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->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].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
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].flags = TTM_PL_FLAG_TT;
156 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
157 places[c].flags |= TTM_PL_FLAG_WC |
158 TTM_PL_FLAG_UNCACHED;
160 places[c].flags |= TTM_PL_FLAG_CACHED;
164 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
167 places[c].flags = TTM_PL_FLAG_SYSTEM;
168 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
169 places[c].flags |= TTM_PL_FLAG_WC |
170 TTM_PL_FLAG_UNCACHED;
172 places[c].flags |= TTM_PL_FLAG_CACHED;
176 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
179 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
183 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
186 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
190 if (domain & AMDGPU_GEM_DOMAIN_OA) {
193 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
200 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
204 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
206 placement->num_placement = c;
207 placement->placement = places;
209 placement->num_busy_placement = c;
210 placement->busy_placement = places;
214 * amdgpu_bo_create_reserved - create reserved BO for kernel use
216 * @adev: amdgpu device object
217 * @size: size for the new BO
218 * @align: alignment for the new BO
219 * @domain: where to place it
220 * @bo_ptr: used to initialize BOs in structures
221 * @gpu_addr: GPU addr of the pinned BO
222 * @cpu_addr: optional CPU address mapping
224 * Allocates and pins a BO for kernel internal use, and returns it still
227 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
230 * 0 on success, negative error code otherwise.
232 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
233 unsigned long size, int align,
234 u32 domain, struct amdgpu_bo **bo_ptr,
235 u64 *gpu_addr, void **cpu_addr)
237 struct amdgpu_bo_param bp;
242 amdgpu_bo_unref(bo_ptr);
246 memset(&bp, 0, sizeof(bp));
248 bp.byte_align = align;
250 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
251 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
252 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
253 bp.type = ttm_bo_type_kernel;
257 r = amdgpu_bo_create(adev, &bp, bo_ptr);
259 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
266 r = amdgpu_bo_reserve(*bo_ptr, false);
268 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
272 r = amdgpu_bo_pin(*bo_ptr, domain);
274 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
275 goto error_unreserve;
278 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
280 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
285 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
288 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
290 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
298 amdgpu_bo_unpin(*bo_ptr);
300 amdgpu_bo_unreserve(*bo_ptr);
304 amdgpu_bo_unref(bo_ptr);
310 * amdgpu_bo_create_kernel - create BO for kernel use
312 * @adev: amdgpu device object
313 * @size: size for the new BO
314 * @align: alignment for the new BO
315 * @domain: where to place it
316 * @bo_ptr: used to initialize BOs in structures
317 * @gpu_addr: GPU addr of the pinned BO
318 * @cpu_addr: optional CPU address mapping
320 * Allocates and pins a BO for kernel internal use.
322 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
325 * 0 on success, negative error code otherwise.
327 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
328 unsigned long size, int align,
329 u32 domain, struct amdgpu_bo **bo_ptr,
330 u64 *gpu_addr, void **cpu_addr)
334 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
341 amdgpu_bo_unreserve(*bo_ptr);
347 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
349 * @adev: amdgpu device object
350 * @offset: offset of the BO
351 * @size: size of the BO
352 * @domain: where to place it
353 * @bo_ptr: used to initialize BOs in structures
354 * @cpu_addr: optional CPU address mapping
356 * Creates a kernel BO at a specific offset in the address space of the domain.
359 * 0 on success, negative error code otherwise.
361 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
362 uint64_t offset, uint64_t size, uint32_t domain,
363 struct amdgpu_bo **bo_ptr, void **cpu_addr)
365 struct ttm_operation_ctx ctx = { false, false };
370 size = ALIGN(size, PAGE_SIZE);
372 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
377 if ((*bo_ptr) == NULL)
381 * Remove the original mem node and create a new one at the request
385 amdgpu_bo_kunmap(*bo_ptr);
387 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.mem);
389 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
390 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
391 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
393 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
394 &(*bo_ptr)->tbo.mem, &ctx);
399 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
404 amdgpu_bo_unreserve(*bo_ptr);
408 amdgpu_bo_unreserve(*bo_ptr);
409 amdgpu_bo_unref(bo_ptr);
414 * amdgpu_bo_free_kernel - free BO for kernel use
416 * @bo: amdgpu BO to free
417 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
418 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
420 * unmaps and unpin a BO for kernel internal use.
422 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
428 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
430 amdgpu_bo_kunmap(*bo);
432 amdgpu_bo_unpin(*bo);
433 amdgpu_bo_unreserve(*bo);
444 /* Validate bo size is bit bigger then the request domain */
445 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
446 unsigned long size, u32 domain)
448 struct ttm_resource_manager *man = NULL;
451 * If GTT is part of requested domains the check must succeed to
452 * allow fall back to GTT
454 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
455 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
457 if (size < (man->size << PAGE_SHIFT))
463 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
464 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
466 if (size < (man->size << PAGE_SHIFT))
473 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
477 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
478 man->size << PAGE_SHIFT);
482 bool amdgpu_bo_support_uswc(u64 bo_flags)
486 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
487 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
490 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
491 /* Don't try to enable write-combining when it can't work, or things
493 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
496 #ifndef CONFIG_COMPILE_TEST
497 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
498 thanks to write-combining
501 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
502 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
503 "better performance thanks to write-combining\n");
506 /* For architectures that don't support WC memory,
507 * mask out the WC flag from the BO
509 if (!drm_arch_can_wc_memory())
516 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
517 struct amdgpu_bo_param *bp,
518 struct amdgpu_bo **bo_ptr)
520 struct ttm_operation_ctx ctx = {
521 .interruptible = (bp->type != ttm_bo_type_kernel),
522 .no_wait_gpu = bp->no_wait_gpu,
524 .flags = bp->type != ttm_bo_type_kernel ?
525 TTM_OPT_FLAG_ALLOW_RES_EVICT : 0
527 struct amdgpu_bo *bo;
528 unsigned long page_align, size = bp->size;
532 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
533 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
534 /* GWS and OA don't need any alignment. */
535 page_align = bp->byte_align;
537 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
538 /* Both size and alignment must be a multiple of 4. */
539 page_align = ALIGN(bp->byte_align, 4);
540 size = ALIGN(size, 4) << PAGE_SHIFT;
542 /* Memory should be aligned at least to a page size. */
543 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
544 size = ALIGN(size, PAGE_SIZE);
547 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
552 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
553 sizeof(struct amdgpu_bo));
555 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
558 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
559 INIT_LIST_HEAD(&bo->shadow_list);
561 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
563 bo->allowed_domains = bo->preferred_domains;
564 if (bp->type != ttm_bo_type_kernel &&
565 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
566 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
568 bo->flags = bp->flags;
570 if (!amdgpu_bo_support_uswc(bo->flags))
571 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
573 bo->tbo.bdev = &adev->mman.bdev;
574 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
575 AMDGPU_GEM_DOMAIN_GDS))
576 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
578 amdgpu_bo_placement_from_domain(bo, bp->domain);
579 if (bp->type == ttm_bo_type_kernel)
580 bo->tbo.priority = 1;
582 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
583 &bo->placement, page_align, &ctx, acc_size,
584 NULL, bp->resv, &amdgpu_bo_destroy);
585 if (unlikely(r != 0))
588 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
589 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
590 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
591 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
594 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
596 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
597 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
598 struct dma_fence *fence;
600 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
604 amdgpu_bo_fence(bo, fence, false);
605 dma_fence_put(bo->tbo.moving);
606 bo->tbo.moving = dma_fence_get(fence);
607 dma_fence_put(fence);
610 amdgpu_bo_unreserve(bo);
613 trace_amdgpu_bo_create(bo);
615 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
616 if (bp->type == ttm_bo_type_device)
617 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
623 dma_resv_unlock(bo->tbo.base.resv);
624 amdgpu_bo_unref(&bo);
628 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
630 struct amdgpu_bo *bo)
632 struct amdgpu_bo_param bp;
638 memset(&bp, 0, sizeof(bp));
640 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
641 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
642 AMDGPU_GEM_CREATE_SHADOW;
643 bp.type = ttm_bo_type_kernel;
644 bp.resv = bo->tbo.base.resv;
646 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
648 bo->shadow->parent = amdgpu_bo_ref(bo);
649 mutex_lock(&adev->shadow_list_lock);
650 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
651 mutex_unlock(&adev->shadow_list_lock);
658 * amdgpu_bo_create - create an &amdgpu_bo buffer object
659 * @adev: amdgpu device object
660 * @bp: parameters to be used for the buffer object
661 * @bo_ptr: pointer to the buffer object pointer
663 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
665 * Shadow object is used to backup the original buffer object, and is always
669 * 0 for success or a negative error code on failure.
671 int amdgpu_bo_create(struct amdgpu_device *adev,
672 struct amdgpu_bo_param *bp,
673 struct amdgpu_bo **bo_ptr)
675 u64 flags = bp->flags;
678 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
679 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
683 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
685 WARN_ON(dma_resv_lock((*bo_ptr)->tbo.base.resv,
688 r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
691 dma_resv_unlock((*bo_ptr)->tbo.base.resv);
694 amdgpu_bo_unref(bo_ptr);
701 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
702 * @bo: pointer to the buffer object
704 * Sets placement according to domain; and changes placement and caching
705 * policy of the buffer object according to the placement.
706 * This is used for validating shadow bos. It calls ttm_bo_validate() to
707 * make sure the buffer is resident where it needs to be.
710 * 0 for success or a negative error code on failure.
712 int amdgpu_bo_validate(struct amdgpu_bo *bo)
714 struct ttm_operation_ctx ctx = { false, false };
721 domain = bo->preferred_domains;
724 amdgpu_bo_placement_from_domain(bo, domain);
725 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
726 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
727 domain = bo->allowed_domains;
735 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
737 * @shadow: &amdgpu_bo shadow to be restored
738 * @fence: dma_fence associated with the operation
740 * Copies a buffer object's shadow content back to the object.
741 * This is used for recovering a buffer from its shadow in case of a gpu
742 * reset where vram context may be lost.
745 * 0 for success or a negative error code on failure.
747 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
750 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
751 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
752 uint64_t shadow_addr, parent_addr;
754 shadow_addr = amdgpu_bo_gpu_offset(shadow);
755 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
757 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
758 amdgpu_bo_size(shadow), NULL, fence,
763 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
764 * @bo: &amdgpu_bo buffer object to be mapped
765 * @ptr: kernel virtual address to be returned
767 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
768 * amdgpu_bo_kptr() to get the kernel virtual address.
771 * 0 for success or a negative error code on failure.
773 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
778 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
781 kptr = amdgpu_bo_kptr(bo);
788 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false, false,
789 MAX_SCHEDULE_TIMEOUT);
793 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
798 *ptr = amdgpu_bo_kptr(bo);
804 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
805 * @bo: &amdgpu_bo buffer object
807 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
810 * the virtual address of a buffer object area.
812 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
816 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
820 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
821 * @bo: &amdgpu_bo buffer object to be unmapped
823 * Unmaps a kernel map set up by amdgpu_bo_kmap().
825 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
828 ttm_bo_kunmap(&bo->kmap);
832 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
833 * @bo: &amdgpu_bo buffer object
835 * References the contained &ttm_buffer_object.
838 * a refcounted pointer to the &amdgpu_bo buffer object.
840 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
845 ttm_bo_get(&bo->tbo);
850 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
851 * @bo: &amdgpu_bo buffer object
853 * Unreferences the contained &ttm_buffer_object and clear the pointer
855 void amdgpu_bo_unref(struct amdgpu_bo **bo)
857 struct ttm_buffer_object *tbo;
868 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
869 * @bo: &amdgpu_bo buffer object to be pinned
870 * @domain: domain to be pinned to
871 * @min_offset: the start of requested address range
872 * @max_offset: the end of requested address range
874 * Pins the buffer object according to requested domain and address range. If
875 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
876 * pin_count and pin_size accordingly.
878 * Pinning means to lock pages in memory along with keeping them at a fixed
879 * offset. It is required when a buffer can not be moved, for example, when
880 * a display buffer is being scanned out.
882 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
883 * where to pin a buffer if there are specific restrictions on where a buffer
887 * 0 for success or a negative error code on failure.
889 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
890 u64 min_offset, u64 max_offset)
892 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
893 struct ttm_operation_ctx ctx = { false, false };
896 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
899 if (WARN_ON_ONCE(min_offset > max_offset))
902 /* A shared bo cannot be migrated to VRAM */
903 if (bo->prime_shared_count) {
904 if (domain & AMDGPU_GEM_DOMAIN_GTT)
905 domain = AMDGPU_GEM_DOMAIN_GTT;
910 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
911 * See function amdgpu_display_supported_domains()
913 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
916 uint32_t mem_type = bo->tbo.mem.mem_type;
918 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
923 if (max_offset != 0) {
924 u64 domain_start = amdgpu_ttm_domain_start(adev,
926 WARN_ON_ONCE(max_offset <
927 (amdgpu_bo_gpu_offset(bo) - domain_start));
933 if (bo->tbo.base.import_attach)
934 dma_buf_pin(bo->tbo.base.import_attach);
936 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
937 /* force to pin into visible video ram */
938 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
939 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
940 amdgpu_bo_placement_from_domain(bo, domain);
941 for (i = 0; i < bo->placement.num_placement; i++) {
944 fpfn = min_offset >> PAGE_SHIFT;
945 lpfn = max_offset >> PAGE_SHIFT;
947 if (fpfn > bo->placements[i].fpfn)
948 bo->placements[i].fpfn = fpfn;
949 if (!bo->placements[i].lpfn ||
950 (lpfn && lpfn < bo->placements[i].lpfn))
951 bo->placements[i].lpfn = lpfn;
952 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
955 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
957 dev_err(adev->dev, "%p pin failed\n", bo);
963 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
964 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
965 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
966 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
967 &adev->visible_pin_size);
968 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
969 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
977 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
978 * @bo: &amdgpu_bo buffer object to be pinned
979 * @domain: domain to be pinned to
981 * A simple wrapper to amdgpu_bo_pin_restricted().
982 * Provides a simpler API for buffers that do not have any strict restrictions
983 * on where a buffer must be located.
986 * 0 for success or a negative error code on failure.
988 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
990 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
994 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
995 * @bo: &amdgpu_bo buffer object to be unpinned
997 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
998 * Changes placement and pin size accordingly.
1001 * 0 for success or a negative error code on failure.
1003 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
1005 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1006 struct ttm_operation_ctx ctx = { false, false };
1009 if (WARN_ON_ONCE(!bo->pin_count)) {
1010 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
1017 amdgpu_bo_subtract_pin_size(bo);
1019 if (bo->tbo.base.import_attach)
1020 dma_buf_unpin(bo->tbo.base.import_attach);
1022 for (i = 0; i < bo->placement.num_placement; i++) {
1023 bo->placements[i].lpfn = 0;
1024 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
1026 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1028 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
1034 * amdgpu_bo_evict_vram - evict VRAM buffers
1035 * @adev: amdgpu device object
1037 * Evicts all VRAM buffers on the lru list of the memory type.
1038 * Mainly used for evicting vram at suspend time.
1041 * 0 for success or a negative error code on failure.
1043 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1045 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
1046 #ifndef CONFIG_HIBERNATION
1047 if (adev->flags & AMD_IS_APU) {
1048 /* Useless to evict on IGP chips */
1052 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
1055 static const char *amdgpu_vram_names[] = {
1069 * amdgpu_bo_init - initialize memory manager
1070 * @adev: amdgpu device object
1072 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1075 * 0 for success or a negative error code on failure.
1077 int amdgpu_bo_init(struct amdgpu_device *adev)
1079 /* reserve PAT memory space to WC for VRAM */
1080 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1081 adev->gmc.aper_size);
1083 /* Add an MTRR for the VRAM */
1084 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1085 adev->gmc.aper_size);
1086 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1087 adev->gmc.mc_vram_size >> 20,
1088 (unsigned long long)adev->gmc.aper_size >> 20);
1089 DRM_INFO("RAM width %dbits %s\n",
1090 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1091 return amdgpu_ttm_init(adev);
1095 * amdgpu_bo_late_init - late init
1096 * @adev: amdgpu device object
1098 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1102 * 0 for success or a negative error code on failure.
1104 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1106 amdgpu_ttm_late_init(adev);
1112 * amdgpu_bo_fini - tear down memory manager
1113 * @adev: amdgpu device object
1115 * Reverses amdgpu_bo_init() to tear down memory manager.
1117 void amdgpu_bo_fini(struct amdgpu_device *adev)
1119 amdgpu_ttm_fini(adev);
1120 arch_phys_wc_del(adev->gmc.vram_mtrr);
1121 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1125 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1126 * @bo: &amdgpu_bo buffer object
1127 * @vma: vma as input from the fbdev mmap method
1129 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1132 * 0 for success or a negative error code on failure.
1134 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1135 struct vm_area_struct *vma)
1137 if (vma->vm_pgoff != 0)
1140 return ttm_bo_mmap_obj(vma, &bo->tbo);
1144 * amdgpu_bo_set_tiling_flags - set tiling flags
1145 * @bo: &amdgpu_bo buffer object
1146 * @tiling_flags: new flags
1148 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1149 * kernel driver to set the tiling flags on a buffer.
1152 * 0 for success or a negative error code on failure.
1154 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1156 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1158 if (adev->family <= AMDGPU_FAMILY_CZ &&
1159 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1162 bo->tiling_flags = tiling_flags;
1167 * amdgpu_bo_get_tiling_flags - get tiling flags
1168 * @bo: &amdgpu_bo buffer object
1169 * @tiling_flags: returned flags
1171 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1172 * set the tiling flags on a buffer.
1174 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1176 dma_resv_assert_held(bo->tbo.base.resv);
1179 *tiling_flags = bo->tiling_flags;
1183 * amdgpu_bo_set_metadata - set metadata
1184 * @bo: &amdgpu_bo buffer object
1185 * @metadata: new metadata
1186 * @metadata_size: size of the new metadata
1187 * @flags: flags of the new metadata
1189 * Sets buffer object's metadata, its size and flags.
1190 * Used via GEM ioctl.
1193 * 0 for success or a negative error code on failure.
1195 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1196 uint32_t metadata_size, uint64_t flags)
1200 if (!metadata_size) {
1201 if (bo->metadata_size) {
1202 kfree(bo->metadata);
1203 bo->metadata = NULL;
1204 bo->metadata_size = 0;
1209 if (metadata == NULL)
1212 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1216 kfree(bo->metadata);
1217 bo->metadata_flags = flags;
1218 bo->metadata = buffer;
1219 bo->metadata_size = metadata_size;
1225 * amdgpu_bo_get_metadata - get metadata
1226 * @bo: &amdgpu_bo buffer object
1227 * @buffer: returned metadata
1228 * @buffer_size: size of the buffer
1229 * @metadata_size: size of the returned metadata
1230 * @flags: flags of the returned metadata
1232 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1233 * less than metadata_size.
1234 * Used via GEM ioctl.
1237 * 0 for success or a negative error code on failure.
1239 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1240 size_t buffer_size, uint32_t *metadata_size,
1243 if (!buffer && !metadata_size)
1247 if (buffer_size < bo->metadata_size)
1250 if (bo->metadata_size)
1251 memcpy(buffer, bo->metadata, bo->metadata_size);
1255 *metadata_size = bo->metadata_size;
1257 *flags = bo->metadata_flags;
1263 * amdgpu_bo_move_notify - notification about a memory move
1264 * @bo: pointer to a buffer object
1265 * @evict: if this move is evicting the buffer from the graphics address space
1266 * @new_mem: new information of the bufer object
1268 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1270 * TTM driver callback which is called when ttm moves a buffer.
1272 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1274 struct ttm_resource *new_mem)
1276 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1277 struct amdgpu_bo *abo;
1278 struct ttm_resource *old_mem = &bo->mem;
1280 if (!amdgpu_bo_is_amdgpu_bo(bo))
1283 abo = ttm_to_amdgpu_bo(bo);
1284 amdgpu_vm_bo_invalidate(adev, abo, evict);
1286 amdgpu_bo_kunmap(abo);
1288 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1289 bo->mem.mem_type != TTM_PL_SYSTEM)
1290 dma_buf_move_notify(abo->tbo.base.dma_buf);
1292 /* remember the eviction */
1294 atomic64_inc(&adev->num_evictions);
1296 /* update statistics */
1300 /* move_notify is called before move happens */
1301 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1305 * amdgpu_bo_release_notify - notification about a BO being released
1306 * @bo: pointer to a buffer object
1308 * Wipes VRAM buffers whose contents should not be leaked before the
1309 * memory is released.
1311 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1313 struct dma_fence *fence = NULL;
1314 struct amdgpu_bo *abo;
1317 if (!amdgpu_bo_is_amdgpu_bo(bo))
1320 abo = ttm_to_amdgpu_bo(bo);
1323 amdgpu_amdkfd_unreserve_memory_limit(abo);
1325 /* We only remove the fence if the resv has individualized. */
1326 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1327 && bo->base.resv != &bo->base._resv);
1328 if (bo->base.resv == &bo->base._resv)
1329 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1331 if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
1332 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1335 dma_resv_lock(bo->base.resv, NULL);
1337 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1339 amdgpu_bo_fence(abo, fence, false);
1340 dma_fence_put(fence);
1343 dma_resv_unlock(bo->base.resv);
1347 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1348 * @bo: pointer to a buffer object
1350 * Notifies the driver we are taking a fault on this BO and have reserved it,
1351 * also performs bookkeeping.
1352 * TTM driver callback for dealing with vm faults.
1355 * 0 for success or a negative error code on failure.
1357 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1359 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1360 struct ttm_operation_ctx ctx = { false, false };
1361 struct amdgpu_bo *abo;
1362 unsigned long offset, size;
1365 if (!amdgpu_bo_is_amdgpu_bo(bo))
1368 abo = ttm_to_amdgpu_bo(bo);
1370 /* Remember that this BO was accessed by the CPU */
1371 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1373 if (bo->mem.mem_type != TTM_PL_VRAM)
1376 size = bo->mem.num_pages << PAGE_SHIFT;
1377 offset = bo->mem.start << PAGE_SHIFT;
1378 if ((offset + size) <= adev->gmc.visible_vram_size)
1381 /* Can't move a pinned BO to visible VRAM */
1382 if (abo->pin_count > 0)
1385 /* hurrah the memory is not visible ! */
1386 atomic64_inc(&adev->num_vram_cpu_page_faults);
1387 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1388 AMDGPU_GEM_DOMAIN_GTT);
1390 /* Avoid costly evictions; only set GTT as a busy placement */
1391 abo->placement.num_busy_placement = 1;
1392 abo->placement.busy_placement = &abo->placements[1];
1394 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1395 if (unlikely(r != 0))
1398 offset = bo->mem.start << PAGE_SHIFT;
1399 /* this should never happen */
1400 if (bo->mem.mem_type == TTM_PL_VRAM &&
1401 (offset + size) > adev->gmc.visible_vram_size)
1408 * amdgpu_bo_fence - add fence to buffer object
1410 * @bo: buffer object in question
1411 * @fence: fence to add
1412 * @shared: true if fence should be added shared
1415 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1418 struct dma_resv *resv = bo->tbo.base.resv;
1421 dma_resv_add_shared_fence(resv, fence);
1423 dma_resv_add_excl_fence(resv, fence);
1427 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1429 * @adev: amdgpu device pointer
1430 * @resv: reservation object to sync to
1431 * @sync_mode: synchronization mode
1432 * @owner: fence owner
1433 * @intr: Whether the wait is interruptible
1435 * Extract the fences from the reservation object and waits for them to finish.
1438 * 0 on success, errno otherwise.
1440 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1441 enum amdgpu_sync_mode sync_mode, void *owner,
1444 struct amdgpu_sync sync;
1447 amdgpu_sync_create(&sync);
1448 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1449 r = amdgpu_sync_wait(&sync, intr);
1450 amdgpu_sync_free(&sync);
1455 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1456 * @bo: buffer object to wait for
1457 * @owner: fence owner
1458 * @intr: Whether the wait is interruptible
1460 * Wrapper to wait for fences in a BO.
1462 * 0 on success, errno otherwise.
1464 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1466 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1468 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1469 AMDGPU_SYNC_NE_OWNER, owner, intr);
1473 * amdgpu_bo_gpu_offset - return GPU offset of bo
1474 * @bo: amdgpu object for which we query the offset
1476 * Note: object should either be pinned or reserved when calling this
1477 * function, it might be useful to add check for this for debugging.
1480 * current GPU offset of the object.
1482 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1484 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1485 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1486 !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1487 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1488 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1489 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1491 return amdgpu_bo_gpu_offset_no_check(bo);
1495 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1496 * @bo: amdgpu object for which we query the offset
1499 * current GPU offset of the object without raising warnings.
1501 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1503 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1506 offset = (bo->tbo.mem.start << PAGE_SHIFT) +
1507 amdgpu_ttm_domain_start(adev, bo->tbo.mem.mem_type);
1509 return amdgpu_gmc_sign_extend(offset);
1513 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1514 * @adev: amdgpu device object
1515 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1518 * Which of the allowed domains is preferred for pinning the BO for scanout.
1520 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1523 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1524 domain = AMDGPU_GEM_DOMAIN_VRAM;
1525 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1526 domain = AMDGPU_GEM_DOMAIN_GTT;