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) {
167 places[c].flags = TTM_PL_FLAG_TT;
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_CPU) {
179 places[c].flags = TTM_PL_FLAG_SYSTEM;
180 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
181 places[c].flags |= TTM_PL_FLAG_WC |
182 TTM_PL_FLAG_UNCACHED;
184 places[c].flags |= TTM_PL_FLAG_CACHED;
188 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
191 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
195 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
198 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
202 if (domain & AMDGPU_GEM_DOMAIN_OA) {
205 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
212 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
216 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
218 placement->num_placement = c;
219 placement->placement = places;
221 placement->num_busy_placement = c;
222 placement->busy_placement = places;
226 * amdgpu_bo_create_reserved - create reserved BO for kernel use
228 * @adev: amdgpu device object
229 * @size: size for the new BO
230 * @align: alignment for the new BO
231 * @domain: where to place it
232 * @bo_ptr: used to initialize BOs in structures
233 * @gpu_addr: GPU addr of the pinned BO
234 * @cpu_addr: optional CPU address mapping
236 * Allocates and pins a BO for kernel internal use, and returns it still
239 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
242 * 0 on success, negative error code otherwise.
244 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
245 unsigned long size, int align,
246 u32 domain, struct amdgpu_bo **bo_ptr,
247 u64 *gpu_addr, void **cpu_addr)
249 struct amdgpu_bo_param bp;
254 amdgpu_bo_unref(bo_ptr);
258 memset(&bp, 0, sizeof(bp));
260 bp.byte_align = align;
262 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
263 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
264 bp.type = ttm_bo_type_kernel;
268 r = amdgpu_bo_create(adev, &bp, bo_ptr);
270 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
277 r = amdgpu_bo_reserve(*bo_ptr, false);
279 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
283 r = amdgpu_bo_pin(*bo_ptr, domain);
285 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
286 goto error_unreserve;
289 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
291 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
296 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
299 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
301 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
309 amdgpu_bo_unpin(*bo_ptr);
311 amdgpu_bo_unreserve(*bo_ptr);
315 amdgpu_bo_unref(bo_ptr);
321 * amdgpu_bo_create_kernel - create BO for kernel use
323 * @adev: amdgpu device object
324 * @size: size for the new BO
325 * @align: alignment for the new BO
326 * @domain: where to place it
327 * @bo_ptr: used to initialize BOs in structures
328 * @gpu_addr: GPU addr of the pinned BO
329 * @cpu_addr: optional CPU address mapping
331 * Allocates and pins a BO for kernel internal use.
333 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
336 * 0 on success, negative error code otherwise.
338 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
339 unsigned long size, int align,
340 u32 domain, struct amdgpu_bo **bo_ptr,
341 u64 *gpu_addr, void **cpu_addr)
345 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
352 amdgpu_bo_unreserve(*bo_ptr);
358 * amdgpu_bo_free_kernel - free BO for kernel use
360 * @bo: amdgpu BO to free
361 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
362 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
364 * unmaps and unpin a BO for kernel internal use.
366 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
372 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
374 amdgpu_bo_kunmap(*bo);
376 amdgpu_bo_unpin(*bo);
377 amdgpu_bo_unreserve(*bo);
388 /* Validate bo size is bit bigger then the request domain */
389 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
390 unsigned long size, u32 domain)
392 struct ttm_mem_type_manager *man = NULL;
395 * If GTT is part of requested domains the check must succeed to
396 * allow fall back to GTT
398 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
399 man = &adev->mman.bdev.man[TTM_PL_TT];
401 if (size < (man->size << PAGE_SHIFT))
407 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
408 man = &adev->mman.bdev.man[TTM_PL_VRAM];
410 if (size < (man->size << PAGE_SHIFT))
417 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
421 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
422 man->size << PAGE_SHIFT);
426 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
427 struct amdgpu_bo_param *bp,
428 struct amdgpu_bo **bo_ptr)
430 struct ttm_operation_ctx ctx = {
431 .interruptible = (bp->type != ttm_bo_type_kernel),
432 .no_wait_gpu = false,
434 .flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
436 struct amdgpu_bo *bo;
437 unsigned long page_align, size = bp->size;
441 page_align = roundup(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
442 size = ALIGN(size, PAGE_SIZE);
444 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
449 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
450 sizeof(struct amdgpu_bo));
452 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
455 drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
456 INIT_LIST_HEAD(&bo->shadow_list);
458 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
460 bo->allowed_domains = bo->preferred_domains;
461 if (bp->type != ttm_bo_type_kernel &&
462 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
463 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
465 bo->flags = bp->flags;
468 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
469 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
471 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
472 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
473 /* Don't try to enable write-combining when it can't work, or things
475 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
478 #ifndef CONFIG_COMPILE_TEST
479 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
480 thanks to write-combining
483 if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
484 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
485 "better performance thanks to write-combining\n");
486 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
488 /* For architectures that don't support WC memory,
489 * mask out the WC flag from the BO
491 if (!drm_arch_can_wc_memory())
492 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
495 bo->tbo.bdev = &adev->mman.bdev;
496 amdgpu_bo_placement_from_domain(bo, bp->domain);
497 if (bp->type == ttm_bo_type_kernel)
498 bo->tbo.priority = 1;
500 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
501 &bo->placement, page_align, &ctx, acc_size,
502 NULL, bp->resv, &amdgpu_bo_destroy);
503 if (unlikely(r != 0))
506 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
507 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
508 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
509 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
512 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
514 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
515 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
516 struct dma_fence *fence;
518 r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
522 amdgpu_bo_fence(bo, fence, false);
523 dma_fence_put(bo->tbo.moving);
524 bo->tbo.moving = dma_fence_get(fence);
525 dma_fence_put(fence);
528 amdgpu_bo_unreserve(bo);
531 trace_amdgpu_bo_create(bo);
533 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
534 if (bp->type == ttm_bo_type_device)
535 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
541 ww_mutex_unlock(&bo->tbo.resv->lock);
542 amdgpu_bo_unref(&bo);
546 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
547 unsigned long size, int byte_align,
548 struct amdgpu_bo *bo)
550 struct amdgpu_bo_param bp;
556 memset(&bp, 0, sizeof(bp));
558 bp.byte_align = byte_align;
559 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
560 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
561 AMDGPU_GEM_CREATE_SHADOW;
562 bp.type = ttm_bo_type_kernel;
563 bp.resv = bo->tbo.resv;
565 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
567 bo->shadow->parent = amdgpu_bo_ref(bo);
568 mutex_lock(&adev->shadow_list_lock);
569 list_add_tail(&bo->shadow_list, &adev->shadow_list);
570 mutex_unlock(&adev->shadow_list_lock);
577 * amdgpu_bo_create - create an &amdgpu_bo buffer object
578 * @adev: amdgpu device object
579 * @bp: parameters to be used for the buffer object
580 * @bo_ptr: pointer to the buffer object pointer
582 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
584 * Shadow object is used to backup the original buffer object, and is always
588 * 0 for success or a negative error code on failure.
590 int amdgpu_bo_create(struct amdgpu_device *adev,
591 struct amdgpu_bo_param *bp,
592 struct amdgpu_bo **bo_ptr)
594 u64 flags = bp->flags;
597 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
598 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
602 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && amdgpu_bo_need_backup(adev)) {
604 WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
607 r = amdgpu_bo_create_shadow(adev, bp->size, bp->byte_align, (*bo_ptr));
610 reservation_object_unlock((*bo_ptr)->tbo.resv);
613 amdgpu_bo_unref(bo_ptr);
620 * amdgpu_bo_backup_to_shadow - Backs up an &amdgpu_bo buffer object
621 * @adev: amdgpu device object
622 * @ring: amdgpu_ring for the engine handling the buffer operations
623 * @bo: &amdgpu_bo buffer to be backed up
624 * @resv: reservation object with embedded fence
625 * @fence: dma_fence associated with the operation
626 * @direct: whether to submit the job directly
628 * Copies an &amdgpu_bo buffer object to its shadow object.
632 * 0 for success or a negative error code on failure.
634 int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
635 struct amdgpu_ring *ring,
636 struct amdgpu_bo *bo,
637 struct reservation_object *resv,
638 struct dma_fence **fence,
642 struct amdgpu_bo *shadow = bo->shadow;
643 uint64_t bo_addr, shadow_addr;
649 bo_addr = amdgpu_bo_gpu_offset(bo);
650 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
652 r = reservation_object_reserve_shared(bo->tbo.resv);
656 r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
657 amdgpu_bo_size(bo), resv, fence,
660 amdgpu_bo_fence(bo, *fence, true);
667 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
668 * @bo: pointer to the buffer object
670 * Sets placement according to domain; and changes placement and caching
671 * policy of the buffer object according to the placement.
672 * This is used for validating shadow bos. It calls ttm_bo_validate() to
673 * make sure the buffer is resident where it needs to be.
676 * 0 for success or a negative error code on failure.
678 int amdgpu_bo_validate(struct amdgpu_bo *bo)
680 struct ttm_operation_ctx ctx = { false, false };
687 domain = bo->preferred_domains;
690 amdgpu_bo_placement_from_domain(bo, domain);
691 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
692 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
693 domain = bo->allowed_domains;
701 * amdgpu_bo_restore_from_shadow - restore an &amdgpu_bo buffer object
702 * @adev: amdgpu device object
703 * @ring: amdgpu_ring for the engine handling the buffer operations
704 * @bo: &amdgpu_bo buffer to be restored
705 * @resv: reservation object with embedded fence
706 * @fence: dma_fence associated with the operation
707 * @direct: whether to submit the job directly
709 * Copies a buffer object's shadow content back to the object.
710 * This is used for recovering a buffer from its shadow in case of a gpu
711 * reset where vram context may be lost.
714 * 0 for success or a negative error code on failure.
716 int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
717 struct amdgpu_ring *ring,
718 struct amdgpu_bo *bo,
719 struct reservation_object *resv,
720 struct dma_fence **fence,
724 struct amdgpu_bo *shadow = bo->shadow;
725 uint64_t bo_addr, shadow_addr;
731 bo_addr = amdgpu_bo_gpu_offset(bo);
732 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
734 r = reservation_object_reserve_shared(bo->tbo.resv);
738 r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
739 amdgpu_bo_size(bo), resv, fence,
742 amdgpu_bo_fence(bo, *fence, true);
749 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
750 * @bo: &amdgpu_bo buffer object to be mapped
751 * @ptr: kernel virtual address to be returned
753 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
754 * amdgpu_bo_kptr() to get the kernel virtual address.
757 * 0 for success or a negative error code on failure.
759 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
764 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
767 kptr = amdgpu_bo_kptr(bo);
774 r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
775 MAX_SCHEDULE_TIMEOUT);
779 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
784 *ptr = amdgpu_bo_kptr(bo);
790 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
791 * @bo: &amdgpu_bo buffer object
793 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
796 * the virtual address of a buffer object area.
798 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
802 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
806 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
807 * @bo: &amdgpu_bo buffer object to be unmapped
809 * Unmaps a kernel map set up by amdgpu_bo_kmap().
811 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
814 ttm_bo_kunmap(&bo->kmap);
818 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
819 * @bo: &amdgpu_bo buffer object
821 * References the contained &ttm_buffer_object.
824 * a refcounted pointer to the &amdgpu_bo buffer object.
826 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
831 ttm_bo_get(&bo->tbo);
836 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
837 * @bo: &amdgpu_bo buffer object
839 * Unreferences the contained &ttm_buffer_object and clear the pointer
841 void amdgpu_bo_unref(struct amdgpu_bo **bo)
843 struct ttm_buffer_object *tbo;
854 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
855 * @bo: &amdgpu_bo buffer object to be pinned
856 * @domain: domain to be pinned to
857 * @min_offset: the start of requested address range
858 * @max_offset: the end of requested address range
860 * Pins the buffer object according to requested domain and address range. If
861 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
862 * pin_count and pin_size accordingly.
864 * Pinning means to lock pages in memory along with keeping them at a fixed
865 * offset. It is required when a buffer can not be moved, for example, when
866 * a display buffer is being scanned out.
868 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
869 * where to pin a buffer if there are specific restrictions on where a buffer
873 * 0 for success or a negative error code on failure.
875 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
876 u64 min_offset, u64 max_offset)
878 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
879 struct ttm_operation_ctx ctx = { false, false };
882 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
885 if (WARN_ON_ONCE(min_offset > max_offset))
888 /* A shared bo cannot be migrated to VRAM */
889 if (bo->prime_shared_count) {
890 if (domain & AMDGPU_GEM_DOMAIN_GTT)
891 domain = AMDGPU_GEM_DOMAIN_GTT;
896 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
897 * See function amdgpu_display_supported_domains()
899 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
902 uint32_t mem_type = bo->tbo.mem.mem_type;
904 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
909 if (max_offset != 0) {
910 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
911 WARN_ON_ONCE(max_offset <
912 (amdgpu_bo_gpu_offset(bo) - domain_start));
918 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
919 /* force to pin into visible video ram */
920 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
921 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
922 amdgpu_bo_placement_from_domain(bo, domain);
923 for (i = 0; i < bo->placement.num_placement; i++) {
926 fpfn = min_offset >> PAGE_SHIFT;
927 lpfn = max_offset >> PAGE_SHIFT;
929 if (fpfn > bo->placements[i].fpfn)
930 bo->placements[i].fpfn = fpfn;
931 if (!bo->placements[i].lpfn ||
932 (lpfn && lpfn < bo->placements[i].lpfn))
933 bo->placements[i].lpfn = lpfn;
934 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
937 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
939 dev_err(adev->dev, "%p pin failed\n", bo);
945 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
946 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
947 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
948 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
949 &adev->visible_pin_size);
950 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
951 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
959 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
960 * @bo: &amdgpu_bo buffer object to be pinned
961 * @domain: domain to be pinned to
963 * A simple wrapper to amdgpu_bo_pin_restricted().
964 * Provides a simpler API for buffers that do not have any strict restrictions
965 * on where a buffer must be located.
968 * 0 for success or a negative error code on failure.
970 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
972 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
976 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
977 * @bo: &amdgpu_bo buffer object to be unpinned
979 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
980 * Changes placement and pin size accordingly.
983 * 0 for success or a negative error code on failure.
985 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
987 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
988 struct ttm_operation_ctx ctx = { false, false };
991 if (!bo->pin_count) {
992 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
999 amdgpu_bo_subtract_pin_size(bo);
1001 for (i = 0; i < bo->placement.num_placement; i++) {
1002 bo->placements[i].lpfn = 0;
1003 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
1005 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1007 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
1013 * amdgpu_bo_evict_vram - evict VRAM buffers
1014 * @adev: amdgpu device object
1016 * Evicts all VRAM buffers on the lru list of the memory type.
1017 * Mainly used for evicting vram at suspend time.
1020 * 0 for success or a negative error code on failure.
1022 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1024 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
1025 #ifndef CONFIG_HIBERNATION
1026 if (adev->flags & AMD_IS_APU) {
1027 /* Useless to evict on IGP chips */
1031 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
1034 static const char *amdgpu_vram_names[] = {
1047 * amdgpu_bo_init - initialize memory manager
1048 * @adev: amdgpu device object
1050 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1053 * 0 for success or a negative error code on failure.
1055 int amdgpu_bo_init(struct amdgpu_device *adev)
1057 /* reserve PAT memory space to WC for VRAM */
1058 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1059 adev->gmc.aper_size);
1061 /* Add an MTRR for the VRAM */
1062 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1063 adev->gmc.aper_size);
1064 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1065 adev->gmc.mc_vram_size >> 20,
1066 (unsigned long long)adev->gmc.aper_size >> 20);
1067 DRM_INFO("RAM width %dbits %s\n",
1068 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1069 return amdgpu_ttm_init(adev);
1073 * amdgpu_bo_late_init - late init
1074 * @adev: amdgpu device object
1076 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1080 * 0 for success or a negative error code on failure.
1082 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1084 amdgpu_ttm_late_init(adev);
1090 * amdgpu_bo_fini - tear down memory manager
1091 * @adev: amdgpu device object
1093 * Reverses amdgpu_bo_init() to tear down memory manager.
1095 void amdgpu_bo_fini(struct amdgpu_device *adev)
1097 amdgpu_ttm_fini(adev);
1098 arch_phys_wc_del(adev->gmc.vram_mtrr);
1099 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1103 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1104 * @bo: &amdgpu_bo buffer object
1105 * @vma: vma as input from the fbdev mmap method
1107 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1110 * 0 for success or a negative error code on failure.
1112 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1113 struct vm_area_struct *vma)
1115 return ttm_fbdev_mmap(vma, &bo->tbo);
1119 * amdgpu_bo_set_tiling_flags - set tiling flags
1120 * @bo: &amdgpu_bo buffer object
1121 * @tiling_flags: new flags
1123 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1124 * kernel driver to set the tiling flags on a buffer.
1127 * 0 for success or a negative error code on failure.
1129 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1131 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1133 if (adev->family <= AMDGPU_FAMILY_CZ &&
1134 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1137 bo->tiling_flags = tiling_flags;
1142 * amdgpu_bo_get_tiling_flags - get tiling flags
1143 * @bo: &amdgpu_bo buffer object
1144 * @tiling_flags: returned flags
1146 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1147 * set the tiling flags on a buffer.
1149 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1151 lockdep_assert_held(&bo->tbo.resv->lock.base);
1154 *tiling_flags = bo->tiling_flags;
1158 * amdgpu_bo_set_metadata - set metadata
1159 * @bo: &amdgpu_bo buffer object
1160 * @metadata: new metadata
1161 * @metadata_size: size of the new metadata
1162 * @flags: flags of the new metadata
1164 * Sets buffer object's metadata, its size and flags.
1165 * Used via GEM ioctl.
1168 * 0 for success or a negative error code on failure.
1170 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1171 uint32_t metadata_size, uint64_t flags)
1175 if (!metadata_size) {
1176 if (bo->metadata_size) {
1177 kfree(bo->metadata);
1178 bo->metadata = NULL;
1179 bo->metadata_size = 0;
1184 if (metadata == NULL)
1187 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1191 kfree(bo->metadata);
1192 bo->metadata_flags = flags;
1193 bo->metadata = buffer;
1194 bo->metadata_size = metadata_size;
1200 * amdgpu_bo_get_metadata - get metadata
1201 * @bo: &amdgpu_bo buffer object
1202 * @buffer: returned metadata
1203 * @buffer_size: size of the buffer
1204 * @metadata_size: size of the returned metadata
1205 * @flags: flags of the returned metadata
1207 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1208 * less than metadata_size.
1209 * Used via GEM ioctl.
1212 * 0 for success or a negative error code on failure.
1214 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1215 size_t buffer_size, uint32_t *metadata_size,
1218 if (!buffer && !metadata_size)
1222 if (buffer_size < bo->metadata_size)
1225 if (bo->metadata_size)
1226 memcpy(buffer, bo->metadata, bo->metadata_size);
1230 *metadata_size = bo->metadata_size;
1232 *flags = bo->metadata_flags;
1238 * amdgpu_bo_move_notify - notification about a memory move
1239 * @bo: pointer to a buffer object
1240 * @evict: if this move is evicting the buffer from the graphics address space
1241 * @new_mem: new information of the bufer object
1243 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1245 * TTM driver callback which is called when ttm moves a buffer.
1247 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1249 struct ttm_mem_reg *new_mem)
1251 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1252 struct amdgpu_bo *abo;
1253 struct ttm_mem_reg *old_mem = &bo->mem;
1255 if (!amdgpu_bo_is_amdgpu_bo(bo))
1258 abo = ttm_to_amdgpu_bo(bo);
1259 amdgpu_vm_bo_invalidate(adev, abo, evict);
1261 amdgpu_bo_kunmap(abo);
1263 /* remember the eviction */
1265 atomic64_inc(&adev->num_evictions);
1267 /* update statistics */
1271 /* move_notify is called before move happens */
1272 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1276 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1277 * @bo: pointer to a buffer object
1279 * Notifies the driver we are taking a fault on this BO and have reserved it,
1280 * also performs bookkeeping.
1281 * TTM driver callback for dealing with vm faults.
1284 * 0 for success or a negative error code on failure.
1286 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1288 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1289 struct ttm_operation_ctx ctx = { false, false };
1290 struct amdgpu_bo *abo;
1291 unsigned long offset, size;
1294 if (!amdgpu_bo_is_amdgpu_bo(bo))
1297 abo = ttm_to_amdgpu_bo(bo);
1299 /* Remember that this BO was accessed by the CPU */
1300 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1302 if (bo->mem.mem_type != TTM_PL_VRAM)
1305 size = bo->mem.num_pages << PAGE_SHIFT;
1306 offset = bo->mem.start << PAGE_SHIFT;
1307 if ((offset + size) <= adev->gmc.visible_vram_size)
1310 /* Can't move a pinned BO to visible VRAM */
1311 if (abo->pin_count > 0)
1314 /* hurrah the memory is not visible ! */
1315 atomic64_inc(&adev->num_vram_cpu_page_faults);
1316 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1317 AMDGPU_GEM_DOMAIN_GTT);
1319 /* Avoid costly evictions; only set GTT as a busy placement */
1320 abo->placement.num_busy_placement = 1;
1321 abo->placement.busy_placement = &abo->placements[1];
1323 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1324 if (unlikely(r != 0))
1327 offset = bo->mem.start << PAGE_SHIFT;
1328 /* this should never happen */
1329 if (bo->mem.mem_type == TTM_PL_VRAM &&
1330 (offset + size) > adev->gmc.visible_vram_size)
1337 * amdgpu_bo_fence - add fence to buffer object
1339 * @bo: buffer object in question
1340 * @fence: fence to add
1341 * @shared: true if fence should be added shared
1344 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1347 struct reservation_object *resv = bo->tbo.resv;
1350 reservation_object_add_shared_fence(resv, fence);
1352 reservation_object_add_excl_fence(resv, fence);
1356 * amdgpu_bo_gpu_offset - return GPU offset of bo
1357 * @bo: amdgpu object for which we query the offset
1359 * Note: object should either be pinned or reserved when calling this
1360 * function, it might be useful to add check for this for debugging.
1363 * current GPU offset of the object.
1365 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1367 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1368 WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1369 !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1370 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1371 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1372 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1374 return amdgpu_gmc_sign_extend(bo->tbo.offset);
1378 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1379 * @adev: amdgpu device object
1380 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1383 * Which of the allowed domains is preferred for pinning the BO for scanout.
1385 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1388 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1389 domain = AMDGPU_GEM_DOMAIN_VRAM;
1390 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1391 domain = AMDGPU_GEM_DOMAIN_GTT;