu_bo_placement_from_domain - set buffer's placement
101 * @abo: &amdgpu_bo buffer object whose placement is to be set
102 * @domain: requested domain
104 * Sets buffer's placement according to requested domain and the buffer's
107 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
109 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
110 struct ttm_placement *placement = &abo->placement;
111 struct ttm_place *places = abo->placements;
112 u64 flags = abo->flags;
115 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
116 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
120 places[c].mem_type = TTM_PL_VRAM;
123 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
124 places[c].lpfn = visible_pfn;
126 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
128 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
129 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
133 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
137 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
138 AMDGPU_PL_PREEMPT : TTM_PL_TT;
143 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
146 places[c].mem_type = TTM_PL_SYSTEM;
151 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
154 places[c].mem_type = AMDGPU_PL_GDS;
159 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
162 places[c].mem_type = AMDGPU_PL_GWS;
167 if (domain & AMDGPU_GEM_DOMAIN_OA) {
170 places[c].mem_type = AMDGPU_PL_OA;
178 places[c].mem_type = TTM_PL_SYSTEM;
183 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
185 placement->num_placement = c;
186 placement->placement = places;
188 placement->num_busy_placement = c;
189 placement->busy_placement = places;
193 * amdgpu_bo_create_reserved - create reserved BO for kernel use
195 * @adev: amdgpu device object
196 * @size: size for the new BO
197 * @align: alignment for the new BO
198 * @domain: where to place it
199 * @bo_ptr: used to initialize BOs in structures
200 * @gpu_addr: GPU addr of the pinned BO
201 * @cpu_addr: optional CPU address mapping
203 * Allocates and pins a BO for kernel internal use, and returns it still
206 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
209 * 0 on success, negative error code otherwise.
211 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
212 unsigned long size, int align,
213 u32 domain, struct amdgpu_bo **bo_ptr,
214 u64 *gpu_addr, void **cpu_addr)
216 struct amdgpu_bo_param bp;
221 amdgpu_bo_unref(bo_ptr);
225 memset(&bp, 0, sizeof(bp));
227 bp.byte_align = align;
229 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
230 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
231 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
232 bp.type = ttm_bo_type_kernel;
234 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
237 r = amdgpu_bo_create(adev, &bp, bo_ptr);
239 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
246 r = amdgpu_bo_reserve(*bo_ptr, false);
248 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
252 r = amdgpu_bo_pin(*bo_ptr, domain);
254 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
255 goto error_unreserve;
258 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
260 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
265 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
268 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
270 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
278 amdgpu_bo_unpin(*bo_ptr);
280 amdgpu_bo_unreserve(*bo_ptr);
284 amdgpu_bo_unref(bo_ptr);
290 * amdgpu_bo_create_kernel - create BO for kernel use
292 * @adev: amdgpu device object
293 * @size: size for the new BO
294 * @align: alignment for the new BO
295 * @domain: where to place it
296 * @bo_ptr: used to initialize BOs in structures
297 * @gpu_addr: GPU addr of the pinned BO
298 * @cpu_addr: optional CPU address mapping
300 * Allocates and pins a BO for kernel internal use.
302 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
305 * 0 on success, negative error code otherwise.
307 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
308 unsigned long size, int align,
309 u32 domain, struct amdgpu_bo **bo_ptr,
310 u64 *gpu_addr, void **cpu_addr)
314 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
321 amdgpu_bo_unreserve(*bo_ptr);
327 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
329 * @adev: amdgpu device object
330 * @offset: offset of the BO
331 * @size: size of the BO
332 * @domain: where to place it
333 * @bo_ptr: used to initialize BOs in structures
334 * @cpu_addr: optional CPU address mapping
336 * Creates a kernel BO at a specific offset in the address space of the domain.
339 * 0 on success, negative error code otherwise.
341 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
342 uint64_t offset, uint64_t size, uint32_t domain,
343 struct amdgpu_bo **bo_ptr, void **cpu_addr)
345 struct ttm_operation_ctx ctx = { false, false };
350 size = ALIGN(size, PAGE_SIZE);
352 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
357 if ((*bo_ptr) == NULL)
361 * Remove the original mem node and create a new one at the request
365 amdgpu_bo_kunmap(*bo_ptr);
367 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.mem);
369 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
370 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
371 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
373 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
374 &(*bo_ptr)->tbo.mem, &ctx);
379 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
384 amdgpu_bo_unreserve(*bo_ptr);
388 amdgpu_bo_unreserve(*bo_ptr);
389 amdgpu_bo_unref(bo_ptr);
394 * amdgpu_bo_free_kernel - free BO for kernel use
396 * @bo: amdgpu BO to free
397 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
398 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
400 * unmaps and unpin a BO for kernel internal use.
402 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
408 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
410 amdgpu_bo_kunmap(*bo);
412 amdgpu_bo_unpin(*bo);
413 amdgpu_bo_unreserve(*bo);
424 /* Validate bo size is bit bigger then the request domain */
425 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
426 unsigned long size, u32 domain)
428 struct ttm_resource_manager *man = NULL;
431 * If GTT is part of requested domains the check must succeed to
432 * allow fall back to GTT
434 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
435 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
437 if (size < (man->size << PAGE_SHIFT))
443 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
444 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
446 if (size < (man->size << PAGE_SHIFT))
453 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
457 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
458 man->size << PAGE_SHIFT);
462 bool amdgpu_bo_support_uswc(u64 bo_flags)
466 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
467 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
470 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
471 /* Don't try to enable write-combining when it can't work, or things
473 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
476 #ifndef CONFIG_COMPILE_TEST
477 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
478 thanks to write-combining
481 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
482 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
483 "better performance thanks to write-combining\n");
486 /* For architectures that don't support WC memory,
487 * mask out the WC flag from the BO
489 if (!drm_arch_can_wc_memory())
497 * amdgpu_bo_create - create an &amdgpu_bo buffer object
498 * @adev: amdgpu device object
499 * @bp: parameters to be used for the buffer object
500 * @bo_ptr: pointer to the buffer object pointer
502 * Creates an &amdgpu_bo buffer object.
505 * 0 for success or a negative error code on failure.
507 int amdgpu_bo_create(struct amdgpu_device *adev,
508 struct amdgpu_bo_param *bp,
509 struct amdgpu_bo **bo_ptr)
511 struct ttm_operation_ctx ctx = {
512 .interruptible = (bp->type != ttm_bo_type_kernel),
513 .no_wait_gpu = bp->no_wait_gpu,
514 /* We opt to avoid OOM on system pages allocations */
515 .gfp_retry_mayfail = true,
516 .allow_res_evict = bp->type != ttm_bo_type_kernel,
519 struct amdgpu_bo *bo;
520 unsigned long page_align, size = bp->size;
523 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
524 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
525 /* GWS and OA don't need any alignment. */
526 page_align = bp->byte_align;
528 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
529 /* Both size and alignment must be a multiple of 4. */
530 page_align = ALIGN(bp->byte_align, 4);
531 size = ALIGN(size, 4) << PAGE_SHIFT;
533 /* Memory should be aligned at least to a page size. */
534 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
535 size = ALIGN(size, PAGE_SIZE);
538 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
541 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
544 bo = kzalloc(bp->bo_ptr_size, GFP_KERNEL);
547 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
548 INIT_LIST_HEAD(&bo->shadow_list);
550 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
552 bo->allowed_domains = bo->preferred_domains;
553 if (bp->type != ttm_bo_type_kernel &&
554 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
555 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
557 bo->flags = bp->flags;
559 if (!amdgpu_bo_support_uswc(bo->flags))
560 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
562 bo->tbo.bdev = &adev->mman.bdev;
563 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
564 AMDGPU_GEM_DOMAIN_GDS))
565 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
567 amdgpu_bo_placement_from_domain(bo, bp->domain);
568 if (bp->type == ttm_bo_type_kernel)
569 bo->tbo.priority = 1;
571 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
572 &bo->placement, page_align, &ctx, NULL,
573 bp->resv, &amdgpu_bo_destroy);
574 if (unlikely(r != 0))
577 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
578 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
579 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
580 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
583 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
585 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
586 bo->tbo.mem.mem_type == TTM_PL_VRAM) {
587 struct dma_fence *fence;
589 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
593 amdgpu_bo_fence(bo, fence, false);
594 dma_fence_put(bo->tbo.moving);
595 bo->tbo.moving = dma_fence_get(fence);
596 dma_fence_put(fence);
599 amdgpu_bo_unreserve(bo);
602 trace_amdgpu_bo_create(bo);
604 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
605 if (bp->type == ttm_bo_type_device)
606 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
612 dma_resv_unlock(bo->tbo.base.resv);
613 amdgpu_bo_unref(&bo);
617 int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
619 struct amdgpu_bo *bo)
621 struct amdgpu_bo_param bp;
627 memset(&bp, 0, sizeof(bp));
629 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
630 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
631 bp.type = ttm_bo_type_kernel;
632 bp.resv = bo->tbo.base.resv;
633 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
635 r = amdgpu_bo_create(adev, &bp, &bo->shadow);
637 bo->shadow->parent = amdgpu_bo_ref(bo);
638 mutex_lock(&adev->shadow_list_lock);
639 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
640 mutex_unlock(&adev->shadow_list_lock);
647 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
648 * @adev: amdgpu device object
649 * @bp: parameters to be used for the buffer object
650 * @ubo_ptr: pointer to the buffer object pointer
652 * Create a BO to be used by user application;
655 * 0 for success or a negative error code on failure.
658 int amdgpu_bo_create_user(struct amdgpu_device *adev,
659 struct amdgpu_bo_param *bp,
660 struct amdgpu_bo_user **ubo_ptr)
662 struct amdgpu_bo *bo_ptr;
665 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
666 r = amdgpu_bo_create(adev, bp, &bo_ptr);
670 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
674 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
675 * @bo: pointer to the buffer object
677 * Sets placement according to domain; and changes placement and caching
678 * policy of the buffer object according to the placement.
679 * This is used for validating shadow bos. It calls ttm_bo_validate() to
680 * make sure the buffer is resident where it needs to be.
683 * 0 for success or a negative error code on failure.
685 int amdgpu_bo_validate(struct amdgpu_bo *bo)
687 struct ttm_operation_ctx ctx = { false, false };
691 if (bo->tbo.pin_count)
694 domain = bo->preferred_domains;
697 amdgpu_bo_placement_from_domain(bo, domain);
698 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
699 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
700 domain = bo->allowed_domains;
708 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
710 * @shadow: &amdgpu_bo shadow to be restored
711 * @fence: dma_fence associated with the operation
713 * Copies a buffer object's shadow content back to the object.
714 * This is used for recovering a buffer from its shadow in case of a gpu
715 * reset where vram context may be lost.
718 * 0 for success or a negative error code on failure.
720 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
723 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
724 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
725 uint64_t shadow_addr, parent_addr;
727 shadow_addr = amdgpu_bo_gpu_offset(shadow);
728 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
730 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
731 amdgpu_bo_size(shadow), NULL, fence,
736 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
737 * @bo: &amdgpu_bo buffer object to be mapped
738 * @ptr: kernel virtual address to be returned
740 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
741 * amdgpu_bo_kptr() to get the kernel virtual address.
744 * 0 for success or a negative error code on failure.
746 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
751 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
754 kptr = amdgpu_bo_kptr(bo);
761 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false, false,
762 MAX_SCHEDULE_TIMEOUT);
766 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.mem.num_pages, &bo->kmap);
771 *ptr = amdgpu_bo_kptr(bo);
777 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
778 * @bo: &amdgpu_bo buffer object
780 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
783 * the virtual address of a buffer object area.
785 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
789 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
793 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
794 * @bo: &amdgpu_bo buffer object to be unmapped
796 * Unmaps a kernel map set up by amdgpu_bo_kmap().
798 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
801 ttm_bo_kunmap(&bo->kmap);
805 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
806 * @bo: &amdgpu_bo buffer object
808 * References the contained &ttm_buffer_object.
811 * a refcounted pointer to the &amdgpu_bo buffer object.
813 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
818 ttm_bo_get(&bo->tbo);
823 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
824 * @bo: &amdgpu_bo buffer object
826 * Unreferences the contained &ttm_buffer_object and clear the pointer
828 void amdgpu_bo_unref(struct amdgpu_bo **bo)
830 struct ttm_buffer_object *tbo;
841 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
842 * @bo: &amdgpu_bo buffer object to be pinned
843 * @domain: domain to be pinned to
844 * @min_offset: the start of requested address range
845 * @max_offset: the end of requested address range
847 * Pins the buffer object according to requested domain and address range. If
848 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
849 * pin_count and pin_size accordingly.
851 * Pinning means to lock pages in memory along with keeping them at a fixed
852 * offset. It is required when a buffer can not be moved, for example, when
853 * a display buffer is being scanned out.
855 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
856 * where to pin a buffer if there are specific restrictions on where a buffer
860 * 0 for success or a negative error code on failure.
862 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
863 u64 min_offset, u64 max_offset)
865 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
866 struct ttm_operation_ctx ctx = { false, false };
869 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
872 if (WARN_ON_ONCE(min_offset > max_offset))
875 /* A shared bo cannot be migrated to VRAM */
876 if (bo->prime_shared_count || bo->tbo.base.import_attach) {
877 if (domain & AMDGPU_GEM_DOMAIN_GTT)
878 domain = AMDGPU_GEM_DOMAIN_GTT;
883 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
884 * See function amdgpu_display_supported_domains()
886 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
888 if (bo->tbo.pin_count) {
889 uint32_t mem_type = bo->tbo.mem.mem_type;
890 uint32_t mem_flags = bo->tbo.mem.placement;
892 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
895 if ((mem_type == TTM_PL_VRAM) &&
896 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
897 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
900 ttm_bo_pin(&bo->tbo);
902 if (max_offset != 0) {
903 u64 domain_start = amdgpu_ttm_domain_start(adev,
905 WARN_ON_ONCE(max_offset <
906 (amdgpu_bo_gpu_offset(bo) - domain_start));
912 if (bo->tbo.base.import_attach)
913 dma_buf_pin(bo->tbo.base.import_attach);
915 /* force to pin into visible video ram */
916 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
917 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
918 amdgpu_bo_placement_from_domain(bo, domain);
919 for (i = 0; i < bo->placement.num_placement; i++) {
922 fpfn = min_offset >> PAGE_SHIFT;
923 lpfn = max_offset >> PAGE_SHIFT;
925 if (fpfn > bo->placements[i].fpfn)
926 bo->placements[i].fpfn = fpfn;
927 if (!bo->placements[i].lpfn ||
928 (lpfn && lpfn < bo->placements[i].lpfn))
929 bo->placements[i].lpfn = lpfn;
932 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
934 dev_err(adev->dev, "%p pin failed\n", bo);
938 ttm_bo_pin(&bo->tbo);
940 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
941 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
942 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
943 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
944 &adev->visible_pin_size);
945 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
946 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
954 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
955 * @bo: &amdgpu_bo buffer object to be pinned
956 * @domain: domain to be pinned to
958 * A simple wrapper to amdgpu_bo_pin_restricted().
959 * Provides a simpler API for buffers that do not have any strict restrictions
960 * on where a buffer must be located.
963 * 0 for success or a negative error code on failure.
965 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
967 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
968 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
972 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
973 * @bo: &amdgpu_bo buffer object to be unpinned
975 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
976 * Changes placement and pin size accordingly.
979 * 0 for success or a negative error code on failure.
981 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
983 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
985 ttm_bo_unpin(&bo->tbo);
986 if (bo->tbo.pin_count)
989 if (bo->tbo.base.import_attach)
990 dma_buf_unpin(bo->tbo.base.import_attach);
992 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
993 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
994 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
995 &adev->visible_pin_size);
996 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
997 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1002 * amdgpu_bo_evict_vram - evict VRAM buffers
1003 * @adev: amdgpu device object
1005 * Evicts all VRAM buffers on the lru list of the memory type.
1006 * Mainly used for evicting vram at suspend time.
1009 * 0 for success or a negative error code on failure.
1011 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1013 struct ttm_resource_manager *man;
1015 if (adev->in_s3 && (adev->flags & AMD_IS_APU)) {
1016 /* No need to evict vram on APUs for suspend to ram */
1020 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1021 return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
1024 static const char *amdgpu_vram_names[] = {
1039 * amdgpu_bo_init - initialize memory manager
1040 * @adev: amdgpu device object
1042 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1045 * 0 for success or a negative error code on failure.
1047 int amdgpu_bo_init(struct amdgpu_device *adev)
1049 /* On A+A platform, VRAM can be mapped as WB */
1050 if (!adev->gmc.xgmi.connected_to_cpu) {
1051 /* reserve PAT memory space to WC for VRAM */
1052 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1053 adev->gmc.aper_size);
1055 /* Add an MTRR for the VRAM */
1056 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1057 adev->gmc.aper_size);
1060 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1061 adev->gmc.mc_vram_size >> 20,
1062 (unsigned long long)adev->gmc.aper_size >> 20);
1063 DRM_INFO("RAM width %dbits %s\n",
1064 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1065 return amdgpu_ttm_init(adev);
1069 * amdgpu_bo_fini - tear down memory manager
1070 * @adev: amdgpu device object
1072 * Reverses amdgpu_bo_init() to tear down memory manager.
1074 void amdgpu_bo_fini(struct amdgpu_device *adev)
1076 amdgpu_ttm_fini(adev);
1077 if (!adev->gmc.xgmi.connected_to_cpu) {
1078 arch_phys_wc_del(adev->gmc.vram_mtrr);
1079 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1084 * amdgpu_bo_set_tiling_flags - set tiling flags
1085 * @bo: &amdgpu_bo buffer object
1086 * @tiling_flags: new flags
1088 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1089 * kernel driver to set the tiling flags on a buffer.
1092 * 0 for success or a negative error code on failure.
1094 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1096 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1097 struct amdgpu_bo_user *ubo;
1099 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1100 if (adev->family <= AMDGPU_FAMILY_CZ &&
1101 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1104 ubo = to_amdgpu_bo_user(bo);
1105 ubo->tiling_flags = tiling_flags;
1110 * amdgpu_bo_get_tiling_flags - get tiling flags
1111 * @bo: &amdgpu_bo buffer object
1112 * @tiling_flags: returned flags
1114 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1115 * set the tiling flags on a buffer.
1117 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1119 struct amdgpu_bo_user *ubo;
1121 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1122 dma_resv_assert_held(bo->tbo.base.resv);
1123 ubo = to_amdgpu_bo_user(bo);
1126 *tiling_flags = ubo->tiling_flags;
1130 * amdgpu_bo_set_metadata - set metadata
1131 * @bo: &amdgpu_bo buffer object
1132 * @metadata: new metadata
1133 * @metadata_size: size of the new metadata
1134 * @flags: flags of the new metadata
1136 * Sets buffer object's metadata, its size and flags.
1137 * Used via GEM ioctl.
1140 * 0 for success or a negative error code on failure.
1142 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1143 uint32_t metadata_size, uint64_t flags)
1145 struct amdgpu_bo_user *ubo;
1148 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1149 ubo = to_amdgpu_bo_user(bo);
1150 if (!metadata_size) {
1151 if (ubo->metadata_size) {
1152 kfree(ubo->metadata);
1153 ubo->metadata = NULL;
1154 ubo->metadata_size = 0;
1159 if (metadata == NULL)
1162 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1166 kfree(ubo->metadata);
1167 ubo->metadata_flags = flags;
1168 ubo->metadata = buffer;
1169 ubo->metadata_size = metadata_size;
1175 * amdgpu_bo_get_metadata - get metadata
1176 * @bo: &amdgpu_bo buffer object
1177 * @buffer: returned metadata
1178 * @buffer_size: size of the buffer
1179 * @metadata_size: size of the returned metadata
1180 * @flags: flags of the returned metadata
1182 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1183 * less than metadata_size.
1184 * Used via GEM ioctl.
1187 * 0 for success or a negative error code on failure.
1189 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1190 size_t buffer_size, uint32_t *metadata_size,
1193 struct amdgpu_bo_user *ubo;
1195 if (!buffer && !metadata_size)
1198 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1199 ubo = to_amdgpu_bo_user(bo);
1201 *metadata_size = ubo->metadata_size;
1204 if (buffer_size < ubo->metadata_size)
1207 if (ubo->metadata_size)
1208 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1212 *flags = ubo->metadata_flags;
1218 * amdgpu_bo_move_notify - notification about a memory move
1219 * @bo: pointer to a buffer object
1220 * @evict: if this move is evicting the buffer from the graphics address space
1221 * @new_mem: new information of the bufer object
1223 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1225 * TTM driver callback which is called when ttm moves a buffer.
1227 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1229 struct ttm_resource *new_mem)
1231 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1232 struct amdgpu_bo *abo;
1233 struct ttm_resource *old_mem = &bo->mem;
1235 if (!amdgpu_bo_is_amdgpu_bo(bo))
1238 abo = ttm_to_amdgpu_bo(bo);
1239 amdgpu_vm_bo_invalidate(adev, abo, evict);
1241 amdgpu_bo_kunmap(abo);
1243 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1244 bo->mem.mem_type != TTM_PL_SYSTEM)
1245 dma_buf_move_notify(abo->tbo.base.dma_buf);
1247 /* remember the eviction */
1249 atomic64_inc(&adev->num_evictions);
1251 /* update statistics */
1255 /* move_notify is called before move happens */
1256 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1259 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1260 uint64_t *gtt_mem, uint64_t *cpu_mem)
1262 unsigned int domain;
1264 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
1266 case AMDGPU_GEM_DOMAIN_VRAM:
1267 *vram_mem += amdgpu_bo_size(bo);
1269 case AMDGPU_GEM_DOMAIN_GTT:
1270 *gtt_mem += amdgpu_bo_size(bo);
1272 case AMDGPU_GEM_DOMAIN_CPU:
1274 *cpu_mem += amdgpu_bo_size(bo);
1280 * amdgpu_bo_release_notify - notification about a BO being released
1281 * @bo: pointer to a buffer object
1283 * Wipes VRAM buffers whose contents should not be leaked before the
1284 * memory is released.
1286 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1288 struct dma_fence *fence = NULL;
1289 struct amdgpu_bo *abo;
1292 if (!amdgpu_bo_is_amdgpu_bo(bo))
1295 abo = ttm_to_amdgpu_bo(bo);
1298 amdgpu_amdkfd_unreserve_memory_limit(abo);
1300 /* We only remove the fence if the resv has individualized. */
1301 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1302 && bo->base.resv != &bo->base._resv);
1303 if (bo->base.resv == &bo->base._resv)
1304 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1306 if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
1307 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1310 dma_resv_lock(bo->base.resv, NULL);
1312 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1314 amdgpu_bo_fence(abo, fence, false);
1315 dma_fence_put(fence);
1318 dma_resv_unlock(bo->base.resv);
1322 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1323 * @bo: pointer to a buffer object
1325 * Notifies the driver we are taking a fault on this BO and have reserved it,
1326 * also performs bookkeeping.
1327 * TTM driver callback for dealing with vm faults.
1330 * 0 for success or a negative error code on failure.
1332 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1334 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1335 struct ttm_operation_ctx ctx = { false, false };
1336 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1337 unsigned long offset;
1340 /* Remember that this BO was accessed by the CPU */
1341 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1343 if (bo->mem.mem_type != TTM_PL_VRAM)
1346 offset = bo->mem.start << PAGE_SHIFT;
1347 if ((offset + bo->base.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 + bo->base.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, VRAM_CONTIGUOUS);
1567 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1568 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);