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.
55 * amdgpu_bo_subtract_pin_size - Remove BO from pin_size accounting
57 * @bo: &amdgpu_bo buffer object
59 * This function is called when a BO stops being pinned, and updates the
60 * &amdgpu_device pin_size values accordingly.
62 static void amdgpu_bo_subtract_pin_size(struct amdgpu_bo *bo)
64 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
66 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
67 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
68 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
69 &adev->visible_pin_size);
70 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
71 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
75 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
77 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
78 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
80 if (bo->pin_count > 0)
81 amdgpu_bo_subtract_pin_size(bo);
85 if (bo->tbo.base.import_attach)
86 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
87 drm_gem_object_release(&bo->tbo.base);
88 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
89 if (!list_empty(&bo->shadow_list)) {
90 mutex_lock(&adev->shadow_list_lock);
91 list_del_init(&bo->shadow_list);
92 mutex_unlock(&adev->shadow_list_lock);
94 amdgpu_bo_unref(&bo->parent);
101 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
102 * @bo: buffer object to be checked
104 * Uses destroy function associated with the object to determine if this is
108 * true if the object belongs to &amdgpu_bo, false if not.
110 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
112 if (bo->destroy == &amdgpu_bo_destroy)
118 * amdgpu_bo_placement_from_domain - set buffer's placement
119 * @abo: &amdgpu_bo buffer object whose placement is to be set
120 * @domain: requested domain
122 * Sets buffer's placement according to requested domain and the buffer's
125 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
127 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
128 struct ttm_placement *placement = &abo->placement;
129 struct ttm_place *places = abo->placements;
130 u64 flags = abo->flags;
133 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
134 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
138 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
141 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
142 places[c].lpfn = visible_pfn;
144 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
146 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
147 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
151 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
154 places[c].flags = TTM_PL_FLAG_TT;
155 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
156 places[c].flags |= TTM_PL_FLAG_WC |
157 TTM_PL_FLAG_UNCACHED;
159 places[c].flags |= TTM_PL_FLAG_CACHED;
163 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
166 places[c].flags = TTM_PL_FLAG_SYSTEM;
167 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
168 places[c].flags |= TTM_PL_FLAG_WC |
169 TTM_PL_FLAG_UNCACHED;
171 places[c].flags |= TTM_PL_FLAG_CACHED;
175 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
178 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
182 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
185 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
189 if (domain & AMDGPU_GEM_DOMAIN_OA) {
192 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
199 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
203 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
205 placement->num_placement = c;
206 placement->placement = places;
208 placement->num_busy_placement = c;
209 placement->busy_placement = places;
213 * amdgpu_bo_create_reserved - create reserved BO for kernel use
215 * @adev: amdgpu device object
216 * @size: size for the new BO
217 * @align: alignment for the new BO
218 * @domain: where to place it
219 * @bo_ptr: used to initialize BOs in structures
220 * @gpu_addr: GPU addr of the pinned BO
221 * @cpu_addr: optional CPU address mapping
223 * Allocates and pins a BO for kernel internal use, and returns it still
226 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
229 * 0 on success, negative error code otherwise.
231 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
232 unsigned long size, int align,
233 u32 domain, struct amdgpu_bo **bo_ptr,
234 u64 *gpu_addr, void **cpu_addr)
236 struct amdgpu_bo_param bp;
241 amdgpu_bo_unref(bo_ptr);
245 memset(&bp, 0, sizeof(bp));
247 bp.byte_align = align;
249 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
250 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
251 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
252 bp.type = ttm_bo_type_kernel;
256 r = amdgpu_bo_create(adev, &bp, bo_ptr);
258 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
265 r = amdgpu_bo_reserve(*bo_ptr, false);
267 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
271 r = amdgpu_bo_pin(*bo_ptr, domain);
273 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
274 goto error_unreserve;
277 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
279 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
284 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
287 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
289 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
297 amdgpu_bo_unpin(*bo_ptr);
299 amdgpu_bo_unreserve(*bo_ptr);
303 amdgpu_bo_unref(bo_ptr);
309 * amdgpu_bo_create_kernel - create BO for kernel use
311 * @adev: amdgpu device object
312 * @size: size for the new BO
313 * @align: alignment for the new BO
314 * @domain: where to place it
315 * @bo_ptr: used to initialize BOs in structures
316 * @gpu_addr: GPU addr of the pinned BO
317 * @cpu_addr: optional CPU address mapping
319 * Allocates and pins a BO for kernel internal use.
321 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
324 * 0 on success, negative error code otherwise.
326 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
327 unsigned long size, int align,
328 u32 domain, struct amdgpu_bo **bo_ptr,
329 u64 *gpu_addr, void **cpu_addr)
333 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
340 amdgpu_bo_unreserve(*bo_ptr);
346 * amdgpu_bo_free_kernel - free BO for kernel use
348 * @bo: amdgpu BO to free
349 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
350 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
352 * unmaps and unpin a BO for kernel internal use.
354 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
360 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
362 amdgpu_bo_kunmap(*bo);
364 amdgpu_bo_unpin(*bo);
365 amdgpu_bo_unreserve(*bo);
376 /* Validate bo size is bit bigger then the request domain */
377 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
378 unsigned long size, u32 domain)
380 struct ttm_mem_type_manager *man = NULL;
383 * If GTT is part of requested domains the check must succeed to
384 * allow fall back to GTT
386 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
387 man = &adev->mman.bdev.man[TTM_PL_TT];
389 if (size < (man->size << PAGE_SHIFT))
395 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
396 man = &adev->mman.bdev.man[TTM_PL_VRAM];
398 if (size < (man->size << PAGE_SHIFT))
405 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
409 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
410 man->size << PAGE_SHIFT);
414 bool amdgpu_bo_support_uswc(u64 bo_flags)
418 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
419 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
422 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
423 /* Don't try to enable write-combining when it can't work, or things
425 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
428 #ifndef CONFIG_COMPILE_TEST
429 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
430 thanks to write-combining
433 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
434 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
435 "better performance thanks to write-combining\n");
438 /* For architectures that don't support WC memory,
439 * mask out the WC flag from the BO
441 if (!drm_arch_can_wc_memory())
448 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
449 struct amdgpu_bo_param *bp,
450 struct amdgpu_bo **bo_ptr)
452 struct ttm_operation_ctx ctx = {
453 .interruptible = (bp->type != ttm_bo_type_kernel),
454 .no_wait_gpu = false,
456 .flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
458 struct amdgpu_bo *bo;
459 unsigned long page_align, size = bp->size;
463 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
464 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
465 /* GWS and OA don't need any alignment. */
466 page_align = bp->byte_align;
468 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
469 /* Both size and alignment must be a multiple of 4. */
470 page_align = ALIGN(bp->byte_align, 4);
471 size = ALIGN(size, 4) << PAGE_SHIFT;
473 /* Memory should be aligned at least to a page size. */
474 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
475 size = ALIGN(size, PAGE_SIZE);
478 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
483 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
484 sizeof(struct amdgpu_bo));
486 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
489 drm_gem_private_object_init(adev->ddev, &bo->tbo.base, size);
490 INIT_LIST_HEAD(&bo->shadow_list);
492 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
494 bo->allowed_domains = bo->preferred_domains;
495 if (bp->type != ttm_bo_type_kernel &&
496 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
497 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
499 bo->flags = bp->flags;
501 if (!amdgpu_bo_support_uswc(bo->flags))
502 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
504 bo->tbo.bdev = &adev->mman.bdev;
505 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
506 AMDGPU_GEM_DOMAIN_GDS))
507 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
509 amdgpu_bo_placement_from_domain(bo, bp->domain);
510 if (bp->type == ttm_bo_type_kernel)
511 bo->tbo.priority = 1;
513 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
514 &bo->placement, page_align, &ctx, acc_size,
515 NULL, bp->resv, &amdgpu_bo_destroy);
516 if (unlikely(r != 0))
519 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
520 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
521 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
522 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
525 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
527 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
528 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
529 struct dma_fence *fence;
531 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
535 amdgpu_bo_fence(bo, fence, false);
536 dma_fence_put(bo->tbo.moving);
537 bo->tbo.moving = dma_fence_get(fence);
538 dma_fence_put(fence);
541 amdgpu_bo_unreserve(bo);
544 trace_amdgpu_bo_create(bo);
546 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
547 if (bp->type == ttm_bo_type_device)
548 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
554 dma_resv_unlock(bo->tbo.base.resv);
555 amdgpu_bo_unref(&bo);
559 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
561 struct amdgpu_bo *bo)
563 struct amdgpu_bo_param bp;
569 memset(&bp, 0, sizeof(bp));
571 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
572 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
573 AMDGPU_GEM_CREATE_SHADOW;
574 bp.type = ttm_bo_type_kernel;
575 bp.resv = bo->tbo.base.resv;
577 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
579 bo->shadow->parent = amdgpu_bo_ref(bo);
580 mutex_lock(&adev->shadow_list_lock);
581 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
582 mutex_unlock(&adev->shadow_list_lock);
589 * amdgpu_bo_create - create an &amdgpu_bo buffer object
590 * @adev: amdgpu device object
591 * @bp: parameters to be used for the buffer object
592 * @bo_ptr: pointer to the buffer object pointer
594 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
596 * Shadow object is used to backup the original buffer object, and is always
600 * 0 for success or a negative error code on failure.
602 int amdgpu_bo_create(struct amdgpu_device *adev,
603 struct amdgpu_bo_param *bp,
604 struct amdgpu_bo **bo_ptr)
606 u64 flags = bp->flags;
609 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
610 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
614 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
616 WARN_ON(dma_resv_lock((*bo_ptr)->tbo.base.resv,
619 r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
622 dma_resv_unlock((*bo_ptr)->tbo.base.resv);
625 amdgpu_bo_unref(bo_ptr);
632 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
633 * @bo: pointer to the buffer object
635 * Sets placement according to domain; and changes placement and caching
636 * policy of the buffer object according to the placement.
637 * This is used for validating shadow bos. It calls ttm_bo_validate() to
638 * make sure the buffer is resident where it needs to be.
641 * 0 for success or a negative error code on failure.
643 int amdgpu_bo_validate(struct amdgpu_bo *bo)
645 struct ttm_operation_ctx ctx = { false, false };
652 domain = bo->preferred_domains;
655 amdgpu_bo_placement_from_domain(bo, domain);
656 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
657 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
658 domain = bo->allowed_domains;
666 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
668 * @shadow: &amdgpu_bo shadow to be restored
669 * @fence: dma_fence associated with the operation
671 * Copies a buffer object's shadow content back to the object.
672 * This is used for recovering a buffer from its shadow in case of a gpu
673 * reset where vram context may be lost.
676 * 0 for success or a negative error code on failure.
678 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
681 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
682 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
683 uint64_t shadow_addr, parent_addr;
685 shadow_addr = amdgpu_bo_gpu_offset(shadow);
686 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
688 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
689 amdgpu_bo_size(shadow), NULL, fence,
694 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
695 * @bo: &amdgpu_bo buffer object to be mapped
696 * @ptr: kernel virtual address to be returned
698 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
699 * amdgpu_bo_kptr() to get the kernel virtual address.
702 * 0 for success or a negative error code on failure.
704 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
709 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
712 kptr = amdgpu_bo_kptr(bo);
719 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false, false,
720 MAX_SCHEDULE_TIMEOUT);
724 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
729 *ptr = amdgpu_bo_kptr(bo);
735 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
736 * @bo: &amdgpu_bo buffer object
738 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
741 * the virtual address of a buffer object area.
743 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
747 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
751 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
752 * @bo: &amdgpu_bo buffer object to be unmapped
754 * Unmaps a kernel map set up by amdgpu_bo_kmap().
756 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
759 ttm_bo_kunmap(&bo->kmap);
763 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
764 * @bo: &amdgpu_bo buffer object
766 * References the contained &ttm_buffer_object.
769 * a refcounted pointer to the &amdgpu_bo buffer object.
771 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
776 ttm_bo_get(&bo->tbo);
781 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
782 * @bo: &amdgpu_bo buffer object
784 * Unreferences the contained &ttm_buffer_object and clear the pointer
786 void amdgpu_bo_unref(struct amdgpu_bo **bo)
788 struct ttm_buffer_object *tbo;
799 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
800 * @bo: &amdgpu_bo buffer object to be pinned
801 * @domain: domain to be pinned to
802 * @min_offset: the start of requested address range
803 * @max_offset: the end of requested address range
805 * Pins the buffer object according to requested domain and address range. If
806 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
807 * pin_count and pin_size accordingly.
809 * Pinning means to lock pages in memory along with keeping them at a fixed
810 * offset. It is required when a buffer can not be moved, for example, when
811 * a display buffer is being scanned out.
813 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
814 * where to pin a buffer if there are specific restrictions on where a buffer
818 * 0 for success or a negative error code on failure.
820 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
821 u64 min_offset, u64 max_offset)
823 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
824 struct ttm_operation_ctx ctx = { false, false };
827 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
830 if (WARN_ON_ONCE(min_offset > max_offset))
833 /* A shared bo cannot be migrated to VRAM */
834 if (bo->prime_shared_count) {
835 if (domain & AMDGPU_GEM_DOMAIN_GTT)
836 domain = AMDGPU_GEM_DOMAIN_GTT;
841 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
842 * See function amdgpu_display_supported_domains()
844 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
847 uint32_t mem_type = bo->tbo.mem.mem_type;
849 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
854 if (max_offset != 0) {
855 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
856 WARN_ON_ONCE(max_offset <
857 (amdgpu_bo_gpu_offset(bo) - domain_start));
863 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
864 /* force to pin into visible video ram */
865 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
866 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
867 amdgpu_bo_placement_from_domain(bo, domain);
868 for (i = 0; i < bo->placement.num_placement; i++) {
871 fpfn = min_offset >> PAGE_SHIFT;
872 lpfn = max_offset >> PAGE_SHIFT;
874 if (fpfn > bo->placements[i].fpfn)
875 bo->placements[i].fpfn = fpfn;
876 if (!bo->placements[i].lpfn ||
877 (lpfn && lpfn < bo->placements[i].lpfn))
878 bo->placements[i].lpfn = lpfn;
879 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
882 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
884 dev_err(adev->dev, "%p pin failed\n", bo);
890 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
891 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
892 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
893 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
894 &adev->visible_pin_size);
895 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
896 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
904 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
905 * @bo: &amdgpu_bo buffer object to be pinned
906 * @domain: domain to be pinned to
908 * A simple wrapper to amdgpu_bo_pin_restricted().
909 * Provides a simpler API for buffers that do not have any strict restrictions
910 * on where a buffer must be located.
913 * 0 for success or a negative error code on failure.
915 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
917 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
921 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
922 * @bo: &amdgpu_bo buffer object to be unpinned
924 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
925 * Changes placement and pin size accordingly.
928 * 0 for success or a negative error code on failure.
930 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
932 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
933 struct ttm_operation_ctx ctx = { false, false };
936 if (WARN_ON_ONCE(!bo->pin_count)) {
937 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
944 amdgpu_bo_subtract_pin_size(bo);
946 for (i = 0; i < bo->placement.num_placement; i++) {
947 bo->placements[i].lpfn = 0;
948 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
950 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
952 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
958 * amdgpu_bo_evict_vram - evict VRAM buffers
959 * @adev: amdgpu device object
961 * Evicts all VRAM buffers on the lru list of the memory type.
962 * Mainly used for evicting vram at suspend time.
965 * 0 for success or a negative error code on failure.
967 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
969 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
970 #ifndef CONFIG_HIBERNATION
971 if (adev->flags & AMD_IS_APU) {
972 /* Useless to evict on IGP chips */
976 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
979 static const char *amdgpu_vram_names[] = {
993 * amdgpu_bo_init - initialize memory manager
994 * @adev: amdgpu device object
996 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
999 * 0 for success or a negative error code on failure.
1001 int amdgpu_bo_init(struct amdgpu_device *adev)
1003 /* reserve PAT memory space to WC for VRAM */
1004 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1005 adev->gmc.aper_size);
1007 /* Add an MTRR for the VRAM */
1008 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1009 adev->gmc.aper_size);
1010 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1011 adev->gmc.mc_vram_size >> 20,
1012 (unsigned long long)adev->gmc.aper_size >> 20);
1013 DRM_INFO("RAM width %dbits %s\n",
1014 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1015 return amdgpu_ttm_init(adev);
1019 * amdgpu_bo_late_init - late init
1020 * @adev: amdgpu device object
1022 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1026 * 0 for success or a negative error code on failure.
1028 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1030 amdgpu_ttm_late_init(adev);
1036 * amdgpu_bo_fini - tear down memory manager
1037 * @adev: amdgpu device object
1039 * Reverses amdgpu_bo_init() to tear down memory manager.
1041 void amdgpu_bo_fini(struct amdgpu_device *adev)
1043 amdgpu_ttm_fini(adev);
1044 arch_phys_wc_del(adev->gmc.vram_mtrr);
1045 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1049 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1050 * @bo: &amdgpu_bo buffer object
1051 * @vma: vma as input from the fbdev mmap method
1053 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1056 * 0 for success or a negative error code on failure.
1058 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1059 struct vm_area_struct *vma)
1061 return ttm_fbdev_mmap(vma, &bo->tbo);
1065 * amdgpu_bo_set_tiling_flags - set tiling flags
1066 * @bo: &amdgpu_bo buffer object
1067 * @tiling_flags: new flags
1069 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1070 * kernel driver to set the tiling flags on a buffer.
1073 * 0 for success or a negative error code on failure.
1075 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1077 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1079 if (adev->family <= AMDGPU_FAMILY_CZ &&
1080 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1083 bo->tiling_flags = tiling_flags;
1088 * amdgpu_bo_get_tiling_flags - get tiling flags
1089 * @bo: &amdgpu_bo buffer object
1090 * @tiling_flags: returned flags
1092 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1093 * set the tiling flags on a buffer.
1095 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1097 dma_resv_assert_held(bo->tbo.base.resv);
1100 *tiling_flags = bo->tiling_flags;
1104 * amdgpu_bo_set_metadata - set metadata
1105 * @bo: &amdgpu_bo buffer object
1106 * @metadata: new metadata
1107 * @metadata_size: size of the new metadata
1108 * @flags: flags of the new metadata
1110 * Sets buffer object's metadata, its size and flags.
1111 * Used via GEM ioctl.
1114 * 0 for success or a negative error code on failure.
1116 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1117 uint32_t metadata_size, uint64_t flags)
1121 if (!metadata_size) {
1122 if (bo->metadata_size) {
1123 kfree(bo->metadata);
1124 bo->metadata = NULL;
1125 bo->metadata_size = 0;
1130 if (metadata == NULL)
1133 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1137 kfree(bo->metadata);
1138 bo->metadata_flags = flags;
1139 bo->metadata = buffer;
1140 bo->metadata_size = metadata_size;
1146 * amdgpu_bo_get_metadata - get metadata
1147 * @bo: &amdgpu_bo buffer object
1148 * @buffer: returned metadata
1149 * @buffer_size: size of the buffer
1150 * @metadata_size: size of the returned metadata
1151 * @flags: flags of the returned metadata
1153 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1154 * less than metadata_size.
1155 * Used via GEM ioctl.
1158 * 0 for success or a negative error code on failure.
1160 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1161 size_t buffer_size, uint32_t *metadata_size,
1164 if (!buffer && !metadata_size)
1168 if (buffer_size < bo->metadata_size)
1171 if (bo->metadata_size)
1172 memcpy(buffer, bo->metadata, bo->metadata_size);
1176 *metadata_size = bo->metadata_size;
1178 *flags = bo->metadata_flags;
1184 * amdgpu_bo_move_notify - notification about a memory move
1185 * @bo: pointer to a buffer object
1186 * @evict: if this move is evicting the buffer from the graphics address space
1187 * @new_mem: new information of the bufer object
1189 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1191 * TTM driver callback which is called when ttm moves a buffer.
1193 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1195 struct ttm_mem_reg *new_mem)
1197 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1198 struct amdgpu_bo *abo;
1199 struct ttm_mem_reg *old_mem = &bo->mem;
1201 if (!amdgpu_bo_is_amdgpu_bo(bo))
1204 abo = ttm_to_amdgpu_bo(bo);
1205 amdgpu_vm_bo_invalidate(adev, abo, evict);
1207 amdgpu_bo_kunmap(abo);
1209 /* remember the eviction */
1211 atomic64_inc(&adev->num_evictions);
1213 /* update statistics */
1217 /* move_notify is called before move happens */
1218 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1222 * amdgpu_bo_move_notify - notification about a BO being released
1223 * @bo: pointer to a buffer object
1225 * Wipes VRAM buffers whose contents should not be leaked before the
1226 * memory is released.
1228 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1230 struct dma_fence *fence = NULL;
1231 struct amdgpu_bo *abo;
1234 if (!amdgpu_bo_is_amdgpu_bo(bo))
1237 abo = ttm_to_amdgpu_bo(bo);
1240 amdgpu_amdkfd_unreserve_memory_limit(abo);
1242 if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
1243 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1246 dma_resv_lock(bo->base.resv, NULL);
1248 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1250 amdgpu_bo_fence(abo, fence, false);
1251 dma_fence_put(fence);
1254 dma_resv_unlock(bo->base.resv);
1258 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1259 * @bo: pointer to a buffer object
1261 * Notifies the driver we are taking a fault on this BO and have reserved it,
1262 * also performs bookkeeping.
1263 * TTM driver callback for dealing with vm faults.
1266 * 0 for success or a negative error code on failure.
1268 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1270 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1271 struct ttm_operation_ctx ctx = { false, false };
1272 struct amdgpu_bo *abo;
1273 unsigned long offset, size;
1276 if (!amdgpu_bo_is_amdgpu_bo(bo))
1279 abo = ttm_to_amdgpu_bo(bo);
1281 /* Remember that this BO was accessed by the CPU */
1282 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1284 if (bo->mem.mem_type != TTM_PL_VRAM)
1287 size = bo->mem.num_pages << PAGE_SHIFT;
1288 offset = bo->mem.start << PAGE_SHIFT;
1289 if ((offset + size) <= adev->gmc.visible_vram_size)
1292 /* Can't move a pinned BO to visible VRAM */
1293 if (abo->pin_count > 0)
1296 /* hurrah the memory is not visible ! */
1297 atomic64_inc(&adev->num_vram_cpu_page_faults);
1298 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1299 AMDGPU_GEM_DOMAIN_GTT);
1301 /* Avoid costly evictions; only set GTT as a busy placement */
1302 abo->placement.num_busy_placement = 1;
1303 abo->placement.busy_placement = &abo->placements[1];
1305 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1306 if (unlikely(r != 0))
1309 offset = bo->mem.start << PAGE_SHIFT;
1310 /* this should never happen */
1311 if (bo->mem.mem_type == TTM_PL_VRAM &&
1312 (offset + size) > adev->gmc.visible_vram_size)
1319 * amdgpu_bo_fence - add fence to buffer object
1321 * @bo: buffer object in question
1322 * @fence: fence to add
1323 * @shared: true if fence should be added shared
1326 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1329 struct dma_resv *resv = bo->tbo.base.resv;
1332 dma_resv_add_shared_fence(resv, fence);
1334 dma_resv_add_excl_fence(resv, fence);
1338 * amdgpu_sync_wait_resv - Wait for BO reservation fences
1340 * @bo: buffer object
1341 * @owner: fence owner
1342 * @intr: Whether the wait is interruptible
1345 * 0 on success, errno otherwise.
1347 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1349 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1350 struct amdgpu_sync sync;
1353 amdgpu_sync_create(&sync);
1354 amdgpu_sync_resv(adev, &sync, bo->tbo.base.resv, owner, false);
1355 r = amdgpu_sync_wait(&sync, intr);
1356 amdgpu_sync_free(&sync);
1362 * amdgpu_bo_gpu_offset - return GPU offset of bo
1363 * @bo: amdgpu object for which we query the offset
1365 * Note: object should either be pinned or reserved when calling this
1366 * function, it might be useful to add check for this for debugging.
1369 * current GPU offset of the object.
1371 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1373 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1374 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1375 !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1376 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1377 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1378 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1380 return amdgpu_gmc_sign_extend(bo->tbo.offset);
1384 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1385 * @adev: amdgpu device object
1386 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1389 * Which of the allowed domains is preferred for pinning the BO for scanout.
1391 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1394 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1395 domain = AMDGPU_GEM_DOMAIN_VRAM;
1396 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1397 domain = AMDGPU_GEM_DOMAIN_GTT;