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_system_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 amdgpu_bo_unref(&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);
102 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
103 * @bo: buffer object to be checked
105 * Uses destroy function associated with the object to determine if this is
109 * true if the object belongs to &amdgpu_bo, false if not.
111 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
113 if (bo->destroy == &amdgpu_bo_destroy)
119 * amdgpu_bo_placement_from_domain - set buffer's placement
120 * @abo: &amdgpu_bo buffer object whose placement is to be set
121 * @domain: requested domain
123 * Sets buffer's placement according to requested domain and the buffer's
126 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
128 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
129 struct ttm_placement *placement = &abo->placement;
130 struct ttm_place *places = abo->placements;
131 u64 flags = abo->flags;
134 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
135 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
139 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
142 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
143 places[c].lpfn = visible_pfn;
145 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
147 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
148 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
152 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
155 places[c].flags = TTM_PL_FLAG_TT;
156 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
157 places[c].flags |= TTM_PL_FLAG_WC |
158 TTM_PL_FLAG_UNCACHED;
160 places[c].flags |= TTM_PL_FLAG_CACHED;
164 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
167 places[c].flags = TTM_PL_FLAG_SYSTEM;
168 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
169 places[c].flags |= TTM_PL_FLAG_WC |
170 TTM_PL_FLAG_UNCACHED;
172 places[c].flags |= TTM_PL_FLAG_CACHED;
176 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
179 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
183 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
186 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
190 if (domain & AMDGPU_GEM_DOMAIN_OA) {
193 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
200 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
204 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
206 placement->num_placement = c;
207 placement->placement = places;
209 placement->num_busy_placement = c;
210 placement->busy_placement = places;
214 * amdgpu_bo_create_reserved - create reserved BO for kernel use
216 * @adev: amdgpu device object
217 * @size: size for the new BO
218 * @align: alignment for the new BO
219 * @domain: where to place it
220 * @bo_ptr: used to initialize BOs in structures
221 * @gpu_addr: GPU addr of the pinned BO
222 * @cpu_addr: optional CPU address mapping
224 * Allocates and pins a BO for kernel internal use, and returns it still
227 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
230 * 0 on success, negative error code otherwise.
232 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
233 unsigned long size, int align,
234 u32 domain, struct amdgpu_bo **bo_ptr,
235 u64 *gpu_addr, void **cpu_addr)
237 struct amdgpu_bo_param bp;
242 amdgpu_bo_unref(bo_ptr);
246 memset(&bp, 0, sizeof(bp));
248 bp.byte_align = align;
250 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
251 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 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
415 struct amdgpu_bo_param *bp,
416 struct amdgpu_bo **bo_ptr)
418 struct ttm_operation_ctx ctx = {
419 .interruptible = (bp->type != ttm_bo_type_kernel),
420 .no_wait_gpu = false,
422 .flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
424 struct amdgpu_bo *bo;
425 unsigned long page_align, size = bp->size;
429 page_align = roundup(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
430 size = ALIGN(size, PAGE_SIZE);
432 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
437 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
438 sizeof(struct amdgpu_bo));
440 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
443 drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
444 INIT_LIST_HEAD(&bo->shadow_list);
446 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
448 bo->allowed_domains = bo->preferred_domains;
449 if (bp->type != ttm_bo_type_kernel &&
450 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
451 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
453 bo->flags = bp->flags;
456 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
457 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
459 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
460 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
461 /* Don't try to enable write-combining when it can't work, or things
463 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
466 #ifndef CONFIG_COMPILE_TEST
467 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
468 thanks to write-combining
471 if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
472 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
473 "better performance thanks to write-combining\n");
474 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
476 /* For architectures that don't support WC memory,
477 * mask out the WC flag from the BO
479 if (!drm_arch_can_wc_memory())
480 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
483 bo->tbo.bdev = &adev->mman.bdev;
484 amdgpu_bo_placement_from_domain(bo, bp->domain);
485 if (bp->type == ttm_bo_type_kernel)
486 bo->tbo.priority = 1;
488 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
489 &bo->placement, page_align, &ctx, acc_size,
490 NULL, bp->resv, &amdgpu_bo_destroy);
491 if (unlikely(r != 0))
494 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
495 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
496 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
497 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
500 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
502 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
503 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
504 struct dma_fence *fence;
506 r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
510 amdgpu_bo_fence(bo, fence, false);
511 dma_fence_put(bo->tbo.moving);
512 bo->tbo.moving = dma_fence_get(fence);
513 dma_fence_put(fence);
516 amdgpu_bo_unreserve(bo);
519 trace_amdgpu_bo_create(bo);
521 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
522 if (bp->type == ttm_bo_type_device)
523 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
529 ww_mutex_unlock(&bo->tbo.resv->lock);
530 amdgpu_bo_unref(&bo);
534 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
535 unsigned long size, int byte_align,
536 struct amdgpu_bo *bo)
538 struct amdgpu_bo_param bp;
544 memset(&bp, 0, sizeof(bp));
546 bp.byte_align = byte_align;
547 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
548 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
549 AMDGPU_GEM_CREATE_SHADOW;
550 bp.type = ttm_bo_type_kernel;
551 bp.resv = bo->tbo.resv;
553 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
555 bo->shadow->parent = amdgpu_bo_ref(bo);
556 mutex_lock(&adev->shadow_list_lock);
557 list_add_tail(&bo->shadow_list, &adev->shadow_list);
558 mutex_unlock(&adev->shadow_list_lock);
565 * amdgpu_bo_create - create an &amdgpu_bo buffer object
566 * @adev: amdgpu device object
567 * @bp: parameters to be used for the buffer object
568 * @bo_ptr: pointer to the buffer object pointer
570 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
572 * Shadow object is used to backup the original buffer object, and is always
576 * 0 for success or a negative error code on failure.
578 int amdgpu_bo_create(struct amdgpu_device *adev,
579 struct amdgpu_bo_param *bp,
580 struct amdgpu_bo **bo_ptr)
582 u64 flags = bp->flags;
585 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
586 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
590 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
592 WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
595 r = amdgpu_bo_create_shadow(adev, bp->size, bp->byte_align, (*bo_ptr));
598 reservation_object_unlock((*bo_ptr)->tbo.resv);
601 amdgpu_bo_unref(bo_ptr);
608 * amdgpu_bo_backup_to_shadow - Backs up an &amdgpu_bo buffer object
609 * @adev: amdgpu device object
610 * @ring: amdgpu_ring for the engine handling the buffer operations
611 * @bo: &amdgpu_bo buffer to be backed up
612 * @resv: reservation object with embedded fence
613 * @fence: dma_fence associated with the operation
614 * @direct: whether to submit the job directly
616 * Copies an &amdgpu_bo buffer object to its shadow object.
620 * 0 for success or a negative error code on failure.
622 int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
623 struct amdgpu_ring *ring,
624 struct amdgpu_bo *bo,
625 struct reservation_object *resv,
626 struct dma_fence **fence,
630 struct amdgpu_bo *shadow = bo->shadow;
631 uint64_t bo_addr, shadow_addr;
637 bo_addr = amdgpu_bo_gpu_offset(bo);
638 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
640 r = reservation_object_reserve_shared(bo->tbo.resv);
644 r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
645 amdgpu_bo_size(bo), resv, fence,
648 amdgpu_bo_fence(bo, *fence, true);
655 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
656 * @bo: pointer to the buffer object
658 * Sets placement according to domain; and changes placement and caching
659 * policy of the buffer object according to the placement.
660 * This is used for validating shadow bos. It calls ttm_bo_validate() to
661 * make sure the buffer is resident where it needs to be.
664 * 0 for success or a negative error code on failure.
666 int amdgpu_bo_validate(struct amdgpu_bo *bo)
668 struct ttm_operation_ctx ctx = { false, false };
675 domain = bo->preferred_domains;
678 amdgpu_bo_placement_from_domain(bo, domain);
679 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
680 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
681 domain = bo->allowed_domains;
689 * amdgpu_bo_restore_from_shadow - restore an &amdgpu_bo buffer object
690 * @adev: amdgpu device object
691 * @ring: amdgpu_ring for the engine handling the buffer operations
692 * @bo: &amdgpu_bo buffer to be restored
693 * @resv: reservation object with embedded fence
694 * @fence: dma_fence associated with the operation
695 * @direct: whether to submit the job directly
697 * Copies a buffer object's shadow content back to the object.
698 * This is used for recovering a buffer from its shadow in case of a gpu
699 * reset where vram context may be lost.
702 * 0 for success or a negative error code on failure.
704 int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
705 struct amdgpu_ring *ring,
706 struct amdgpu_bo *bo,
707 struct reservation_object *resv,
708 struct dma_fence **fence,
712 struct amdgpu_bo *shadow = bo->shadow;
713 uint64_t bo_addr, shadow_addr;
719 bo_addr = amdgpu_bo_gpu_offset(bo);
720 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
722 r = reservation_object_reserve_shared(bo->tbo.resv);
726 r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
727 amdgpu_bo_size(bo), resv, fence,
730 amdgpu_bo_fence(bo, *fence, true);
737 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
738 * @bo: &amdgpu_bo buffer object to be mapped
739 * @ptr: kernel virtual address to be returned
741 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
742 * amdgpu_bo_kptr() to get the kernel virtual address.
745 * 0 for success or a negative error code on failure.
747 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
752 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
755 kptr = amdgpu_bo_kptr(bo);
762 r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
763 MAX_SCHEDULE_TIMEOUT);
767 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
772 *ptr = amdgpu_bo_kptr(bo);
778 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
779 * @bo: &amdgpu_bo buffer object
781 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
784 * the virtual address of a buffer object area.
786 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
790 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
794 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
795 * @bo: &amdgpu_bo buffer object to be unmapped
797 * Unmaps a kernel map set up by amdgpu_bo_kmap().
799 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
802 ttm_bo_kunmap(&bo->kmap);
806 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
807 * @bo: &amdgpu_bo buffer object
809 * References the contained &ttm_buffer_object.
812 * a refcounted pointer to the &amdgpu_bo buffer object.
814 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
819 ttm_bo_get(&bo->tbo);
824 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
825 * @bo: &amdgpu_bo buffer object
827 * Unreferences the contained &ttm_buffer_object and clear the pointer
829 void amdgpu_bo_unref(struct amdgpu_bo **bo)
831 struct ttm_buffer_object *tbo;
842 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
843 * @bo: &amdgpu_bo buffer object to be pinned
844 * @domain: domain to be pinned to
845 * @min_offset: the start of requested address range
846 * @max_offset: the end of requested address range
848 * Pins the buffer object according to requested domain and address range. If
849 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
850 * pin_count and pin_size accordingly.
852 * Pinning means to lock pages in memory along with keeping them at a fixed
853 * offset. It is required when a buffer can not be moved, for example, when
854 * a display buffer is being scanned out.
856 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
857 * where to pin a buffer if there are specific restrictions on where a buffer
861 * 0 for success or a negative error code on failure.
863 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
864 u64 min_offset, u64 max_offset)
866 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
867 struct ttm_operation_ctx ctx = { false, false };
870 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
873 if (WARN_ON_ONCE(min_offset > max_offset))
876 /* A shared bo cannot be migrated to VRAM */
877 if (bo->prime_shared_count) {
878 if (domain & AMDGPU_GEM_DOMAIN_GTT)
879 domain = AMDGPU_GEM_DOMAIN_GTT;
884 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
885 * See function amdgpu_display_supported_domains()
887 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
890 uint32_t mem_type = bo->tbo.mem.mem_type;
892 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
897 if (max_offset != 0) {
898 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
899 WARN_ON_ONCE(max_offset <
900 (amdgpu_bo_gpu_offset(bo) - domain_start));
906 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
907 /* force to pin into visible video ram */
908 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
909 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
910 amdgpu_bo_placement_from_domain(bo, domain);
911 for (i = 0; i < bo->placement.num_placement; i++) {
914 fpfn = min_offset >> PAGE_SHIFT;
915 lpfn = max_offset >> PAGE_SHIFT;
917 if (fpfn > bo->placements[i].fpfn)
918 bo->placements[i].fpfn = fpfn;
919 if (!bo->placements[i].lpfn ||
920 (lpfn && lpfn < bo->placements[i].lpfn))
921 bo->placements[i].lpfn = lpfn;
922 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
925 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
927 dev_err(adev->dev, "%p pin failed\n", bo);
933 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
934 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
935 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
936 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
937 &adev->visible_pin_size);
938 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
939 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
947 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
948 * @bo: &amdgpu_bo buffer object to be pinned
949 * @domain: domain to be pinned to
951 * A simple wrapper to amdgpu_bo_pin_restricted().
952 * Provides a simpler API for buffers that do not have any strict restrictions
953 * on where a buffer must be located.
956 * 0 for success or a negative error code on failure.
958 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
960 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
964 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
965 * @bo: &amdgpu_bo buffer object to be unpinned
967 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
968 * Changes placement and pin size accordingly.
971 * 0 for success or a negative error code on failure.
973 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
975 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
976 struct ttm_operation_ctx ctx = { false, false };
979 if (!bo->pin_count) {
980 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
987 amdgpu_bo_subtract_pin_size(bo);
989 for (i = 0; i < bo->placement.num_placement; i++) {
990 bo->placements[i].lpfn = 0;
991 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
993 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
995 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
1001 * amdgpu_bo_evict_vram - evict VRAM buffers
1002 * @adev: amdgpu device object
1004 * Evicts all VRAM buffers on the lru list of the memory type.
1005 * Mainly used for evicting vram at suspend time.
1008 * 0 for success or a negative error code on failure.
1010 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1012 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
1013 #ifndef CONFIG_HIBERNATION
1014 if (adev->flags & AMD_IS_APU) {
1015 /* Useless to evict on IGP chips */
1019 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
1022 static const char *amdgpu_vram_names[] = {
1035 * amdgpu_bo_init - initialize memory manager
1036 * @adev: amdgpu device object
1038 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1041 * 0 for success or a negative error code on failure.
1043 int amdgpu_bo_init(struct amdgpu_device *adev)
1045 /* reserve PAT memory space to WC for VRAM */
1046 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1047 adev->gmc.aper_size);
1049 /* Add an MTRR for the VRAM */
1050 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1051 adev->gmc.aper_size);
1052 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1053 adev->gmc.mc_vram_size >> 20,
1054 (unsigned long long)adev->gmc.aper_size >> 20);
1055 DRM_INFO("RAM width %dbits %s\n",
1056 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1057 return amdgpu_ttm_init(adev);
1061 * amdgpu_bo_late_init - late init
1062 * @adev: amdgpu device object
1064 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1068 * 0 for success or a negative error code on failure.
1070 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1072 amdgpu_ttm_late_init(adev);
1078 * amdgpu_bo_fini - tear down memory manager
1079 * @adev: amdgpu device object
1081 * Reverses amdgpu_bo_init() to tear down memory manager.
1083 void amdgpu_bo_fini(struct amdgpu_device *adev)
1085 amdgpu_ttm_fini(adev);
1086 arch_phys_wc_del(adev->gmc.vram_mtrr);
1087 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1091 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1092 * @bo: &amdgpu_bo buffer object
1093 * @vma: vma as input from the fbdev mmap method
1095 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1098 * 0 for success or a negative error code on failure.
1100 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1101 struct vm_area_struct *vma)
1103 return ttm_fbdev_mmap(vma, &bo->tbo);
1107 * amdgpu_bo_set_tiling_flags - set tiling flags
1108 * @bo: &amdgpu_bo buffer object
1109 * @tiling_flags: new flags
1111 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1112 * kernel driver to set the tiling flags on a buffer.
1115 * 0 for success or a negative error code on failure.
1117 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1119 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1121 if (adev->family <= AMDGPU_FAMILY_CZ &&
1122 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1125 bo->tiling_flags = tiling_flags;
1130 * amdgpu_bo_get_tiling_flags - get tiling flags
1131 * @bo: &amdgpu_bo buffer object
1132 * @tiling_flags: returned flags
1134 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1135 * set the tiling flags on a buffer.
1137 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1139 lockdep_assert_held(&bo->tbo.resv->lock.base);
1142 *tiling_flags = bo->tiling_flags;
1146 * amdgpu_bo_set_metadata - set metadata
1147 * @bo: &amdgpu_bo buffer object
1148 * @metadata: new metadata
1149 * @metadata_size: size of the new metadata
1150 * @flags: flags of the new metadata
1152 * Sets buffer object's metadata, its size and flags.
1153 * Used via GEM ioctl.
1156 * 0 for success or a negative error code on failure.
1158 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1159 uint32_t metadata_size, uint64_t flags)
1163 if (!metadata_size) {
1164 if (bo->metadata_size) {
1165 kfree(bo->metadata);
1166 bo->metadata = NULL;
1167 bo->metadata_size = 0;
1172 if (metadata == NULL)
1175 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1179 kfree(bo->metadata);
1180 bo->metadata_flags = flags;
1181 bo->metadata = buffer;
1182 bo->metadata_size = metadata_size;
1188 * amdgpu_bo_get_metadata - get metadata
1189 * @bo: &amdgpu_bo buffer object
1190 * @buffer: returned metadata
1191 * @buffer_size: size of the buffer
1192 * @metadata_size: size of the returned metadata
1193 * @flags: flags of the returned metadata
1195 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1196 * less than metadata_size.
1197 * Used via GEM ioctl.
1200 * 0 for success or a negative error code on failure.
1202 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1203 size_t buffer_size, uint32_t *metadata_size,
1206 if (!buffer && !metadata_size)
1210 if (buffer_size < bo->metadata_size)
1213 if (bo->metadata_size)
1214 memcpy(buffer, bo->metadata, bo->metadata_size);
1218 *metadata_size = bo->metadata_size;
1220 *flags = bo->metadata_flags;
1226 * amdgpu_bo_move_notify - notification about a memory move
1227 * @bo: pointer to a buffer object
1228 * @evict: if this move is evicting the buffer from the graphics address space
1229 * @new_mem: new information of the bufer object
1231 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1233 * TTM driver callback which is called when ttm moves a buffer.
1235 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1237 struct ttm_mem_reg *new_mem)
1239 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1240 struct amdgpu_bo *abo;
1241 struct ttm_mem_reg *old_mem = &bo->mem;
1243 if (!amdgpu_bo_is_amdgpu_bo(bo))
1246 abo = ttm_to_amdgpu_bo(bo);
1247 amdgpu_vm_bo_invalidate(adev, abo, evict);
1249 amdgpu_bo_kunmap(abo);
1251 /* remember the eviction */
1253 atomic64_inc(&adev->num_evictions);
1255 /* update statistics */
1259 /* move_notify is called before move happens */
1260 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1264 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1265 * @bo: pointer to a buffer object
1267 * Notifies the driver we are taking a fault on this BO and have reserved it,
1268 * also performs bookkeeping.
1269 * TTM driver callback for dealing with vm faults.
1272 * 0 for success or a negative error code on failure.
1274 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1276 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1277 struct ttm_operation_ctx ctx = { false, false };
1278 struct amdgpu_bo *abo;
1279 unsigned long offset, size;
1282 if (!amdgpu_bo_is_amdgpu_bo(bo))
1285 abo = ttm_to_amdgpu_bo(bo);
1287 /* Remember that this BO was accessed by the CPU */
1288 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1290 if (bo->mem.mem_type != TTM_PL_VRAM)
1293 size = bo->mem.num_pages << PAGE_SHIFT;
1294 offset = bo->mem.start << PAGE_SHIFT;
1295 if ((offset + size) <= adev->gmc.visible_vram_size)
1298 /* Can't move a pinned BO to visible VRAM */
1299 if (abo->pin_count > 0)
1302 /* hurrah the memory is not visible ! */
1303 atomic64_inc(&adev->num_vram_cpu_page_faults);
1304 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1305 AMDGPU_GEM_DOMAIN_GTT);
1307 /* Avoid costly evictions; only set GTT as a busy placement */
1308 abo->placement.num_busy_placement = 1;
1309 abo->placement.busy_placement = &abo->placements[1];
1311 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1312 if (unlikely(r != 0))
1315 offset = bo->mem.start << PAGE_SHIFT;
1316 /* this should never happen */
1317 if (bo->mem.mem_type == TTM_PL_VRAM &&
1318 (offset + size) > adev->gmc.visible_vram_size)
1325 * amdgpu_bo_fence - add fence to buffer object
1327 * @bo: buffer object in question
1328 * @fence: fence to add
1329 * @shared: true if fence should be added shared
1332 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1335 struct reservation_object *resv = bo->tbo.resv;
1338 reservation_object_add_shared_fence(resv, fence);
1340 reservation_object_add_excl_fence(resv, fence);
1344 * amdgpu_bo_gpu_offset - return GPU offset of bo
1345 * @bo: amdgpu object for which we query the offset
1347 * Note: object should either be pinned or reserved when calling this
1348 * function, it might be useful to add check for this for debugging.
1351 * current GPU offset of the object.
1353 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1355 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1356 WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1357 !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1358 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1359 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1360 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1362 return amdgpu_gmc_sign_extend(bo->tbo.offset);
1366 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1367 * @adev: amdgpu device object
1368 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1371 * Which of the allowed domains is preferred for pinning the BO for scanout.
1373 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1376 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1377 domain = AMDGPU_GEM_DOMAIN_VRAM;
1378 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1379 domain = AMDGPU_GEM_DOMAIN_GTT;