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>
35 #include <drm/amdgpu_drm.h>
36 #include <drm/drm_cache.h>
38 #include "amdgpu_trace.h"
39 #include "amdgpu_amdkfd.h"
44 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
45 * represents memory used by driver (VRAM, system memory, etc.). The driver
46 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
47 * to create/destroy/set buffer object which are then managed by the kernel TTM
49 * The interfaces are also used internally by kernel clients, including gfx,
50 * uvd, etc. for kernel managed allocations used by the GPU.
54 static bool amdgpu_bo_need_backup(struct amdgpu_device *adev)
56 if (adev->flags & AMD_IS_APU)
59 if (amdgpu_gpu_recovery == 0 ||
60 (amdgpu_gpu_recovery == -1 && !amdgpu_sriov_vf(adev)))
67 * amdgpu_bo_subtract_pin_size - Remove BO from pin_size accounting
69 * @bo: &amdgpu_bo buffer object
71 * This function is called when a BO stops being pinned, and updates the
72 * &amdgpu_device pin_size values accordingly.
74 static void amdgpu_bo_subtract_pin_size(struct amdgpu_bo *bo)
76 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
78 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
79 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
80 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
81 &adev->visible_pin_size);
82 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
83 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
87 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
89 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
90 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
92 if (bo->pin_count > 0)
93 amdgpu_bo_subtract_pin_size(bo);
96 amdgpu_amdkfd_unreserve_system_memory_limit(bo);
100 if (bo->gem_base.import_attach)
101 drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
102 drm_gem_object_release(&bo->gem_base);
103 amdgpu_bo_unref(&bo->parent);
104 if (!list_empty(&bo->shadow_list)) {
105 mutex_lock(&adev->shadow_list_lock);
106 list_del_init(&bo->shadow_list);
107 mutex_unlock(&adev->shadow_list_lock);
114 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
115 * @bo: buffer object to be checked
117 * Uses destroy function associated with the object to determine if this is
121 * true if the object belongs to &amdgpu_bo, false if not.
123 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
125 if (bo->destroy == &amdgpu_bo_destroy)
131 * amdgpu_bo_placement_from_domain - set buffer's placement
132 * @abo: &amdgpu_bo buffer object whose placement is to be set
133 * @domain: requested domain
135 * Sets buffer's placement according to requested domain and the buffer's
138 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
140 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
141 struct ttm_placement *placement = &abo->placement;
142 struct ttm_place *places = abo->placements;
143 u64 flags = abo->flags;
146 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
147 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
151 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
154 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
155 places[c].lpfn = visible_pfn;
157 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
159 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
160 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
164 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
166 if (flags & AMDGPU_GEM_CREATE_SHADOW)
167 places[c].lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
170 places[c].flags = TTM_PL_FLAG_TT;
171 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
172 places[c].flags |= TTM_PL_FLAG_WC |
173 TTM_PL_FLAG_UNCACHED;
175 places[c].flags |= TTM_PL_FLAG_CACHED;
179 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
182 places[c].flags = TTM_PL_FLAG_SYSTEM;
183 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
184 places[c].flags |= TTM_PL_FLAG_WC |
185 TTM_PL_FLAG_UNCACHED;
187 places[c].flags |= TTM_PL_FLAG_CACHED;
191 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
194 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
198 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
201 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
205 if (domain & AMDGPU_GEM_DOMAIN_OA) {
208 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
215 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
219 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
221 placement->num_placement = c;
222 placement->placement = places;
224 placement->num_busy_placement = c;
225 placement->busy_placement = places;
229 * amdgpu_bo_create_reserved - create reserved BO for kernel use
231 * @adev: amdgpu device object
232 * @size: size for the new BO
233 * @align: alignment for the new BO
234 * @domain: where to place it
235 * @bo_ptr: used to initialize BOs in structures
236 * @gpu_addr: GPU addr of the pinned BO
237 * @cpu_addr: optional CPU address mapping
239 * Allocates and pins a BO for kernel internal use, and returns it still
242 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
245 * 0 on success, negative error code otherwise.
247 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
248 unsigned long size, int align,
249 u32 domain, struct amdgpu_bo **bo_ptr,
250 u64 *gpu_addr, void **cpu_addr)
252 struct amdgpu_bo_param bp;
256 memset(&bp, 0, sizeof(bp));
258 bp.byte_align = align;
260 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
261 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
262 bp.type = ttm_bo_type_kernel;
266 r = amdgpu_bo_create(adev, &bp, bo_ptr);
268 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
275 r = amdgpu_bo_reserve(*bo_ptr, false);
277 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
281 r = amdgpu_bo_pin(*bo_ptr, domain);
283 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
284 goto error_unreserve;
287 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
289 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
294 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
297 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
299 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
307 amdgpu_bo_unpin(*bo_ptr);
309 amdgpu_bo_unreserve(*bo_ptr);
313 amdgpu_bo_unref(bo_ptr);
319 * amdgpu_bo_create_kernel - create BO for kernel use
321 * @adev: amdgpu device object
322 * @size: size for the new BO
323 * @align: alignment for the new BO
324 * @domain: where to place it
325 * @bo_ptr: used to initialize BOs in structures
326 * @gpu_addr: GPU addr of the pinned BO
327 * @cpu_addr: optional CPU address mapping
329 * Allocates and pins a BO for kernel internal use.
331 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
334 * 0 on success, negative error code otherwise.
336 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
337 unsigned long size, int align,
338 u32 domain, struct amdgpu_bo **bo_ptr,
339 u64 *gpu_addr, void **cpu_addr)
343 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
349 amdgpu_bo_unreserve(*bo_ptr);
355 * amdgpu_bo_free_kernel - free BO for kernel use
357 * @bo: amdgpu BO to free
358 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
359 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
361 * unmaps and unpin a BO for kernel internal use.
363 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
369 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
371 amdgpu_bo_kunmap(*bo);
373 amdgpu_bo_unpin(*bo);
374 amdgpu_bo_unreserve(*bo);
385 /* Validate bo size is bit bigger then the request domain */
386 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
387 unsigned long size, u32 domain)
389 struct ttm_mem_type_manager *man = NULL;
392 * If GTT is part of requested domains the check must succeed to
393 * allow fall back to GTT
395 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
396 man = &adev->mman.bdev.man[TTM_PL_TT];
398 if (size < (man->size << PAGE_SHIFT))
404 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
405 man = &adev->mman.bdev.man[TTM_PL_VRAM];
407 if (size < (man->size << PAGE_SHIFT))
414 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
418 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
419 man->size << PAGE_SHIFT);
423 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
424 struct amdgpu_bo_param *bp,
425 struct amdgpu_bo **bo_ptr)
427 struct ttm_operation_ctx ctx = {
428 .interruptible = (bp->type != ttm_bo_type_kernel),
429 .no_wait_gpu = false,
431 .flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
433 struct amdgpu_bo *bo;
434 unsigned long page_align, size = bp->size;
438 page_align = roundup(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
439 size = ALIGN(size, PAGE_SIZE);
441 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
446 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
447 sizeof(struct amdgpu_bo));
449 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
452 drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
453 INIT_LIST_HEAD(&bo->shadow_list);
454 INIT_LIST_HEAD(&bo->va);
455 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
457 bo->allowed_domains = bo->preferred_domains;
458 if (bp->type != ttm_bo_type_kernel &&
459 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
460 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
462 bo->flags = bp->flags;
465 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
466 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
468 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
469 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
470 /* Don't try to enable write-combining when it can't work, or things
472 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
475 #ifndef CONFIG_COMPILE_TEST
476 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
477 thanks to write-combining
480 if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
481 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
482 "better performance thanks to write-combining\n");
483 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
485 /* For architectures that don't support WC memory,
486 * mask out the WC flag from the BO
488 if (!drm_arch_can_wc_memory())
489 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
492 bo->tbo.bdev = &adev->mman.bdev;
493 amdgpu_bo_placement_from_domain(bo, bp->domain);
494 if (bp->type == ttm_bo_type_kernel)
495 bo->tbo.priority = 1;
497 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
498 &bo->placement, page_align, &ctx, acc_size,
499 NULL, bp->resv, &amdgpu_bo_destroy);
500 if (unlikely(r != 0))
503 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
504 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
505 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
506 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
509 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
511 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
512 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
513 struct dma_fence *fence;
515 r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
519 amdgpu_bo_fence(bo, fence, false);
520 dma_fence_put(bo->tbo.moving);
521 bo->tbo.moving = dma_fence_get(fence);
522 dma_fence_put(fence);
525 amdgpu_bo_unreserve(bo);
528 trace_amdgpu_bo_create(bo);
530 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
531 if (bp->type == ttm_bo_type_device)
532 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
538 ww_mutex_unlock(&bo->tbo.resv->lock);
539 amdgpu_bo_unref(&bo);
543 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
544 unsigned long size, int byte_align,
545 struct amdgpu_bo *bo)
547 struct amdgpu_bo_param bp;
553 memset(&bp, 0, sizeof(bp));
555 bp.byte_align = byte_align;
556 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
557 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
558 AMDGPU_GEM_CREATE_SHADOW;
559 bp.type = ttm_bo_type_kernel;
560 bp.resv = bo->tbo.resv;
562 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
564 bo->shadow->parent = amdgpu_bo_ref(bo);
565 mutex_lock(&adev->shadow_list_lock);
566 list_add_tail(&bo->shadow_list, &adev->shadow_list);
567 mutex_unlock(&adev->shadow_list_lock);
574 * amdgpu_bo_create - create an &amdgpu_bo buffer object
575 * @adev: amdgpu device object
576 * @bp: parameters to be used for the buffer object
577 * @bo_ptr: pointer to the buffer object pointer
579 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
581 * Shadow object is used to backup the original buffer object, and is always
585 * 0 for success or a negative error code on failure.
587 int amdgpu_bo_create(struct amdgpu_device *adev,
588 struct amdgpu_bo_param *bp,
589 struct amdgpu_bo **bo_ptr)
591 u64 flags = bp->flags;
594 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
595 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
599 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && amdgpu_bo_need_backup(adev)) {
601 WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
604 r = amdgpu_bo_create_shadow(adev, bp->size, bp->byte_align, (*bo_ptr));
607 reservation_object_unlock((*bo_ptr)->tbo.resv);
610 amdgpu_bo_unref(bo_ptr);
617 * amdgpu_bo_backup_to_shadow - Backs up an &amdgpu_bo buffer object
618 * @adev: amdgpu device object
619 * @ring: amdgpu_ring for the engine handling the buffer operations
620 * @bo: &amdgpu_bo buffer to be backed up
621 * @resv: reservation object with embedded fence
622 * @fence: dma_fence associated with the operation
623 * @direct: whether to submit the job directly
625 * Copies an &amdgpu_bo buffer object to its shadow object.
629 * 0 for success or a negative error code on failure.
631 int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
632 struct amdgpu_ring *ring,
633 struct amdgpu_bo *bo,
634 struct reservation_object *resv,
635 struct dma_fence **fence,
639 struct amdgpu_bo *shadow = bo->shadow;
640 uint64_t bo_addr, shadow_addr;
646 bo_addr = amdgpu_bo_gpu_offset(bo);
647 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
649 r = reservation_object_reserve_shared(bo->tbo.resv);
653 r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
654 amdgpu_bo_size(bo), resv, fence,
657 amdgpu_bo_fence(bo, *fence, true);
664 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
665 * @bo: pointer to the buffer object
667 * Sets placement according to domain; and changes placement and caching
668 * policy of the buffer object according to the placement.
669 * This is used for validating shadow bos. It calls ttm_bo_validate() to
670 * make sure the buffer is resident where it needs to be.
673 * 0 for success or a negative error code on failure.
675 int amdgpu_bo_validate(struct amdgpu_bo *bo)
677 struct ttm_operation_ctx ctx = { false, false };
684 domain = bo->preferred_domains;
687 amdgpu_bo_placement_from_domain(bo, domain);
688 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
689 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
690 domain = bo->allowed_domains;
698 * amdgpu_bo_restore_from_shadow - restore an &amdgpu_bo buffer object
699 * @adev: amdgpu device object
700 * @ring: amdgpu_ring for the engine handling the buffer operations
701 * @bo: &amdgpu_bo buffer to be restored
702 * @resv: reservation object with embedded fence
703 * @fence: dma_fence associated with the operation
704 * @direct: whether to submit the job directly
706 * Copies a buffer object's shadow content back to the object.
707 * This is used for recovering a buffer from its shadow in case of a gpu
708 * reset where vram context may be lost.
711 * 0 for success or a negative error code on failure.
713 int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
714 struct amdgpu_ring *ring,
715 struct amdgpu_bo *bo,
716 struct reservation_object *resv,
717 struct dma_fence **fence,
721 struct amdgpu_bo *shadow = bo->shadow;
722 uint64_t bo_addr, shadow_addr;
728 bo_addr = amdgpu_bo_gpu_offset(bo);
729 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
731 r = reservation_object_reserve_shared(bo->tbo.resv);
735 r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
736 amdgpu_bo_size(bo), resv, fence,
739 amdgpu_bo_fence(bo, *fence, true);
746 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
747 * @bo: &amdgpu_bo buffer object to be mapped
748 * @ptr: kernel virtual address to be returned
750 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
751 * amdgpu_bo_kptr() to get the kernel virtual address.
754 * 0 for success or a negative error code on failure.
756 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
761 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
764 kptr = amdgpu_bo_kptr(bo);
771 r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
772 MAX_SCHEDULE_TIMEOUT);
776 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
781 *ptr = amdgpu_bo_kptr(bo);
787 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
788 * @bo: &amdgpu_bo buffer object
790 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
793 * the virtual address of a buffer object area.
795 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
799 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
803 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
804 * @bo: &amdgpu_bo buffer object to be unmapped
806 * Unmaps a kernel map set up by amdgpu_bo_kmap().
808 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
811 ttm_bo_kunmap(&bo->kmap);
815 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
816 * @bo: &amdgpu_bo buffer object
818 * References the contained &ttm_buffer_object.
821 * a refcounted pointer to the &amdgpu_bo buffer object.
823 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
828 ttm_bo_get(&bo->tbo);
833 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
834 * @bo: &amdgpu_bo buffer object
836 * Unreferences the contained &ttm_buffer_object and clear the pointer
838 void amdgpu_bo_unref(struct amdgpu_bo **bo)
840 struct ttm_buffer_object *tbo;
851 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
852 * @bo: &amdgpu_bo buffer object to be pinned
853 * @domain: domain to be pinned to
854 * @min_offset: the start of requested address range
855 * @max_offset: the end of requested address range
857 * Pins the buffer object according to requested domain and address range. If
858 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
859 * pin_count and pin_size accordingly.
861 * Pinning means to lock pages in memory along with keeping them at a fixed
862 * offset. It is required when a buffer can not be moved, for example, when
863 * a display buffer is being scanned out.
865 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
866 * where to pin a buffer if there are specific restrictions on where a buffer
870 * 0 for success or a negative error code on failure.
872 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
873 u64 min_offset, u64 max_offset)
875 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
876 struct ttm_operation_ctx ctx = { false, false };
879 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
882 if (WARN_ON_ONCE(min_offset > max_offset))
885 /* A shared bo cannot be migrated to VRAM */
886 if (bo->prime_shared_count) {
887 if (domain & AMDGPU_GEM_DOMAIN_GTT)
888 domain = AMDGPU_GEM_DOMAIN_GTT;
893 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
894 * See function amdgpu_display_supported_domains()
896 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
899 uint32_t mem_type = bo->tbo.mem.mem_type;
901 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
906 if (max_offset != 0) {
907 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
908 WARN_ON_ONCE(max_offset <
909 (amdgpu_bo_gpu_offset(bo) - domain_start));
915 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
916 /* force to pin into visible video ram */
917 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
918 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
919 amdgpu_bo_placement_from_domain(bo, domain);
920 for (i = 0; i < bo->placement.num_placement; i++) {
923 fpfn = min_offset >> PAGE_SHIFT;
924 lpfn = max_offset >> PAGE_SHIFT;
926 if (fpfn > bo->placements[i].fpfn)
927 bo->placements[i].fpfn = fpfn;
928 if (!bo->placements[i].lpfn ||
929 (lpfn && lpfn < bo->placements[i].lpfn))
930 bo->placements[i].lpfn = lpfn;
931 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
934 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
936 dev_err(adev->dev, "%p pin failed\n", bo);
942 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
943 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
944 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
945 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
946 &adev->visible_pin_size);
947 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
948 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
956 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
957 * @bo: &amdgpu_bo buffer object to be pinned
958 * @domain: domain to be pinned to
960 * A simple wrapper to amdgpu_bo_pin_restricted().
961 * Provides a simpler API for buffers that do not have any strict restrictions
962 * on where a buffer must be located.
965 * 0 for success or a negative error code on failure.
967 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
969 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
973 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
974 * @bo: &amdgpu_bo buffer object to be unpinned
976 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
977 * Changes placement and pin size accordingly.
980 * 0 for success or a negative error code on failure.
982 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
984 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
985 struct ttm_operation_ctx ctx = { false, false };
988 if (!bo->pin_count) {
989 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
996 amdgpu_bo_subtract_pin_size(bo);
998 for (i = 0; i < bo->placement.num_placement; i++) {
999 bo->placements[i].lpfn = 0;
1000 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
1002 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1004 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
1010 * amdgpu_bo_evict_vram - evict VRAM buffers
1011 * @adev: amdgpu device object
1013 * Evicts all VRAM buffers on the lru list of the memory type.
1014 * Mainly used for evicting vram at suspend time.
1017 * 0 for success or a negative error code on failure.
1019 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1021 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
1022 if (0 && (adev->flags & AMD_IS_APU)) {
1023 /* Useless to evict on IGP chips */
1026 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
1029 static const char *amdgpu_vram_names[] = {
1042 * amdgpu_bo_init - initialize memory manager
1043 * @adev: amdgpu device object
1045 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1048 * 0 for success or a negative error code on failure.
1050 int amdgpu_bo_init(struct amdgpu_device *adev)
1052 /* reserve PAT memory space to WC for VRAM */
1053 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1054 adev->gmc.aper_size);
1056 /* Add an MTRR for the VRAM */
1057 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1058 adev->gmc.aper_size);
1059 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1060 adev->gmc.mc_vram_size >> 20,
1061 (unsigned long long)adev->gmc.aper_size >> 20);
1062 DRM_INFO("RAM width %dbits %s\n",
1063 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1064 return amdgpu_ttm_init(adev);
1068 * amdgpu_bo_late_init - late init
1069 * @adev: amdgpu device object
1071 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1075 * 0 for success or a negative error code on failure.
1077 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1079 amdgpu_ttm_late_init(adev);
1085 * amdgpu_bo_fini - tear down memory manager
1086 * @adev: amdgpu device object
1088 * Reverses amdgpu_bo_init() to tear down memory manager.
1090 void amdgpu_bo_fini(struct amdgpu_device *adev)
1092 amdgpu_ttm_fini(adev);
1093 arch_phys_wc_del(adev->gmc.vram_mtrr);
1094 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1098 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1099 * @bo: &amdgpu_bo buffer object
1100 * @vma: vma as input from the fbdev mmap method
1102 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1105 * 0 for success or a negative error code on failure.
1107 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1108 struct vm_area_struct *vma)
1110 return ttm_fbdev_mmap(vma, &bo->tbo);
1114 * amdgpu_bo_set_tiling_flags - set tiling flags
1115 * @bo: &amdgpu_bo buffer object
1116 * @tiling_flags: new flags
1118 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1119 * kernel driver to set the tiling flags on a buffer.
1122 * 0 for success or a negative error code on failure.
1124 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1126 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1128 if (adev->family <= AMDGPU_FAMILY_CZ &&
1129 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1132 bo->tiling_flags = tiling_flags;
1137 * amdgpu_bo_get_tiling_flags - get tiling flags
1138 * @bo: &amdgpu_bo buffer object
1139 * @tiling_flags: returned flags
1141 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1142 * set the tiling flags on a buffer.
1144 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1146 lockdep_assert_held(&bo->tbo.resv->lock.base);
1149 *tiling_flags = bo->tiling_flags;
1153 * amdgpu_bo_set_metadata - set metadata
1154 * @bo: &amdgpu_bo buffer object
1155 * @metadata: new metadata
1156 * @metadata_size: size of the new metadata
1157 * @flags: flags of the new metadata
1159 * Sets buffer object's metadata, its size and flags.
1160 * Used via GEM ioctl.
1163 * 0 for success or a negative error code on failure.
1165 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1166 uint32_t metadata_size, uint64_t flags)
1170 if (!metadata_size) {
1171 if (bo->metadata_size) {
1172 kfree(bo->metadata);
1173 bo->metadata = NULL;
1174 bo->metadata_size = 0;
1179 if (metadata == NULL)
1182 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1186 kfree(bo->metadata);
1187 bo->metadata_flags = flags;
1188 bo->metadata = buffer;
1189 bo->metadata_size = metadata_size;
1195 * amdgpu_bo_get_metadata - get metadata
1196 * @bo: &amdgpu_bo buffer object
1197 * @buffer: returned metadata
1198 * @buffer_size: size of the buffer
1199 * @metadata_size: size of the returned metadata
1200 * @flags: flags of the returned metadata
1202 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1203 * less than metadata_size.
1204 * Used via GEM ioctl.
1207 * 0 for success or a negative error code on failure.
1209 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1210 size_t buffer_size, uint32_t *metadata_size,
1213 if (!buffer && !metadata_size)
1217 if (buffer_size < bo->metadata_size)
1220 if (bo->metadata_size)
1221 memcpy(buffer, bo->metadata, bo->metadata_size);
1225 *metadata_size = bo->metadata_size;
1227 *flags = bo->metadata_flags;
1233 * amdgpu_bo_move_notify - notification about a memory move
1234 * @bo: pointer to a buffer object
1235 * @evict: if this move is evicting the buffer from the graphics address space
1236 * @new_mem: new information of the bufer object
1238 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1240 * TTM driver callback which is called when ttm moves a buffer.
1242 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1244 struct ttm_mem_reg *new_mem)
1246 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1247 struct amdgpu_bo *abo;
1248 struct ttm_mem_reg *old_mem = &bo->mem;
1250 if (!amdgpu_bo_is_amdgpu_bo(bo))
1253 abo = ttm_to_amdgpu_bo(bo);
1254 amdgpu_vm_bo_invalidate(adev, abo, evict);
1256 amdgpu_bo_kunmap(abo);
1258 /* remember the eviction */
1260 atomic64_inc(&adev->num_evictions);
1262 /* update statistics */
1266 /* move_notify is called before move happens */
1267 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1271 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1272 * @bo: pointer to a buffer object
1274 * Notifies the driver we are taking a fault on this BO and have reserved it,
1275 * also performs bookkeeping.
1276 * TTM driver callback for dealing with vm faults.
1279 * 0 for success or a negative error code on failure.
1281 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1283 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1284 struct ttm_operation_ctx ctx = { false, false };
1285 struct amdgpu_bo *abo;
1286 unsigned long offset, size;
1289 if (!amdgpu_bo_is_amdgpu_bo(bo))
1292 abo = ttm_to_amdgpu_bo(bo);
1294 /* Remember that this BO was accessed by the CPU */
1295 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1297 if (bo->mem.mem_type != TTM_PL_VRAM)
1300 size = bo->mem.num_pages << PAGE_SHIFT;
1301 offset = bo->mem.start << PAGE_SHIFT;
1302 if ((offset + size) <= adev->gmc.visible_vram_size)
1305 /* Can't move a pinned BO to visible VRAM */
1306 if (abo->pin_count > 0)
1309 /* hurrah the memory is not visible ! */
1310 atomic64_inc(&adev->num_vram_cpu_page_faults);
1311 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1312 AMDGPU_GEM_DOMAIN_GTT);
1314 /* Avoid costly evictions; only set GTT as a busy placement */
1315 abo->placement.num_busy_placement = 1;
1316 abo->placement.busy_placement = &abo->placements[1];
1318 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1319 if (unlikely(r != 0))
1322 offset = bo->mem.start << PAGE_SHIFT;
1323 /* this should never happen */
1324 if (bo->mem.mem_type == TTM_PL_VRAM &&
1325 (offset + size) > adev->gmc.visible_vram_size)
1332 * amdgpu_bo_fence - add fence to buffer object
1334 * @bo: buffer object in question
1335 * @fence: fence to add
1336 * @shared: true if fence should be added shared
1339 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1342 struct reservation_object *resv = bo->tbo.resv;
1345 reservation_object_add_shared_fence(resv, fence);
1347 reservation_object_add_excl_fence(resv, fence);
1351 * amdgpu_bo_gpu_offset - return GPU offset of bo
1352 * @bo: amdgpu object for which we query the offset
1354 * Note: object should either be pinned or reserved when calling this
1355 * function, it might be useful to add check for this for debugging.
1358 * current GPU offset of the object.
1360 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1362 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1363 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_TT &&
1364 !amdgpu_gtt_mgr_has_gart_addr(&bo->tbo.mem));
1365 WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1367 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1368 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1369 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1371 return bo->tbo.offset;
1375 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1376 * @adev: amdgpu device object
1377 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1380 * Which of the allowed domains is preferred for pinning the BO for scanout.
1382 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1385 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1386 domain = AMDGPU_GEM_DOMAIN_VRAM;
1387 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1388 domain = AMDGPU_GEM_DOMAIN_GTT;