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_device) {
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) {
136 places[c].mem_type = TTM_PL_TT;
141 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
144 places[c].mem_type = TTM_PL_SYSTEM;
149 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
152 places[c].mem_type = AMDGPU_PL_GDS;
157 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
160 places[c].mem_type = AMDGPU_PL_GWS;
165 if (domain & AMDGPU_GEM_DOMAIN_OA) {
168 places[c].mem_type = AMDGPU_PL_OA;
176 places[c].mem_type = TTM_PL_SYSTEM;
181 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
183 placement->num_placement = c;
184 placement->placement = places;
186 placement->num_busy_placement = c;
187 placement->busy_placement = places;
191 * amdgpu_bo_create_reserved - create reserved BO for kernel use
193 * @adev: amdgpu device object
194 * @size: size for the new BO
195 * @align: alignment for the new BO
196 * @domain: where to place it
197 * @bo_ptr: used to initialize BOs in structures
198 * @gpu_addr: GPU addr of the pinned BO
199 * @cpu_addr: optional CPU address mapping
201 * Allocates and pins a BO for kernel internal use, and returns it still
204 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
207 * 0 on success, negative error code otherwise.
209 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
210 unsigned long size, int align,
211 u32 domain, struct amdgpu_bo **bo_ptr,
212 u64 *gpu_addr, void **cpu_addr)
214 struct amdgpu_bo_param bp;
219 amdgpu_bo_unref(bo_ptr);
223 memset(&bp, 0, sizeof(bp));
225 bp.byte_align = align;
227 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
228 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
229 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
230 bp.type = ttm_bo_type_kernel;
232 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
235 r = amdgpu_bo_create(adev, &bp, bo_ptr);
237 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
244 r = amdgpu_bo_reserve(*bo_ptr, false);
246 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
250 r = amdgpu_bo_pin(*bo_ptr, domain);
252 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
253 goto error_unreserve;
256 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
258 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
263 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
266 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
268 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
276 amdgpu_bo_unpin(*bo_ptr);
278 amdgpu_bo_unreserve(*bo_ptr);
282 amdgpu_bo_unref(bo_ptr);
288 * amdgpu_bo_create_kernel - create BO for kernel use
290 * @adev: amdgpu device object
291 * @size: size for the new BO
292 * @align: alignment for the new BO
293 * @domain: where to place it
294 * @bo_ptr: used to initialize BOs in structures
295 * @gpu_addr: GPU addr of the pinned BO
296 * @cpu_addr: optional CPU address mapping
298 * Allocates and pins a BO for kernel internal use.
300 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
303 * 0 on success, negative error code otherwise.
305 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
306 unsigned long size, int align,
307 u32 domain, struct amdgpu_bo **bo_ptr,
308 u64 *gpu_addr, void **cpu_addr)
312 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
319 amdgpu_bo_unreserve(*bo_ptr);
325 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
327 * @adev: amdgpu device object
328 * @offset: offset of the BO
329 * @size: size of the BO
330 * @domain: where to place it
331 * @bo_ptr: used to initialize BOs in structures
332 * @cpu_addr: optional CPU address mapping
334 * Creates a kernel BO at a specific offset in the address space of the domain.
337 * 0 on success, negative error code otherwise.
339 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
340 uint64_t offset, uint64_t size, uint32_t domain,
341 struct amdgpu_bo **bo_ptr, void **cpu_addr)
343 struct ttm_operation_ctx ctx = { false, false };
348 size = ALIGN(size, PAGE_SIZE);
350 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
355 if ((*bo_ptr) == NULL)
359 * Remove the original mem node and create a new one at the request
363 amdgpu_bo_kunmap(*bo_ptr);
365 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.mem);
367 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
368 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
369 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
371 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
372 &(*bo_ptr)->tbo.mem, &ctx);
377 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
382 amdgpu_bo_unreserve(*bo_ptr);
386 amdgpu_bo_unreserve(*bo_ptr);
387 amdgpu_bo_unref(bo_ptr);
392 * amdgpu_bo_free_kernel - free BO for kernel use
394 * @bo: amdgpu BO to free
395 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
396 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
398 * unmaps and unpin a BO for kernel internal use.
400 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
406 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
408 amdgpu_bo_kunmap(*bo);
410 amdgpu_bo_unpin(*bo);
411 amdgpu_bo_unreserve(*bo);
422 /* Validate bo size is bit bigger then the request domain */
423 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
424 unsigned long size, u32 domain)
426 struct ttm_resource_manager *man = NULL;
429 * If GTT is part of requested domains the check must succeed to
430 * allow fall back to GTT
432 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
433 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
435 if (size < (man->size << PAGE_SHIFT))
441 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
442 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
444 if (size < (man->size << PAGE_SHIFT))
451 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
455 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
456 man->size << PAGE_SHIFT);
460 bool amdgpu_bo_support_uswc(u64 bo_flags)
464 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
465 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
468 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
469 /* Don't try to enable write-combining when it can't work, or things
471 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
474 #ifndef CONFIG_COMPILE_TEST
475 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
476 thanks to write-combining
479 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
480 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
481 "better performance thanks to write-combining\n");
484 /* For architectures that don't support WC memory,
485 * mask out the WC flag from the BO
487 if (!drm_arch_can_wc_memory())
495 * amdgpu_bo_create - create an &amdgpu_bo buffer object
496 * @adev: amdgpu device object
497 * @bp: parameters to be used for the buffer object
498 * @bo_ptr: pointer to the buffer object pointer
500 * Creates an &amdgpu_bo buffer object.
503 * 0 for success or a negative error code on failure.
505 int amdgpu_bo_create(struct amdgpu_device *adev,
506 struct amdgpu_bo_param *bp,
507 struct amdgpu_bo **bo_ptr)
509 struct ttm_operation_ctx ctx = {
510 .interruptible = (bp->type != ttm_bo_type_kernel),
511 .no_wait_gpu = bp->no_wait_gpu,
512 /* We opt to avoid OOM on system pages allocations */
513 .gfp_retry_mayfail = true,
514 .allow_res_evict = bp->type != ttm_bo_type_kernel,
517 struct amdgpu_bo *bo;
518 unsigned long page_align, size = bp->size;
521 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
522 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
523 /* GWS and OA don't need any alignment. */
524 page_align = bp->byte_align;
526 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
527 /* Both size and alignment must be a multiple of 4. */
528 page_align = ALIGN(bp->byte_align, 4);
529 size = ALIGN(size, 4) << PAGE_SHIFT;
531 /* Memory should be aligned at least to a page size. */
532 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
533 size = ALIGN(size, PAGE_SIZE);
536 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
539 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
542 bo = kzalloc(bp->bo_ptr_size, GFP_KERNEL);
545 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
546 INIT_LIST_HEAD(&bo->shadow_list);
548 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
550 bo->allowed_domains = bo->preferred_domains;
551 if (bp->type != ttm_bo_type_kernel &&
552 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
553 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
555 bo->flags = bp->flags;
557 if (!amdgpu_bo_support_uswc(bo->flags))
558 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
560 bo->tbo.bdev = &adev->mman.bdev;
561 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
562 AMDGPU_GEM_DOMAIN_GDS))
563 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
565 amdgpu_bo_placement_from_domain(bo, bp->domain);
566 if (bp->type == ttm_bo_type_kernel)
567 bo->tbo.priority = 1;
569 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
570 &bo->placement, page_align, &ctx, NULL,
571 bp->resv, &amdgpu_bo_destroy);
572 if (unlikely(r != 0))
575 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
576 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
577 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
578 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
581 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
583 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
584 bo->tbo.mem.mem_type == TTM_PL_VRAM) {
585 struct dma_fence *fence;
587 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
591 amdgpu_bo_fence(bo, fence, false);
592 dma_fence_put(bo->tbo.moving);
593 bo->tbo.moving = dma_fence_get(fence);
594 dma_fence_put(fence);
597 amdgpu_bo_unreserve(bo);
600 trace_amdgpu_bo_create(bo);
602 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
603 if (bp->type == ttm_bo_type_device)
604 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
610 dma_resv_unlock(bo->tbo.base.resv);
611 amdgpu_bo_unref(&bo);
615 int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
617 struct amdgpu_bo *bo)
619 struct amdgpu_bo_param bp;
625 memset(&bp, 0, sizeof(bp));
627 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
628 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
629 bp.type = ttm_bo_type_kernel;
630 bp.resv = bo->tbo.base.resv;
631 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
633 r = amdgpu_bo_create(adev, &bp, &bo->shadow);
635 bo->shadow->parent = amdgpu_bo_ref(bo);
636 mutex_lock(&adev->shadow_list_lock);
637 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
638 mutex_unlock(&adev->shadow_list_lock);
645 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
646 * @adev: amdgpu device object
647 * @bp: parameters to be used for the buffer object
648 * @ubo_ptr: pointer to the buffer object pointer
650 * Create a BO to be used by user application;
653 * 0 for success or a negative error code on failure.
656 int amdgpu_bo_create_user(struct amdgpu_device *adev,
657 struct amdgpu_bo_param *bp,
658 struct amdgpu_bo_user **ubo_ptr)
660 struct amdgpu_bo *bo_ptr;
663 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
664 r = amdgpu_bo_create(adev, bp, &bo_ptr);
668 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
672 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
673 * @bo: pointer to the buffer object
675 * Sets placement according to domain; and changes placement and caching
676 * policy of the buffer object according to the placement.
677 * This is used for validating shadow bos. It calls ttm_bo_validate() to
678 * make sure the buffer is resident where it needs to be.
681 * 0 for success or a negative error code on failure.
683 int amdgpu_bo_validate(struct amdgpu_bo *bo)
685 struct ttm_operation_ctx ctx = { false, false };
689 if (bo->tbo.pin_count)
692 domain = bo->preferred_domains;
695 amdgpu_bo_placement_from_domain(bo, domain);
696 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
697 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
698 domain = bo->allowed_domains;
706 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
708 * @shadow: &amdgpu_bo shadow to be restored
709 * @fence: dma_fence associated with the operation
711 * Copies a buffer object's shadow content back to the object.
712 * This is used for recovering a buffer from its shadow in case of a gpu
713 * reset where vram context may be lost.
716 * 0 for success or a negative error code on failure.
718 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
721 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
722 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
723 uint64_t shadow_addr, parent_addr;
725 shadow_addr = amdgpu_bo_gpu_offset(shadow);
726 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
728 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
729 amdgpu_bo_size(shadow), NULL, fence,
734 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
735 * @bo: &amdgpu_bo buffer object to be mapped
736 * @ptr: kernel virtual address to be returned
738 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
739 * amdgpu_bo_kptr() to get the kernel virtual address.
742 * 0 for success or a negative error code on failure.
744 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
749 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
752 kptr = amdgpu_bo_kptr(bo);
759 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false, false,
760 MAX_SCHEDULE_TIMEOUT);
764 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.mem.num_pages, &bo->kmap);
769 *ptr = amdgpu_bo_kptr(bo);
775 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
776 * @bo: &amdgpu_bo buffer object
778 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
781 * the virtual address of a buffer object area.
783 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
787 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
791 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
792 * @bo: &amdgpu_bo buffer object to be unmapped
794 * Unmaps a kernel map set up by amdgpu_bo_kmap().
796 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
799 ttm_bo_kunmap(&bo->kmap);
803 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
804 * @bo: &amdgpu_bo buffer object
806 * References the contained &ttm_buffer_object.
809 * a refcounted pointer to the &amdgpu_bo buffer object.
811 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
816 ttm_bo_get(&bo->tbo);
821 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
822 * @bo: &amdgpu_bo buffer object
824 * Unreferences the contained &ttm_buffer_object and clear the pointer
826 void amdgpu_bo_unref(struct amdgpu_bo **bo)
828 struct ttm_buffer_object *tbo;
839 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
840 * @bo: &amdgpu_bo buffer object to be pinned
841 * @domain: domain to be pinned to
842 * @min_offset: the start of requested address range
843 * @max_offset: the end of requested address range
845 * Pins the buffer object according to requested domain and address range. If
846 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
847 * pin_count and pin_size accordingly.
849 * Pinning means to lock pages in memory along with keeping them at a fixed
850 * offset. It is required when a buffer can not be moved, for example, when
851 * a display buffer is being scanned out.
853 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
854 * where to pin a buffer if there are specific restrictions on where a buffer
858 * 0 for success or a negative error code on failure.
860 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
861 u64 min_offset, u64 max_offset)
863 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
864 struct ttm_operation_ctx ctx = { false, false };
867 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
870 if (WARN_ON_ONCE(min_offset > max_offset))
873 /* A shared bo cannot be migrated to VRAM */
874 if (bo->prime_shared_count || bo->tbo.base.import_attach) {
875 if (domain & AMDGPU_GEM_DOMAIN_GTT)
876 domain = AMDGPU_GEM_DOMAIN_GTT;
881 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
882 * See function amdgpu_display_supported_domains()
884 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
886 if (bo->tbo.pin_count) {
887 uint32_t mem_type = bo->tbo.mem.mem_type;
888 uint32_t mem_flags = bo->tbo.mem.placement;
890 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
893 if ((mem_type == TTM_PL_VRAM) &&
894 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
895 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
898 ttm_bo_pin(&bo->tbo);
900 if (max_offset != 0) {
901 u64 domain_start = amdgpu_ttm_domain_start(adev,
903 WARN_ON_ONCE(max_offset <
904 (amdgpu_bo_gpu_offset(bo) - domain_start));
910 if (bo->tbo.base.import_attach)
911 dma_buf_pin(bo->tbo.base.import_attach);
913 /* force to pin into visible video ram */
914 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
915 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
916 amdgpu_bo_placement_from_domain(bo, domain);
917 for (i = 0; i < bo->placement.num_placement; i++) {
920 fpfn = min_offset >> PAGE_SHIFT;
921 lpfn = max_offset >> PAGE_SHIFT;
923 if (fpfn > bo->placements[i].fpfn)
924 bo->placements[i].fpfn = fpfn;
925 if (!bo->placements[i].lpfn ||
926 (lpfn && lpfn < bo->placements[i].lpfn))
927 bo->placements[i].lpfn = lpfn;
930 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
932 dev_err(adev->dev, "%p pin failed\n", bo);
936 ttm_bo_pin(&bo->tbo);
938 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
939 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
940 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
941 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
942 &adev->visible_pin_size);
943 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
944 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
952 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
953 * @bo: &amdgpu_bo buffer object to be pinned
954 * @domain: domain to be pinned to
956 * A simple wrapper to amdgpu_bo_pin_restricted().
957 * Provides a simpler API for buffers that do not have any strict restrictions
958 * on where a buffer must be located.
961 * 0 for success or a negative error code on failure.
963 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
965 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
966 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
970 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
971 * @bo: &amdgpu_bo buffer object to be unpinned
973 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
974 * Changes placement and pin size accordingly.
977 * 0 for success or a negative error code on failure.
979 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
981 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
983 ttm_bo_unpin(&bo->tbo);
984 if (bo->tbo.pin_count)
987 if (bo->tbo.base.import_attach)
988 dma_buf_unpin(bo->tbo.base.import_attach);
990 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
991 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
992 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
993 &adev->visible_pin_size);
994 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
995 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1000 * amdgpu_bo_evict_vram - evict VRAM buffers
1001 * @adev: amdgpu device object
1003 * Evicts all VRAM buffers on the lru list of the memory type.
1004 * Mainly used for evicting vram at suspend time.
1007 * 0 for success or a negative error code on failure.
1009 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1011 struct ttm_resource_manager *man;
1013 if (adev->in_s3 && (adev->flags & AMD_IS_APU)) {
1014 /* No need to evict vram on APUs for suspend to ram */
1018 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1019 return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
1022 static const char *amdgpu_vram_names[] = {
1037 * amdgpu_bo_init - initialize memory manager
1038 * @adev: amdgpu device object
1040 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1043 * 0 for success or a negative error code on failure.
1045 int amdgpu_bo_init(struct amdgpu_device *adev)
1047 /* On A+A platform, VRAM can be mapped as WB */
1048 if (!adev->gmc.xgmi.connected_to_cpu) {
1049 /* reserve PAT memory space to WC for VRAM */
1050 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1051 adev->gmc.aper_size);
1053 /* Add an MTRR for the VRAM */
1054 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1055 adev->gmc.aper_size);
1058 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1059 adev->gmc.mc_vram_size >> 20,
1060 (unsigned long long)adev->gmc.aper_size >> 20);
1061 DRM_INFO("RAM width %dbits %s\n",
1062 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1063 return amdgpu_ttm_init(adev);
1067 * amdgpu_bo_fini - tear down memory manager
1068 * @adev: amdgpu device object
1070 * Reverses amdgpu_bo_init() to tear down memory manager.
1072 void amdgpu_bo_fini(struct amdgpu_device *adev)
1074 amdgpu_ttm_fini(adev);
1078 * amdgpu_bo_set_tiling_flags - set tiling flags
1079 * @bo: &amdgpu_bo buffer object
1080 * @tiling_flags: new flags
1082 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1083 * kernel driver to set the tiling flags on a buffer.
1086 * 0 for success or a negative error code on failure.
1088 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1090 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1091 struct amdgpu_bo_user *ubo;
1093 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1094 if (adev->family <= AMDGPU_FAMILY_CZ &&
1095 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1098 ubo = to_amdgpu_bo_user(bo);
1099 ubo->tiling_flags = tiling_flags;
1104 * amdgpu_bo_get_tiling_flags - get tiling flags
1105 * @bo: &amdgpu_bo buffer object
1106 * @tiling_flags: returned flags
1108 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1109 * set the tiling flags on a buffer.
1111 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1113 struct amdgpu_bo_user *ubo;
1115 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1116 dma_resv_assert_held(bo->tbo.base.resv);
1117 ubo = to_amdgpu_bo_user(bo);
1120 *tiling_flags = ubo->tiling_flags;
1124 * amdgpu_bo_set_metadata - set metadata
1125 * @bo: &amdgpu_bo buffer object
1126 * @metadata: new metadata
1127 * @metadata_size: size of the new metadata
1128 * @flags: flags of the new metadata
1130 * Sets buffer object's metadata, its size and flags.
1131 * Used via GEM ioctl.
1134 * 0 for success or a negative error code on failure.
1136 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1137 uint32_t metadata_size, uint64_t flags)
1139 struct amdgpu_bo_user *ubo;
1142 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1143 ubo = to_amdgpu_bo_user(bo);
1144 if (!metadata_size) {
1145 if (ubo->metadata_size) {
1146 kfree(ubo->metadata);
1147 ubo->metadata = NULL;
1148 ubo->metadata_size = 0;
1153 if (metadata == NULL)
1156 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1160 kfree(ubo->metadata);
1161 ubo->metadata_flags = flags;
1162 ubo->metadata = buffer;
1163 ubo->metadata_size = metadata_size;
1169 * amdgpu_bo_get_metadata - get metadata
1170 * @bo: &amdgpu_bo buffer object
1171 * @buffer: returned metadata
1172 * @buffer_size: size of the buffer
1173 * @metadata_size: size of the returned metadata
1174 * @flags: flags of the returned metadata
1176 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1177 * less than metadata_size.
1178 * Used via GEM ioctl.
1181 * 0 for success or a negative error code on failure.
1183 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1184 size_t buffer_size, uint32_t *metadata_size,
1187 struct amdgpu_bo_user *ubo;
1189 if (!buffer && !metadata_size)
1192 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1193 ubo = to_amdgpu_bo_user(bo);
1195 if (buffer_size < ubo->metadata_size)
1198 if (ubo->metadata_size)
1199 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1203 *metadata_size = ubo->metadata_size;
1205 *flags = ubo->metadata_flags;
1211 * amdgpu_bo_move_notify - notification about a memory move
1212 * @bo: pointer to a buffer object
1213 * @evict: if this move is evicting the buffer from the graphics address space
1214 * @new_mem: new information of the bufer object
1216 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1218 * TTM driver callback which is called when ttm moves a buffer.
1220 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1222 struct ttm_resource *new_mem)
1224 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1225 struct amdgpu_bo *abo;
1226 struct ttm_resource *old_mem = &bo->mem;
1228 if (!amdgpu_bo_is_amdgpu_bo(bo))
1231 abo = ttm_to_amdgpu_bo(bo);
1232 amdgpu_vm_bo_invalidate(adev, abo, evict);
1234 amdgpu_bo_kunmap(abo);
1236 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1237 bo->mem.mem_type != TTM_PL_SYSTEM)
1238 dma_buf_move_notify(abo->tbo.base.dma_buf);
1240 /* remember the eviction */
1242 atomic64_inc(&adev->num_evictions);
1244 /* update statistics */
1248 /* move_notify is called before move happens */
1249 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1252 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1253 uint64_t *gtt_mem, uint64_t *cpu_mem)
1255 unsigned int domain;
1257 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
1259 case AMDGPU_GEM_DOMAIN_VRAM:
1260 *vram_mem += amdgpu_bo_size(bo);
1262 case AMDGPU_GEM_DOMAIN_GTT:
1263 *gtt_mem += amdgpu_bo_size(bo);
1265 case AMDGPU_GEM_DOMAIN_CPU:
1267 *cpu_mem += amdgpu_bo_size(bo);
1273 * amdgpu_bo_release_notify - notification about a BO being released
1274 * @bo: pointer to a buffer object
1276 * Wipes VRAM buffers whose contents should not be leaked before the
1277 * memory is released.
1279 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1281 struct dma_fence *fence = NULL;
1282 struct amdgpu_bo *abo;
1285 if (!amdgpu_bo_is_amdgpu_bo(bo))
1288 abo = ttm_to_amdgpu_bo(bo);
1291 amdgpu_amdkfd_unreserve_memory_limit(abo);
1293 /* We only remove the fence if the resv has individualized. */
1294 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1295 && bo->base.resv != &bo->base._resv);
1296 if (bo->base.resv == &bo->base._resv)
1297 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1299 if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
1300 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1303 dma_resv_lock(bo->base.resv, NULL);
1305 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1307 amdgpu_bo_fence(abo, fence, false);
1308 dma_fence_put(fence);
1311 dma_resv_unlock(bo->base.resv);
1315 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1316 * @bo: pointer to a buffer object
1318 * Notifies the driver we are taking a fault on this BO and have reserved it,
1319 * also performs bookkeeping.
1320 * TTM driver callback for dealing with vm faults.
1323 * 0 for success or a negative error code on failure.
1325 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1327 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1328 struct ttm_operation_ctx ctx = { false, false };
1329 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1330 unsigned long offset;
1333 /* Remember that this BO was accessed by the CPU */
1334 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1336 if (bo->mem.mem_type != TTM_PL_VRAM)
1339 offset = bo->mem.start << PAGE_SHIFT;
1340 if ((offset + bo->base.size) <= adev->gmc.visible_vram_size)
1343 /* Can't move a pinned BO to visible VRAM */
1344 if (abo->tbo.pin_count > 0)
1345 return VM_FAULT_SIGBUS;
1347 /* hurrah the memory is not visible ! */
1348 atomic64_inc(&adev->num_vram_cpu_page_faults);
1349 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1350 AMDGPU_GEM_DOMAIN_GTT);
1352 /* Avoid costly evictions; only set GTT as a busy placement */
1353 abo->placement.num_busy_placement = 1;
1354 abo->placement.busy_placement = &abo->placements[1];
1356 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1357 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1358 return VM_FAULT_NOPAGE;
1359 else if (unlikely(r))
1360 return VM_FAULT_SIGBUS;
1362 offset = bo->mem.start << PAGE_SHIFT;
1363 /* this should never happen */
1364 if (bo->mem.mem_type == TTM_PL_VRAM &&
1365 (offset + bo->base.size) > adev->gmc.visible_vram_size)
1366 return VM_FAULT_SIGBUS;
1368 ttm_bo_move_to_lru_tail_unlocked(bo);
1373 * amdgpu_bo_fence - add fence to buffer object
1375 * @bo: buffer object in question
1376 * @fence: fence to add
1377 * @shared: true if fence should be added shared
1380 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1383 struct dma_resv *resv = bo->tbo.base.resv;
1386 dma_resv_add_shared_fence(resv, fence);
1388 dma_resv_add_excl_fence(resv, fence);
1392 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1394 * @adev: amdgpu device pointer
1395 * @resv: reservation object to sync to
1396 * @sync_mode: synchronization mode
1397 * @owner: fence owner
1398 * @intr: Whether the wait is interruptible
1400 * Extract the fences from the reservation object and waits for them to finish.
1403 * 0 on success, errno otherwise.
1405 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1406 enum amdgpu_sync_mode sync_mode, void *owner,
1409 struct amdgpu_sync sync;
1412 amdgpu_sync_create(&sync);
1413 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1414 r = amdgpu_sync_wait(&sync, intr);
1415 amdgpu_sync_free(&sync);
1420 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1421 * @bo: buffer object to wait for
1422 * @owner: fence owner
1423 * @intr: Whether the wait is interruptible
1425 * Wrapper to wait for fences in a BO.
1427 * 0 on success, errno otherwise.
1429 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1431 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1433 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1434 AMDGPU_SYNC_NE_OWNER, owner, intr);
1438 * amdgpu_bo_gpu_offset - return GPU offset of bo
1439 * @bo: amdgpu object for which we query the offset
1441 * Note: object should either be pinned or reserved when calling this
1442 * function, it might be useful to add check for this for debugging.
1445 * current GPU offset of the object.
1447 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1449 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1450 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1451 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1452 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1453 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1454 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1456 return amdgpu_bo_gpu_offset_no_check(bo);
1460 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1461 * @bo: amdgpu object for which we query the offset
1464 * current GPU offset of the object without raising warnings.
1466 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1468 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1471 offset = (bo->tbo.mem.start << PAGE_SHIFT) +
1472 amdgpu_ttm_domain_start(adev, bo->tbo.mem.mem_type);
1474 return amdgpu_gmc_sign_extend(offset);
1478 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1479 * @adev: amdgpu device object
1480 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1483 * Which of the allowed domains is preferred for pinning the BO for scanout.
1485 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1488 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1489 domain = AMDGPU_GEM_DOMAIN_VRAM;
1490 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1491 domain = AMDGPU_GEM_DOMAIN_GTT;
1496 #if defined(CONFIG_DEBUG_FS)
1497 #define amdgpu_bo_print_flag(m, bo, flag) \
1499 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1500 seq_printf((m), " " #flag); \
1505 * amdgpu_bo_print_info - print BO info in debugfs file
1507 * @id: Index or Id of the BO
1508 * @bo: Requested BO for printing info
1511 * Print BO information in debugfs file
1514 * Size of the BO in bytes.
1516 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1518 struct dma_buf_attachment *attachment;
1519 struct dma_buf *dma_buf;
1520 unsigned int domain;
1521 const char *placement;
1522 unsigned int pin_count;
1525 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
1527 case AMDGPU_GEM_DOMAIN_VRAM:
1530 case AMDGPU_GEM_DOMAIN_GTT:
1533 case AMDGPU_GEM_DOMAIN_CPU:
1539 size = amdgpu_bo_size(bo);
1540 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1541 id, size, placement);
1543 pin_count = READ_ONCE(bo->tbo.pin_count);
1545 seq_printf(m, " pin count %d", pin_count);
1547 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1548 attachment = READ_ONCE(bo->tbo.base.import_attach);
1551 seq_printf(m, " imported from %p", dma_buf);
1553 seq_printf(m, " exported as %p", dma_buf);
1555 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1556 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1557 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1558 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1559 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1560 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1561 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);