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);
84 amdgpu_amdkfd_unreserve_memory_limit(bo);
88 if (bo->gem_base.import_attach)
89 drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
90 drm_gem_object_release(&bo->gem_base);
91 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
92 if (!list_empty(&bo->shadow_list)) {
93 mutex_lock(&adev->shadow_list_lock);
94 list_del_init(&bo->shadow_list);
95 mutex_unlock(&adev->shadow_list_lock);
97 amdgpu_bo_unref(&bo->parent);
104 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
105 * @bo: buffer object to be checked
107 * Uses destroy function associated with the object to determine if this is
111 * true if the object belongs to &amdgpu_bo, false if not.
113 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
115 if (bo->destroy == &amdgpu_bo_destroy)
121 * amdgpu_bo_placement_from_domain - set buffer's placement
122 * @abo: &amdgpu_bo buffer object whose placement is to be set
123 * @domain: requested domain
125 * Sets buffer's placement according to requested domain and the buffer's
128 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
130 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
131 struct ttm_placement *placement = &abo->placement;
132 struct ttm_place *places = abo->placements;
133 u64 flags = abo->flags;
136 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
137 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
141 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
144 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
145 places[c].lpfn = visible_pfn;
147 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
149 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
150 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
154 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
157 places[c].flags = TTM_PL_FLAG_TT;
158 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
159 places[c].flags |= TTM_PL_FLAG_WC |
160 TTM_PL_FLAG_UNCACHED;
162 places[c].flags |= TTM_PL_FLAG_CACHED;
166 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
169 places[c].flags = TTM_PL_FLAG_SYSTEM;
170 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
171 places[c].flags |= TTM_PL_FLAG_WC |
172 TTM_PL_FLAG_UNCACHED;
174 places[c].flags |= TTM_PL_FLAG_CACHED;
178 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
181 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
185 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
188 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
192 if (domain & AMDGPU_GEM_DOMAIN_OA) {
195 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
202 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
206 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
208 placement->num_placement = c;
209 placement->placement = places;
211 placement->num_busy_placement = c;
212 placement->busy_placement = places;
216 * amdgpu_bo_create_reserved - create reserved BO for kernel use
218 * @adev: amdgpu device object
219 * @size: size for the new BO
220 * @align: alignment for the new BO
221 * @domain: where to place it
222 * @bo_ptr: used to initialize BOs in structures
223 * @gpu_addr: GPU addr of the pinned BO
224 * @cpu_addr: optional CPU address mapping
226 * Allocates and pins a BO for kernel internal use, and returns it still
229 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
232 * 0 on success, negative error code otherwise.
234 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
235 unsigned long size, int align,
236 u32 domain, struct amdgpu_bo **bo_ptr,
237 u64 *gpu_addr, void **cpu_addr)
239 struct amdgpu_bo_param bp;
244 amdgpu_bo_unref(bo_ptr);
248 memset(&bp, 0, sizeof(bp));
250 bp.byte_align = align;
252 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
253 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
254 bp.type = ttm_bo_type_kernel;
258 r = amdgpu_bo_create(adev, &bp, bo_ptr);
260 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
267 r = amdgpu_bo_reserve(*bo_ptr, false);
269 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
273 r = amdgpu_bo_pin(*bo_ptr, domain);
275 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
276 goto error_unreserve;
279 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
281 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
286 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
289 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
291 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
299 amdgpu_bo_unpin(*bo_ptr);
301 amdgpu_bo_unreserve(*bo_ptr);
305 amdgpu_bo_unref(bo_ptr);
311 * amdgpu_bo_create_kernel - create BO for kernel use
313 * @adev: amdgpu device object
314 * @size: size for the new BO
315 * @align: alignment for the new BO
316 * @domain: where to place it
317 * @bo_ptr: used to initialize BOs in structures
318 * @gpu_addr: GPU addr of the pinned BO
319 * @cpu_addr: optional CPU address mapping
321 * Allocates and pins a BO for kernel internal use.
323 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
326 * 0 on success, negative error code otherwise.
328 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
329 unsigned long size, int align,
330 u32 domain, struct amdgpu_bo **bo_ptr,
331 u64 *gpu_addr, void **cpu_addr)
335 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
342 amdgpu_bo_unreserve(*bo_ptr);
348 * amdgpu_bo_free_kernel - free BO for kernel use
350 * @bo: amdgpu BO to free
351 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
352 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
354 * unmaps and unpin a BO for kernel internal use.
356 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
362 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
364 amdgpu_bo_kunmap(*bo);
366 amdgpu_bo_unpin(*bo);
367 amdgpu_bo_unreserve(*bo);
378 /* Validate bo size is bit bigger then the request domain */
379 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
380 unsigned long size, u32 domain)
382 struct ttm_mem_type_manager *man = NULL;
385 * If GTT is part of requested domains the check must succeed to
386 * allow fall back to GTT
388 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
389 man = &adev->mman.bdev.man[TTM_PL_TT];
391 if (size < (man->size << PAGE_SHIFT))
397 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
398 man = &adev->mman.bdev.man[TTM_PL_VRAM];
400 if (size < (man->size << PAGE_SHIFT))
407 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
411 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
412 man->size << PAGE_SHIFT);
416 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
417 struct amdgpu_bo_param *bp,
418 struct amdgpu_bo **bo_ptr)
420 struct ttm_operation_ctx ctx = {
421 .interruptible = (bp->type != ttm_bo_type_kernel),
422 .no_wait_gpu = false,
424 .flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
426 struct amdgpu_bo *bo;
427 unsigned long page_align, size = bp->size;
431 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
432 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
433 /* GWS and OA don't need any alignment. */
434 page_align = bp->byte_align;
436 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
437 /* Both size and alignment must be a multiple of 4. */
438 page_align = ALIGN(bp->byte_align, 4);
439 size = ALIGN(size, 4) << PAGE_SHIFT;
441 /* Memory should be aligned at least to a page size. */
442 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
443 size = ALIGN(size, PAGE_SIZE);
446 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
451 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
452 sizeof(struct amdgpu_bo));
454 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
457 drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
458 INIT_LIST_HEAD(&bo->shadow_list);
460 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
462 bo->allowed_domains = bo->preferred_domains;
463 if (bp->type != ttm_bo_type_kernel &&
464 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
465 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
467 bo->flags = bp->flags;
470 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
471 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
473 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
474 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
475 /* Don't try to enable write-combining when it can't work, or things
477 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
480 #ifndef CONFIG_COMPILE_TEST
481 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
482 thanks to write-combining
485 if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
486 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
487 "better performance thanks to write-combining\n");
488 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
490 /* For architectures that don't support WC memory,
491 * mask out the WC flag from the BO
493 if (!drm_arch_can_wc_memory())
494 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
497 bo->tbo.bdev = &adev->mman.bdev;
498 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
499 AMDGPU_GEM_DOMAIN_GDS))
500 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
502 amdgpu_bo_placement_from_domain(bo, bp->domain);
503 if (bp->type == ttm_bo_type_kernel)
504 bo->tbo.priority = 1;
506 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
507 &bo->placement, page_align, &ctx, acc_size,
508 NULL, bp->resv, &amdgpu_bo_destroy);
509 if (unlikely(r != 0))
512 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
513 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
514 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
515 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
518 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
520 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
521 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
522 struct dma_fence *fence;
524 r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
528 amdgpu_bo_fence(bo, fence, false);
529 dma_fence_put(bo->tbo.moving);
530 bo->tbo.moving = dma_fence_get(fence);
531 dma_fence_put(fence);
534 amdgpu_bo_unreserve(bo);
537 trace_amdgpu_bo_create(bo);
539 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
540 if (bp->type == ttm_bo_type_device)
541 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
547 ww_mutex_unlock(&bo->tbo.resv->lock);
548 amdgpu_bo_unref(&bo);
552 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
554 struct amdgpu_bo *bo)
556 struct amdgpu_bo_param bp;
562 memset(&bp, 0, sizeof(bp));
564 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
565 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
566 AMDGPU_GEM_CREATE_SHADOW;
567 bp.type = ttm_bo_type_kernel;
568 bp.resv = bo->tbo.resv;
570 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
572 bo->shadow->parent = amdgpu_bo_ref(bo);
573 mutex_lock(&adev->shadow_list_lock);
574 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
575 mutex_unlock(&adev->shadow_list_lock);
582 * amdgpu_bo_create - create an &amdgpu_bo buffer object
583 * @adev: amdgpu device object
584 * @bp: parameters to be used for the buffer object
585 * @bo_ptr: pointer to the buffer object pointer
587 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
589 * Shadow object is used to backup the original buffer object, and is always
593 * 0 for success or a negative error code on failure.
595 int amdgpu_bo_create(struct amdgpu_device *adev,
596 struct amdgpu_bo_param *bp,
597 struct amdgpu_bo **bo_ptr)
599 u64 flags = bp->flags;
602 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
603 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
607 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
609 WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
612 r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
615 reservation_object_unlock((*bo_ptr)->tbo.resv);
618 amdgpu_bo_unref(bo_ptr);
625 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
626 * @bo: pointer to the buffer object
628 * Sets placement according to domain; and changes placement and caching
629 * policy of the buffer object according to the placement.
630 * This is used for validating shadow bos. It calls ttm_bo_validate() to
631 * make sure the buffer is resident where it needs to be.
634 * 0 for success or a negative error code on failure.
636 int amdgpu_bo_validate(struct amdgpu_bo *bo)
638 struct ttm_operation_ctx ctx = { false, false };
645 domain = bo->preferred_domains;
648 amdgpu_bo_placement_from_domain(bo, domain);
649 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
650 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
651 domain = bo->allowed_domains;
659 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
661 * @shadow: &amdgpu_bo shadow to be restored
662 * @fence: dma_fence associated with the operation
664 * Copies a buffer object's shadow content back to the object.
665 * This is used for recovering a buffer from its shadow in case of a gpu
666 * reset where vram context may be lost.
669 * 0 for success or a negative error code on failure.
671 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
674 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
675 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
676 uint64_t shadow_addr, parent_addr;
678 shadow_addr = amdgpu_bo_gpu_offset(shadow);
679 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
681 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
682 amdgpu_bo_size(shadow), NULL, fence,
687 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
688 * @bo: &amdgpu_bo buffer object to be mapped
689 * @ptr: kernel virtual address to be returned
691 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
692 * amdgpu_bo_kptr() to get the kernel virtual address.
695 * 0 for success or a negative error code on failure.
697 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
702 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
705 kptr = amdgpu_bo_kptr(bo);
712 r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
713 MAX_SCHEDULE_TIMEOUT);
717 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
722 *ptr = amdgpu_bo_kptr(bo);
728 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
729 * @bo: &amdgpu_bo buffer object
731 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
734 * the virtual address of a buffer object area.
736 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
740 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
744 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
745 * @bo: &amdgpu_bo buffer object to be unmapped
747 * Unmaps a kernel map set up by amdgpu_bo_kmap().
749 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
752 ttm_bo_kunmap(&bo->kmap);
756 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
757 * @bo: &amdgpu_bo buffer object
759 * References the contained &ttm_buffer_object.
762 * a refcounted pointer to the &amdgpu_bo buffer object.
764 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
769 ttm_bo_get(&bo->tbo);
774 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
775 * @bo: &amdgpu_bo buffer object
777 * Unreferences the contained &ttm_buffer_object and clear the pointer
779 void amdgpu_bo_unref(struct amdgpu_bo **bo)
781 struct ttm_buffer_object *tbo;
792 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
793 * @bo: &amdgpu_bo buffer object to be pinned
794 * @domain: domain to be pinned to
795 * @min_offset: the start of requested address range
796 * @max_offset: the end of requested address range
798 * Pins the buffer object according to requested domain and address range. If
799 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
800 * pin_count and pin_size accordingly.
802 * Pinning means to lock pages in memory along with keeping them at a fixed
803 * offset. It is required when a buffer can not be moved, for example, when
804 * a display buffer is being scanned out.
806 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
807 * where to pin a buffer if there are specific restrictions on where a buffer
811 * 0 for success or a negative error code on failure.
813 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
814 u64 min_offset, u64 max_offset)
816 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
817 struct ttm_operation_ctx ctx = { false, false };
820 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
823 if (WARN_ON_ONCE(min_offset > max_offset))
826 /* A shared bo cannot be migrated to VRAM */
827 if (bo->prime_shared_count) {
828 if (domain & AMDGPU_GEM_DOMAIN_GTT)
829 domain = AMDGPU_GEM_DOMAIN_GTT;
834 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
835 * See function amdgpu_display_supported_domains()
837 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
840 uint32_t mem_type = bo->tbo.mem.mem_type;
842 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
847 if (max_offset != 0) {
848 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
849 WARN_ON_ONCE(max_offset <
850 (amdgpu_bo_gpu_offset(bo) - domain_start));
856 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
857 /* force to pin into visible video ram */
858 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
859 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
860 amdgpu_bo_placement_from_domain(bo, domain);
861 for (i = 0; i < bo->placement.num_placement; i++) {
864 fpfn = min_offset >> PAGE_SHIFT;
865 lpfn = max_offset >> PAGE_SHIFT;
867 if (fpfn > bo->placements[i].fpfn)
868 bo->placements[i].fpfn = fpfn;
869 if (!bo->placements[i].lpfn ||
870 (lpfn && lpfn < bo->placements[i].lpfn))
871 bo->placements[i].lpfn = lpfn;
872 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
875 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
877 dev_err(adev->dev, "%p pin failed\n", bo);
883 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
884 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
885 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
886 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
887 &adev->visible_pin_size);
888 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
889 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
897 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
898 * @bo: &amdgpu_bo buffer object to be pinned
899 * @domain: domain to be pinned to
901 * A simple wrapper to amdgpu_bo_pin_restricted().
902 * Provides a simpler API for buffers that do not have any strict restrictions
903 * on where a buffer must be located.
906 * 0 for success or a negative error code on failure.
908 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
910 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
914 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
915 * @bo: &amdgpu_bo buffer object to be unpinned
917 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
918 * Changes placement and pin size accordingly.
921 * 0 for success or a negative error code on failure.
923 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
925 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
926 struct ttm_operation_ctx ctx = { false, false };
929 if (WARN_ON_ONCE(!bo->pin_count)) {
930 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
937 amdgpu_bo_subtract_pin_size(bo);
939 for (i = 0; i < bo->placement.num_placement; i++) {
940 bo->placements[i].lpfn = 0;
941 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
943 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
945 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
951 * amdgpu_bo_evict_vram - evict VRAM buffers
952 * @adev: amdgpu device object
954 * Evicts all VRAM buffers on the lru list of the memory type.
955 * Mainly used for evicting vram at suspend time.
958 * 0 for success or a negative error code on failure.
960 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
962 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
963 #ifndef CONFIG_HIBERNATION
964 if (adev->flags & AMD_IS_APU) {
965 /* Useless to evict on IGP chips */
969 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
972 static const char *amdgpu_vram_names[] = {
986 * amdgpu_bo_init - initialize memory manager
987 * @adev: amdgpu device object
989 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
992 * 0 for success or a negative error code on failure.
994 int amdgpu_bo_init(struct amdgpu_device *adev)
996 /* reserve PAT memory space to WC for VRAM */
997 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
998 adev->gmc.aper_size);
1000 /* Add an MTRR for the VRAM */
1001 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1002 adev->gmc.aper_size);
1003 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1004 adev->gmc.mc_vram_size >> 20,
1005 (unsigned long long)adev->gmc.aper_size >> 20);
1006 DRM_INFO("RAM width %dbits %s\n",
1007 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1008 return amdgpu_ttm_init(adev);
1012 * amdgpu_bo_late_init - late init
1013 * @adev: amdgpu device object
1015 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1019 * 0 for success or a negative error code on failure.
1021 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1023 amdgpu_ttm_late_init(adev);
1029 * amdgpu_bo_fini - tear down memory manager
1030 * @adev: amdgpu device object
1032 * Reverses amdgpu_bo_init() to tear down memory manager.
1034 void amdgpu_bo_fini(struct amdgpu_device *adev)
1036 amdgpu_ttm_fini(adev);
1037 arch_phys_wc_del(adev->gmc.vram_mtrr);
1038 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1042 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1043 * @bo: &amdgpu_bo buffer object
1044 * @vma: vma as input from the fbdev mmap method
1046 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1049 * 0 for success or a negative error code on failure.
1051 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1052 struct vm_area_struct *vma)
1054 return ttm_fbdev_mmap(vma, &bo->tbo);
1058 * amdgpu_bo_set_tiling_flags - set tiling flags
1059 * @bo: &amdgpu_bo buffer object
1060 * @tiling_flags: new flags
1062 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1063 * kernel driver to set the tiling flags on a buffer.
1066 * 0 for success or a negative error code on failure.
1068 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1070 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1072 if (adev->family <= AMDGPU_FAMILY_CZ &&
1073 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1076 bo->tiling_flags = tiling_flags;
1081 * amdgpu_bo_get_tiling_flags - get tiling flags
1082 * @bo: &amdgpu_bo buffer object
1083 * @tiling_flags: returned flags
1085 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1086 * set the tiling flags on a buffer.
1088 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1090 lockdep_assert_held(&bo->tbo.resv->lock.base);
1093 *tiling_flags = bo->tiling_flags;
1097 * amdgpu_bo_set_metadata - set metadata
1098 * @bo: &amdgpu_bo buffer object
1099 * @metadata: new metadata
1100 * @metadata_size: size of the new metadata
1101 * @flags: flags of the new metadata
1103 * Sets buffer object's metadata, its size and flags.
1104 * Used via GEM ioctl.
1107 * 0 for success or a negative error code on failure.
1109 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1110 uint32_t metadata_size, uint64_t flags)
1114 if (!metadata_size) {
1115 if (bo->metadata_size) {
1116 kfree(bo->metadata);
1117 bo->metadata = NULL;
1118 bo->metadata_size = 0;
1123 if (metadata == NULL)
1126 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1130 kfree(bo->metadata);
1131 bo->metadata_flags = flags;
1132 bo->metadata = buffer;
1133 bo->metadata_size = metadata_size;
1139 * amdgpu_bo_get_metadata - get metadata
1140 * @bo: &amdgpu_bo buffer object
1141 * @buffer: returned metadata
1142 * @buffer_size: size of the buffer
1143 * @metadata_size: size of the returned metadata
1144 * @flags: flags of the returned metadata
1146 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1147 * less than metadata_size.
1148 * Used via GEM ioctl.
1151 * 0 for success or a negative error code on failure.
1153 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1154 size_t buffer_size, uint32_t *metadata_size,
1157 if (!buffer && !metadata_size)
1161 if (buffer_size < bo->metadata_size)
1164 if (bo->metadata_size)
1165 memcpy(buffer, bo->metadata, bo->metadata_size);
1169 *metadata_size = bo->metadata_size;
1171 *flags = bo->metadata_flags;
1177 * amdgpu_bo_move_notify - notification about a memory move
1178 * @bo: pointer to a buffer object
1179 * @evict: if this move is evicting the buffer from the graphics address space
1180 * @new_mem: new information of the bufer object
1182 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1184 * TTM driver callback which is called when ttm moves a buffer.
1186 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1188 struct ttm_mem_reg *new_mem)
1190 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1191 struct amdgpu_bo *abo;
1192 struct ttm_mem_reg *old_mem = &bo->mem;
1194 if (!amdgpu_bo_is_amdgpu_bo(bo))
1197 abo = ttm_to_amdgpu_bo(bo);
1198 amdgpu_vm_bo_invalidate(adev, abo, evict);
1200 amdgpu_bo_kunmap(abo);
1202 /* remember the eviction */
1204 atomic64_inc(&adev->num_evictions);
1206 /* update statistics */
1210 /* move_notify is called before move happens */
1211 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1215 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1216 * @bo: pointer to a buffer object
1218 * Notifies the driver we are taking a fault on this BO and have reserved it,
1219 * also performs bookkeeping.
1220 * TTM driver callback for dealing with vm faults.
1223 * 0 for success or a negative error code on failure.
1225 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1227 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1228 struct ttm_operation_ctx ctx = { false, false };
1229 struct amdgpu_bo *abo;
1230 unsigned long offset, size;
1233 if (!amdgpu_bo_is_amdgpu_bo(bo))
1236 abo = ttm_to_amdgpu_bo(bo);
1238 /* Remember that this BO was accessed by the CPU */
1239 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1241 if (bo->mem.mem_type != TTM_PL_VRAM)
1244 size = bo->mem.num_pages << PAGE_SHIFT;
1245 offset = bo->mem.start << PAGE_SHIFT;
1246 if ((offset + size) <= adev->gmc.visible_vram_size)
1249 /* Can't move a pinned BO to visible VRAM */
1250 if (abo->pin_count > 0)
1253 /* hurrah the memory is not visible ! */
1254 atomic64_inc(&adev->num_vram_cpu_page_faults);
1255 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1256 AMDGPU_GEM_DOMAIN_GTT);
1258 /* Avoid costly evictions; only set GTT as a busy placement */
1259 abo->placement.num_busy_placement = 1;
1260 abo->placement.busy_placement = &abo->placements[1];
1262 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1263 if (unlikely(r != 0))
1266 offset = bo->mem.start << PAGE_SHIFT;
1267 /* this should never happen */
1268 if (bo->mem.mem_type == TTM_PL_VRAM &&
1269 (offset + size) > adev->gmc.visible_vram_size)
1276 * amdgpu_bo_fence - add fence to buffer object
1278 * @bo: buffer object in question
1279 * @fence: fence to add
1280 * @shared: true if fence should be added shared
1283 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1286 struct reservation_object *resv = bo->tbo.resv;
1289 reservation_object_add_shared_fence(resv, fence);
1291 reservation_object_add_excl_fence(resv, fence);
1295 * amdgpu_sync_wait_resv - Wait for BO reservation fences
1297 * @bo: buffer object
1298 * @owner: fence owner
1299 * @intr: Whether the wait is interruptible
1302 * 0 on success, errno otherwise.
1304 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1306 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1307 struct amdgpu_sync sync;
1310 amdgpu_sync_create(&sync);
1311 amdgpu_sync_resv(adev, &sync, bo->tbo.resv, owner, false);
1312 r = amdgpu_sync_wait(&sync, intr);
1313 amdgpu_sync_free(&sync);
1319 * amdgpu_bo_gpu_offset - return GPU offset of bo
1320 * @bo: amdgpu object for which we query the offset
1322 * Note: object should either be pinned or reserved when calling this
1323 * function, it might be useful to add check for this for debugging.
1326 * current GPU offset of the object.
1328 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1330 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1331 WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1332 !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1333 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1334 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1335 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1337 return amdgpu_gmc_sign_extend(bo->tbo.offset);
1341 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1342 * @adev: amdgpu device object
1343 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1346 * Which of the allowed domains is preferred for pinning the BO for scanout.
1348 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1351 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1352 domain = AMDGPU_GEM_DOMAIN_VRAM;
1353 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1354 domain = AMDGPU_GEM_DOMAIN_GTT;