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_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)))
66 static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
68 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
69 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
72 amdgpu_amdkfd_unreserve_system_memory_limit(bo);
76 if (bo->gem_base.import_attach)
77 drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
78 drm_gem_object_release(&bo->gem_base);
79 amdgpu_bo_unref(&bo->parent);
80 if (!list_empty(&bo->shadow_list)) {
81 mutex_lock(&adev->shadow_list_lock);
82 list_del_init(&bo->shadow_list);
83 mutex_unlock(&adev->shadow_list_lock);
90 * amdgpu_ttm_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
91 * @bo: buffer object to be checked
93 * Uses destroy function associated with the object to determine if this is
97 * true if the object belongs to &amdgpu_bo, false if not.
99 bool amdgpu_ttm_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
101 if (bo->destroy == &amdgpu_ttm_bo_destroy)
107 * amdgpu_ttm_placement_from_domain - set buffer's placement
108 * @abo: &amdgpu_bo buffer object whose placement is to be set
109 * @domain: requested domain
111 * Sets buffer's placement according to requested domain and the buffer's
114 void amdgpu_ttm_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
116 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
117 struct ttm_placement *placement = &abo->placement;
118 struct ttm_place *places = abo->placements;
119 u64 flags = abo->flags;
122 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
123 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
127 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
130 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
131 places[c].lpfn = visible_pfn;
133 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
135 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
136 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
140 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
142 if (flags & AMDGPU_GEM_CREATE_SHADOW)
143 places[c].lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
146 places[c].flags = TTM_PL_FLAG_TT;
147 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
148 places[c].flags |= TTM_PL_FLAG_WC |
149 TTM_PL_FLAG_UNCACHED;
151 places[c].flags |= TTM_PL_FLAG_CACHED;
155 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
158 places[c].flags = TTM_PL_FLAG_SYSTEM;
159 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
160 places[c].flags |= TTM_PL_FLAG_WC |
161 TTM_PL_FLAG_UNCACHED;
163 places[c].flags |= TTM_PL_FLAG_CACHED;
167 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
170 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
174 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
177 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
181 if (domain & AMDGPU_GEM_DOMAIN_OA) {
184 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
191 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
195 placement->num_placement = c;
196 placement->placement = places;
198 placement->num_busy_placement = c;
199 placement->busy_placement = places;
203 * amdgpu_bo_create_reserved - create reserved BO for kernel use
205 * @adev: amdgpu device object
206 * @size: size for the new BO
207 * @align: alignment for the new BO
208 * @domain: where to place it
209 * @bo_ptr: used to initialize BOs in structures
210 * @gpu_addr: GPU addr of the pinned BO
211 * @cpu_addr: optional CPU address mapping
213 * Allocates and pins a BO for kernel internal use, and returns it still
216 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
219 * 0 on success, negative error code otherwise.
221 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
222 unsigned long size, int align,
223 u32 domain, struct amdgpu_bo **bo_ptr,
224 u64 *gpu_addr, void **cpu_addr)
226 struct amdgpu_bo_param bp;
230 memset(&bp, 0, sizeof(bp));
232 bp.byte_align = align;
234 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
235 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
236 bp.type = ttm_bo_type_kernel;
240 r = amdgpu_bo_create(adev, &bp, bo_ptr);
242 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
249 r = amdgpu_bo_reserve(*bo_ptr, false);
251 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
255 r = amdgpu_bo_pin(*bo_ptr, domain, gpu_addr);
257 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
258 goto error_unreserve;
262 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
264 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
265 goto error_unreserve;
272 amdgpu_bo_unreserve(*bo_ptr);
276 amdgpu_bo_unref(bo_ptr);
282 * amdgpu_bo_create_kernel - create BO for kernel use
284 * @adev: amdgpu device object
285 * @size: size for the new BO
286 * @align: alignment for the new BO
287 * @domain: where to place it
288 * @bo_ptr: used to initialize BOs in structures
289 * @gpu_addr: GPU addr of the pinned BO
290 * @cpu_addr: optional CPU address mapping
292 * Allocates and pins a BO for kernel internal use.
294 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
297 * 0 on success, negative error code otherwise.
299 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
300 unsigned long size, int align,
301 u32 domain, struct amdgpu_bo **bo_ptr,
302 u64 *gpu_addr, void **cpu_addr)
306 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
312 amdgpu_bo_unreserve(*bo_ptr);
318 * amdgpu_bo_free_kernel - free BO for kernel use
320 * @bo: amdgpu BO to free
321 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
322 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
324 * unmaps and unpin a BO for kernel internal use.
326 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
332 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
334 amdgpu_bo_kunmap(*bo);
336 amdgpu_bo_unpin(*bo);
337 amdgpu_bo_unreserve(*bo);
348 /* Validate bo size is bit bigger then the request domain */
349 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
350 unsigned long size, u32 domain)
352 struct ttm_mem_type_manager *man = NULL;
355 * If GTT is part of requested domains the check must succeed to
356 * allow fall back to GTT
358 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
359 man = &adev->mman.bdev.man[TTM_PL_TT];
361 if (size < (man->size << PAGE_SHIFT))
367 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
368 man = &adev->mman.bdev.man[TTM_PL_VRAM];
370 if (size < (man->size << PAGE_SHIFT))
377 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
381 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
382 man->size << PAGE_SHIFT);
386 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
387 struct amdgpu_bo_param *bp,
388 struct amdgpu_bo **bo_ptr)
390 struct ttm_operation_ctx ctx = {
391 .interruptible = (bp->type != ttm_bo_type_kernel),
392 .no_wait_gpu = false,
394 .flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
396 struct amdgpu_bo *bo;
397 unsigned long page_align, size = bp->size;
401 page_align = roundup(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
402 size = ALIGN(size, PAGE_SIZE);
404 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
409 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
410 sizeof(struct amdgpu_bo));
412 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
415 drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
416 INIT_LIST_HEAD(&bo->shadow_list);
417 INIT_LIST_HEAD(&bo->va);
418 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
420 bo->allowed_domains = bo->preferred_domains;
421 if (bp->type != ttm_bo_type_kernel &&
422 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
423 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
425 bo->flags = bp->flags;
428 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
429 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
431 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
432 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
433 /* Don't try to enable write-combining when it can't work, or things
435 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
438 #ifndef CONFIG_COMPILE_TEST
439 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
440 thanks to write-combining
443 if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
444 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
445 "better performance thanks to write-combining\n");
446 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
448 /* For architectures that don't support WC memory,
449 * mask out the WC flag from the BO
451 if (!drm_arch_can_wc_memory())
452 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
455 bo->tbo.bdev = &adev->mman.bdev;
456 amdgpu_ttm_placement_from_domain(bo, bp->domain);
457 if (bp->type == ttm_bo_type_kernel)
458 bo->tbo.priority = 1;
460 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
461 &bo->placement, page_align, &ctx, acc_size,
462 NULL, bp->resv, &amdgpu_ttm_bo_destroy);
463 if (unlikely(r != 0))
466 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
467 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
468 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
469 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
472 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
474 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
475 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
476 struct dma_fence *fence;
478 r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
482 amdgpu_bo_fence(bo, fence, false);
483 dma_fence_put(bo->tbo.moving);
484 bo->tbo.moving = dma_fence_get(fence);
485 dma_fence_put(fence);
488 amdgpu_bo_unreserve(bo);
491 trace_amdgpu_bo_create(bo);
493 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
494 if (bp->type == ttm_bo_type_device)
495 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
501 ww_mutex_unlock(&bo->tbo.resv->lock);
502 amdgpu_bo_unref(&bo);
506 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
507 unsigned long size, int byte_align,
508 struct amdgpu_bo *bo)
510 struct amdgpu_bo_param bp;
516 memset(&bp, 0, sizeof(bp));
518 bp.byte_align = byte_align;
519 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
520 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
521 AMDGPU_GEM_CREATE_SHADOW;
522 bp.type = ttm_bo_type_kernel;
523 bp.resv = bo->tbo.resv;
525 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
527 bo->shadow->parent = amdgpu_bo_ref(bo);
528 mutex_lock(&adev->shadow_list_lock);
529 list_add_tail(&bo->shadow_list, &adev->shadow_list);
530 mutex_unlock(&adev->shadow_list_lock);
537 * amdgpu_bo_create - create an &amdgpu_bo buffer object
538 * @adev: amdgpu device object
539 * @bp: parameters to be used for the buffer object
540 * @bo_ptr: pointer to the buffer object pointer
542 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
544 * Shadow object is used to backup the original buffer object, and is always
548 * 0 for success or a negative error code on failure.
550 int amdgpu_bo_create(struct amdgpu_device *adev,
551 struct amdgpu_bo_param *bp,
552 struct amdgpu_bo **bo_ptr)
554 u64 flags = bp->flags;
557 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
558 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
562 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && amdgpu_need_backup(adev)) {
564 WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
567 r = amdgpu_bo_create_shadow(adev, bp->size, bp->byte_align, (*bo_ptr));
570 reservation_object_unlock((*bo_ptr)->tbo.resv);
573 amdgpu_bo_unref(bo_ptr);
580 * amdgpu_bo_backup_to_shadow - Backs up an &amdgpu_bo buffer object
581 * @adev: amdgpu device object
582 * @ring: amdgpu_ring for the engine handling the buffer operations
583 * @bo: &amdgpu_bo buffer to be backed up
584 * @resv: reservation object with embedded fence
585 * @fence: dma_fence associated with the operation
586 * @direct: whether to submit the job directly
588 * Copies an &amdgpu_bo buffer object to its shadow object.
592 * 0 for success or a negative error code on failure.
594 int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
595 struct amdgpu_ring *ring,
596 struct amdgpu_bo *bo,
597 struct reservation_object *resv,
598 struct dma_fence **fence,
602 struct amdgpu_bo *shadow = bo->shadow;
603 uint64_t bo_addr, shadow_addr;
609 bo_addr = amdgpu_bo_gpu_offset(bo);
610 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
612 r = reservation_object_reserve_shared(bo->tbo.resv);
616 r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
617 amdgpu_bo_size(bo), resv, fence,
620 amdgpu_bo_fence(bo, *fence, true);
627 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
628 * @bo: pointer to the buffer object
630 * Sets placement according to domain; and changes placement and caching
631 * policy of the buffer object according to the placement.
632 * This is used for validating shadow bos. It calls ttm_bo_validate() to
633 * make sure the buffer is resident where it needs to be.
636 * 0 for success or a negative error code on failure.
638 int amdgpu_bo_validate(struct amdgpu_bo *bo)
640 struct ttm_operation_ctx ctx = { false, false };
647 domain = bo->preferred_domains;
650 amdgpu_ttm_placement_from_domain(bo, domain);
651 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
652 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
653 domain = bo->allowed_domains;
661 * amdgpu_bo_restore_from_shadow - restore an &amdgpu_bo buffer object
662 * @adev: amdgpu device object
663 * @ring: amdgpu_ring for the engine handling the buffer operations
664 * @bo: &amdgpu_bo buffer to be restored
665 * @resv: reservation object with embedded fence
666 * @fence: dma_fence associated with the operation
667 * @direct: whether to submit the job directly
669 * Copies a buffer object's shadow content back to the object.
670 * This is used for recovering a buffer from its shadow in case of a gpu
671 * reset where vram context may be lost.
674 * 0 for success or a negative error code on failure.
676 int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
677 struct amdgpu_ring *ring,
678 struct amdgpu_bo *bo,
679 struct reservation_object *resv,
680 struct dma_fence **fence,
684 struct amdgpu_bo *shadow = bo->shadow;
685 uint64_t bo_addr, shadow_addr;
691 bo_addr = amdgpu_bo_gpu_offset(bo);
692 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
694 r = reservation_object_reserve_shared(bo->tbo.resv);
698 r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
699 amdgpu_bo_size(bo), resv, fence,
702 amdgpu_bo_fence(bo, *fence, true);
709 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
710 * @bo: &amdgpu_bo buffer object to be mapped
711 * @ptr: kernel virtual address to be returned
713 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
714 * amdgpu_bo_kptr() to get the kernel virtual address.
717 * 0 for success or a negative error code on failure.
719 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
724 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
727 kptr = amdgpu_bo_kptr(bo);
734 r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
735 MAX_SCHEDULE_TIMEOUT);
739 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
744 *ptr = amdgpu_bo_kptr(bo);
750 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
751 * @bo: &amdgpu_bo buffer object
753 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
756 * the virtual address of a buffer object area.
758 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
762 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
766 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
767 * @bo: &amdgpu_bo buffer object to be unmapped
769 * Unmaps a kernel map set up by amdgpu_bo_kmap().
771 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
774 ttm_bo_kunmap(&bo->kmap);
778 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
779 * @bo: &amdgpu_bo buffer object
781 * References the contained &ttm_buffer_object.
784 * a refcounted pointer to the &amdgpu_bo buffer object.
786 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
791 ttm_bo_reference(&bo->tbo);
796 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
797 * @bo: &amdgpu_bo buffer object
799 * Unreferences the contained &ttm_buffer_object and clear the pointer
801 void amdgpu_bo_unref(struct amdgpu_bo **bo)
803 struct ttm_buffer_object *tbo;
815 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
816 * @bo: &amdgpu_bo buffer object to be pinned
817 * @domain: domain to be pinned to
818 * @min_offset: the start of requested address range
819 * @max_offset: the end of requested address range
820 * @gpu_addr: GPU offset of the &amdgpu_bo buffer object
822 * Pins the buffer object according to requested domain and address range. If
823 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
824 * pin_count and pin_size accordingly.
826 * Pinning means to lock pages in memory along with keeping them at a fixed
827 * offset. It is required when a buffer can not be moved, for example, when
828 * a display buffer is being scanned out.
830 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
831 * where to pin a buffer if there are specific restrictions on where a buffer
835 * 0 for success or a negative error code on failure.
837 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
838 u64 min_offset, u64 max_offset,
841 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
842 struct ttm_operation_ctx ctx = { false, false };
845 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
848 if (WARN_ON_ONCE(min_offset > max_offset))
851 /* A shared bo cannot be migrated to VRAM */
852 if (bo->prime_shared_count) {
853 if (domain & AMDGPU_GEM_DOMAIN_GTT)
854 domain = AMDGPU_GEM_DOMAIN_GTT;
859 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
860 * See function amdgpu_display_supported_domains()
862 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
865 uint32_t mem_type = bo->tbo.mem.mem_type;
867 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
872 *gpu_addr = amdgpu_bo_gpu_offset(bo);
874 if (max_offset != 0) {
875 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
876 WARN_ON_ONCE(max_offset <
877 (amdgpu_bo_gpu_offset(bo) - domain_start));
883 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
884 /* force to pin into visible video ram */
885 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
886 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
887 amdgpu_ttm_placement_from_domain(bo, domain);
888 for (i = 0; i < bo->placement.num_placement; i++) {
891 fpfn = min_offset >> PAGE_SHIFT;
892 lpfn = max_offset >> PAGE_SHIFT;
894 if (fpfn > bo->placements[i].fpfn)
895 bo->placements[i].fpfn = fpfn;
896 if (!bo->placements[i].lpfn ||
897 (lpfn && lpfn < bo->placements[i].lpfn))
898 bo->placements[i].lpfn = lpfn;
899 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
902 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
904 dev_err(adev->dev, "%p pin failed\n", bo);
908 r = amdgpu_ttm_alloc_gart(&bo->tbo);
910 dev_err(adev->dev, "%p bind failed\n", bo);
915 if (gpu_addr != NULL)
916 *gpu_addr = amdgpu_bo_gpu_offset(bo);
918 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
919 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
920 adev->vram_pin_size += amdgpu_bo_size(bo);
921 adev->invisible_pin_size += amdgpu_vram_mgr_bo_invisible_size(bo);
922 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
923 adev->gart_pin_size += amdgpu_bo_size(bo);
931 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
932 * @bo: &amdgpu_bo buffer object to be pinned
933 * @domain: domain to be pinned to
934 * @gpu_addr: GPU offset of the &amdgpu_bo buffer object
936 * A simple wrapper to amdgpu_bo_pin_restricted().
937 * Provides a simpler API for buffers that do not have any strict restrictions
938 * on where a buffer must be located.
941 * 0 for success or a negative error code on failure.
943 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain, u64 *gpu_addr)
945 return amdgpu_bo_pin_restricted(bo, domain, 0, 0, gpu_addr);
949 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
950 * @bo: &amdgpu_bo buffer object to be unpinned
952 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
953 * Changes placement and pin size accordingly.
956 * 0 for success or a negative error code on failure.
958 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
960 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
961 struct ttm_operation_ctx ctx = { false, false };
964 if (!bo->pin_count) {
965 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
972 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
973 adev->vram_pin_size -= amdgpu_bo_size(bo);
974 adev->invisible_pin_size -= amdgpu_vram_mgr_bo_invisible_size(bo);
975 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
976 adev->gart_pin_size -= amdgpu_bo_size(bo);
979 for (i = 0; i < bo->placement.num_placement; i++) {
980 bo->placements[i].lpfn = 0;
981 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
983 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
985 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
991 * amdgpu_bo_evict_vram - evict VRAM buffers
992 * @adev: amdgpu device object
994 * Evicts all VRAM buffers on the lru list of the memory type.
995 * Mainly used for evicting vram at suspend time.
998 * 0 for success or a negative error code on failure.
1000 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1002 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
1003 if (0 && (adev->flags & AMD_IS_APU)) {
1004 /* Useless to evict on IGP chips */
1007 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
1010 static const char *amdgpu_vram_names[] = {
1023 * amdgpu_bo_init - initialize memory manager
1024 * @adev: amdgpu device object
1026 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1029 * 0 for success or a negative error code on failure.
1031 int amdgpu_bo_init(struct amdgpu_device *adev)
1033 /* reserve PAT memory space to WC for VRAM */
1034 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1035 adev->gmc.aper_size);
1037 /* Add an MTRR for the VRAM */
1038 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1039 adev->gmc.aper_size);
1040 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1041 adev->gmc.mc_vram_size >> 20,
1042 (unsigned long long)adev->gmc.aper_size >> 20);
1043 DRM_INFO("RAM width %dbits %s\n",
1044 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1045 return amdgpu_ttm_init(adev);
1049 * amdgpu_bo_late_init - late init
1050 * @adev: amdgpu device object
1052 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1056 * 0 for success or a negative error code on failure.
1058 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1060 amdgpu_ttm_late_init(adev);
1066 * amdgpu_bo_fini - tear down memory manager
1067 * @adev: amdgpu device object
1069 * Reverses amdgpu_bo_init() to tear down memory manager.
1071 void amdgpu_bo_fini(struct amdgpu_device *adev)
1073 amdgpu_ttm_fini(adev);
1074 arch_phys_wc_del(adev->gmc.vram_mtrr);
1075 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1079 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1080 * @bo: &amdgpu_bo buffer object
1081 * @vma: vma as input from the fbdev mmap method
1083 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1086 * 0 for success or a negative error code on failure.
1088 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1089 struct vm_area_struct *vma)
1091 return ttm_fbdev_mmap(vma, &bo->tbo);
1095 * amdgpu_bo_set_tiling_flags - set tiling flags
1096 * @bo: &amdgpu_bo buffer object
1097 * @tiling_flags: new flags
1099 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1100 * kernel driver to set the tiling flags on a buffer.
1103 * 0 for success or a negative error code on failure.
1105 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1107 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1109 if (adev->family <= AMDGPU_FAMILY_CZ &&
1110 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1113 bo->tiling_flags = tiling_flags;
1118 * amdgpu_bo_get_tiling_flags - get tiling flags
1119 * @bo: &amdgpu_bo buffer object
1120 * @tiling_flags: returned flags
1122 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1123 * set the tiling flags on a buffer.
1125 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1127 lockdep_assert_held(&bo->tbo.resv->lock.base);
1130 *tiling_flags = bo->tiling_flags;
1134 * amdgpu_bo_set_metadata - set metadata
1135 * @bo: &amdgpu_bo buffer object
1136 * @metadata: new metadata
1137 * @metadata_size: size of the new metadata
1138 * @flags: flags of the new metadata
1140 * Sets buffer object's metadata, its size and flags.
1141 * Used via GEM ioctl.
1144 * 0 for success or a negative error code on failure.
1146 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1147 uint32_t metadata_size, uint64_t flags)
1151 if (!metadata_size) {
1152 if (bo->metadata_size) {
1153 kfree(bo->metadata);
1154 bo->metadata = NULL;
1155 bo->metadata_size = 0;
1160 if (metadata == NULL)
1163 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1167 kfree(bo->metadata);
1168 bo->metadata_flags = flags;
1169 bo->metadata = buffer;
1170 bo->metadata_size = metadata_size;
1176 * amdgpu_bo_get_metadata - get metadata
1177 * @bo: &amdgpu_bo buffer object
1178 * @buffer: returned metadata
1179 * @buffer_size: size of the buffer
1180 * @metadata_size: size of the returned metadata
1181 * @flags: flags of the returned metadata
1183 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1184 * less than metadata_size.
1185 * Used via GEM ioctl.
1188 * 0 for success or a negative error code on failure.
1190 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1191 size_t buffer_size, uint32_t *metadata_size,
1194 if (!buffer && !metadata_size)
1198 if (buffer_size < bo->metadata_size)
1201 if (bo->metadata_size)
1202 memcpy(buffer, bo->metadata, bo->metadata_size);
1206 *metadata_size = bo->metadata_size;
1208 *flags = bo->metadata_flags;
1214 * amdgpu_bo_move_notify - notification about a memory move
1215 * @bo: pointer to a buffer object
1216 * @evict: if this move is evicting the buffer from the graphics address space
1217 * @new_mem: new information of the bufer object
1219 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1221 * TTM driver callback which is called when ttm moves a buffer.
1223 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1225 struct ttm_mem_reg *new_mem)
1227 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1228 struct amdgpu_bo *abo;
1229 struct ttm_mem_reg *old_mem = &bo->mem;
1231 if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
1234 abo = ttm_to_amdgpu_bo(bo);
1235 amdgpu_vm_bo_invalidate(adev, abo, evict);
1237 amdgpu_bo_kunmap(abo);
1239 /* remember the eviction */
1241 atomic64_inc(&adev->num_evictions);
1243 /* update statistics */
1247 /* move_notify is called before move happens */
1248 trace_amdgpu_ttm_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1252 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1253 * @bo: pointer to a buffer object
1255 * Notifies the driver we are taking a fault on this BO and have reserved it,
1256 * also performs bookkeeping.
1257 * TTM driver callback for dealing with vm faults.
1260 * 0 for success or a negative error code on failure.
1262 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1264 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1265 struct ttm_operation_ctx ctx = { false, false };
1266 struct amdgpu_bo *abo;
1267 unsigned long offset, size;
1270 if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
1273 abo = ttm_to_amdgpu_bo(bo);
1275 /* Remember that this BO was accessed by the CPU */
1276 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1278 if (bo->mem.mem_type != TTM_PL_VRAM)
1281 size = bo->mem.num_pages << PAGE_SHIFT;
1282 offset = bo->mem.start << PAGE_SHIFT;
1283 if ((offset + size) <= adev->gmc.visible_vram_size)
1286 /* Can't move a pinned BO to visible VRAM */
1287 if (abo->pin_count > 0)
1290 /* hurrah the memory is not visible ! */
1291 atomic64_inc(&adev->num_vram_cpu_page_faults);
1292 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1293 AMDGPU_GEM_DOMAIN_GTT);
1295 /* Avoid costly evictions; only set GTT as a busy placement */
1296 abo->placement.num_busy_placement = 1;
1297 abo->placement.busy_placement = &abo->placements[1];
1299 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1300 if (unlikely(r != 0))
1303 offset = bo->mem.start << PAGE_SHIFT;
1304 /* this should never happen */
1305 if (bo->mem.mem_type == TTM_PL_VRAM &&
1306 (offset + size) > adev->gmc.visible_vram_size)
1313 * amdgpu_bo_fence - add fence to buffer object
1315 * @bo: buffer object in question
1316 * @fence: fence to add
1317 * @shared: true if fence should be added shared
1320 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1323 struct reservation_object *resv = bo->tbo.resv;
1326 reservation_object_add_shared_fence(resv, fence);
1328 reservation_object_add_excl_fence(resv, fence);
1332 * amdgpu_bo_gpu_offset - return GPU offset of bo
1333 * @bo: amdgpu object for which we query the offset
1335 * Note: object should either be pinned or reserved when calling this
1336 * function, it might be useful to add check for this for debugging.
1339 * current GPU offset of the object.
1341 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1343 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1344 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_TT &&
1345 !amdgpu_gtt_mgr_has_gart_addr(&bo->tbo.mem));
1346 WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1348 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1349 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1350 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1352 return bo->tbo.offset;
1356 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1357 * @adev: amdgpu device object
1358 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1361 * Which of the allowed domains is preferred for pinning the BO for scanout.
1363 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1366 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1367 domain = AMDGPU_GEM_DOMAIN_VRAM;
1368 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1369 domain = AMDGPU_GEM_DOMAIN_GTT;