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 if (bp->domain & (AMDGPU_GEM_DOMAIN_GDS | AMDGPU_GEM_DOMAIN_GWS |
431 AMDGPU_GEM_DOMAIN_OA))
434 size = ALIGN(size, PAGE_SIZE);
436 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
441 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
442 sizeof(struct amdgpu_bo));
444 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
447 drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
448 INIT_LIST_HEAD(&bo->shadow_list);
450 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
452 bo->allowed_domains = bo->preferred_domains;
453 if (bp->type != ttm_bo_type_kernel &&
454 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
455 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
457 bo->flags = bp->flags;
460 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
461 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
463 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
464 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
465 /* Don't try to enable write-combining when it can't work, or things
467 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
470 #ifndef CONFIG_COMPILE_TEST
471 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
472 thanks to write-combining
475 if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
476 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
477 "better performance thanks to write-combining\n");
478 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
480 /* For architectures that don't support WC memory,
481 * mask out the WC flag from the BO
483 if (!drm_arch_can_wc_memory())
484 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
487 bo->tbo.bdev = &adev->mman.bdev;
488 amdgpu_bo_placement_from_domain(bo, bp->domain);
489 if (bp->type == ttm_bo_type_kernel)
490 bo->tbo.priority = 1;
492 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
493 &bo->placement, page_align, &ctx, acc_size,
494 NULL, bp->resv, &amdgpu_bo_destroy);
495 if (unlikely(r != 0))
498 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
499 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
500 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
501 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
504 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
506 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
507 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
508 struct dma_fence *fence;
510 r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
514 amdgpu_bo_fence(bo, fence, false);
515 dma_fence_put(bo->tbo.moving);
516 bo->tbo.moving = dma_fence_get(fence);
517 dma_fence_put(fence);
520 amdgpu_bo_unreserve(bo);
523 trace_amdgpu_bo_create(bo);
525 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
526 if (bp->type == ttm_bo_type_device)
527 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
533 ww_mutex_unlock(&bo->tbo.resv->lock);
534 amdgpu_bo_unref(&bo);
538 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
540 struct amdgpu_bo *bo)
542 struct amdgpu_bo_param bp;
548 memset(&bp, 0, sizeof(bp));
550 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
551 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
552 AMDGPU_GEM_CREATE_SHADOW;
553 bp.type = ttm_bo_type_kernel;
554 bp.resv = bo->tbo.resv;
556 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
558 bo->shadow->parent = amdgpu_bo_ref(bo);
559 mutex_lock(&adev->shadow_list_lock);
560 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
561 mutex_unlock(&adev->shadow_list_lock);
568 * amdgpu_bo_create - create an &amdgpu_bo buffer object
569 * @adev: amdgpu device object
570 * @bp: parameters to be used for the buffer object
571 * @bo_ptr: pointer to the buffer object pointer
573 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
575 * Shadow object is used to backup the original buffer object, and is always
579 * 0 for success or a negative error code on failure.
581 int amdgpu_bo_create(struct amdgpu_device *adev,
582 struct amdgpu_bo_param *bp,
583 struct amdgpu_bo **bo_ptr)
585 u64 flags = bp->flags;
588 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
589 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
593 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
595 WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
598 r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
601 reservation_object_unlock((*bo_ptr)->tbo.resv);
604 amdgpu_bo_unref(bo_ptr);
611 * amdgpu_bo_backup_to_shadow - Backs up an &amdgpu_bo buffer object
612 * @adev: amdgpu device object
613 * @ring: amdgpu_ring for the engine handling the buffer operations
614 * @bo: &amdgpu_bo buffer to be backed up
615 * @resv: reservation object with embedded fence
616 * @fence: dma_fence associated with the operation
617 * @direct: whether to submit the job directly
619 * Copies an &amdgpu_bo buffer object to its shadow object.
623 * 0 for success or a negative error code on failure.
625 int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
626 struct amdgpu_ring *ring,
627 struct amdgpu_bo *bo,
628 struct reservation_object *resv,
629 struct dma_fence **fence,
633 struct amdgpu_bo *shadow = bo->shadow;
634 uint64_t bo_addr, shadow_addr;
640 bo_addr = amdgpu_bo_gpu_offset(bo);
641 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
643 r = reservation_object_reserve_shared(bo->tbo.resv);
647 r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
648 amdgpu_bo_size(bo), resv, fence,
651 amdgpu_bo_fence(bo, *fence, true);
658 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
659 * @bo: pointer to the buffer object
661 * Sets placement according to domain; and changes placement and caching
662 * policy of the buffer object according to the placement.
663 * This is used for validating shadow bos. It calls ttm_bo_validate() to
664 * make sure the buffer is resident where it needs to be.
667 * 0 for success or a negative error code on failure.
669 int amdgpu_bo_validate(struct amdgpu_bo *bo)
671 struct ttm_operation_ctx ctx = { false, false };
678 domain = bo->preferred_domains;
681 amdgpu_bo_placement_from_domain(bo, domain);
682 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
683 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
684 domain = bo->allowed_domains;
692 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
694 * @shadow: &amdgpu_bo shadow to be restored
695 * @fence: dma_fence associated with the operation
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_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
707 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
708 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
709 uint64_t shadow_addr, parent_addr;
711 shadow_addr = amdgpu_bo_gpu_offset(shadow);
712 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
714 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
715 amdgpu_bo_size(shadow), NULL, fence,
720 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
721 * @bo: &amdgpu_bo buffer object to be mapped
722 * @ptr: kernel virtual address to be returned
724 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
725 * amdgpu_bo_kptr() to get the kernel virtual address.
728 * 0 for success or a negative error code on failure.
730 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
735 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
738 kptr = amdgpu_bo_kptr(bo);
745 r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
746 MAX_SCHEDULE_TIMEOUT);
750 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
755 *ptr = amdgpu_bo_kptr(bo);
761 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
762 * @bo: &amdgpu_bo buffer object
764 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
767 * the virtual address of a buffer object area.
769 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
773 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
777 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
778 * @bo: &amdgpu_bo buffer object to be unmapped
780 * Unmaps a kernel map set up by amdgpu_bo_kmap().
782 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
785 ttm_bo_kunmap(&bo->kmap);
789 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
790 * @bo: &amdgpu_bo buffer object
792 * References the contained &ttm_buffer_object.
795 * a refcounted pointer to the &amdgpu_bo buffer object.
797 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
802 ttm_bo_get(&bo->tbo);
807 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
808 * @bo: &amdgpu_bo buffer object
810 * Unreferences the contained &ttm_buffer_object and clear the pointer
812 void amdgpu_bo_unref(struct amdgpu_bo **bo)
814 struct ttm_buffer_object *tbo;
825 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
826 * @bo: &amdgpu_bo buffer object to be pinned
827 * @domain: domain to be pinned to
828 * @min_offset: the start of requested address range
829 * @max_offset: the end of requested address range
831 * Pins the buffer object according to requested domain and address range. If
832 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
833 * pin_count and pin_size accordingly.
835 * Pinning means to lock pages in memory along with keeping them at a fixed
836 * offset. It is required when a buffer can not be moved, for example, when
837 * a display buffer is being scanned out.
839 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
840 * where to pin a buffer if there are specific restrictions on where a buffer
844 * 0 for success or a negative error code on failure.
846 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
847 u64 min_offset, u64 max_offset)
849 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
850 struct ttm_operation_ctx ctx = { false, false };
853 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
856 if (WARN_ON_ONCE(min_offset > max_offset))
859 /* A shared bo cannot be migrated to VRAM */
860 if (bo->prime_shared_count) {
861 if (domain & AMDGPU_GEM_DOMAIN_GTT)
862 domain = AMDGPU_GEM_DOMAIN_GTT;
867 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
868 * See function amdgpu_display_supported_domains()
870 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
873 uint32_t mem_type = bo->tbo.mem.mem_type;
875 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
880 if (max_offset != 0) {
881 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
882 WARN_ON_ONCE(max_offset <
883 (amdgpu_bo_gpu_offset(bo) - domain_start));
889 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
890 /* force to pin into visible video ram */
891 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
892 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
893 amdgpu_bo_placement_from_domain(bo, domain);
894 for (i = 0; i < bo->placement.num_placement; i++) {
897 fpfn = min_offset >> PAGE_SHIFT;
898 lpfn = max_offset >> PAGE_SHIFT;
900 if (fpfn > bo->placements[i].fpfn)
901 bo->placements[i].fpfn = fpfn;
902 if (!bo->placements[i].lpfn ||
903 (lpfn && lpfn < bo->placements[i].lpfn))
904 bo->placements[i].lpfn = lpfn;
905 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
908 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
910 dev_err(adev->dev, "%p pin failed\n", bo);
916 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
917 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
918 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
919 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
920 &adev->visible_pin_size);
921 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
922 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
930 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
931 * @bo: &amdgpu_bo buffer object to be pinned
932 * @domain: domain to be pinned to
934 * A simple wrapper to amdgpu_bo_pin_restricted().
935 * Provides a simpler API for buffers that do not have any strict restrictions
936 * on where a buffer must be located.
939 * 0 for success or a negative error code on failure.
941 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
943 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
947 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
948 * @bo: &amdgpu_bo buffer object to be unpinned
950 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
951 * Changes placement and pin size accordingly.
954 * 0 for success or a negative error code on failure.
956 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
958 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
959 struct ttm_operation_ctx ctx = { false, false };
962 if (!bo->pin_count) {
963 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
970 amdgpu_bo_subtract_pin_size(bo);
972 for (i = 0; i < bo->placement.num_placement; i++) {
973 bo->placements[i].lpfn = 0;
974 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
976 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
978 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
984 * amdgpu_bo_evict_vram - evict VRAM buffers
985 * @adev: amdgpu device object
987 * Evicts all VRAM buffers on the lru list of the memory type.
988 * Mainly used for evicting vram at suspend time.
991 * 0 for success or a negative error code on failure.
993 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
995 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
996 #ifndef CONFIG_HIBERNATION
997 if (adev->flags & AMD_IS_APU) {
998 /* Useless to evict on IGP chips */
1002 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
1005 static const char *amdgpu_vram_names[] = {
1018 * amdgpu_bo_init - initialize memory manager
1019 * @adev: amdgpu device object
1021 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1024 * 0 for success or a negative error code on failure.
1026 int amdgpu_bo_init(struct amdgpu_device *adev)
1028 /* reserve PAT memory space to WC for VRAM */
1029 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1030 adev->gmc.aper_size);
1032 /* Add an MTRR for the VRAM */
1033 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1034 adev->gmc.aper_size);
1035 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1036 adev->gmc.mc_vram_size >> 20,
1037 (unsigned long long)adev->gmc.aper_size >> 20);
1038 DRM_INFO("RAM width %dbits %s\n",
1039 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1040 return amdgpu_ttm_init(adev);
1044 * amdgpu_bo_late_init - late init
1045 * @adev: amdgpu device object
1047 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1051 * 0 for success or a negative error code on failure.
1053 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1055 amdgpu_ttm_late_init(adev);
1061 * amdgpu_bo_fini - tear down memory manager
1062 * @adev: amdgpu device object
1064 * Reverses amdgpu_bo_init() to tear down memory manager.
1066 void amdgpu_bo_fini(struct amdgpu_device *adev)
1068 amdgpu_ttm_fini(adev);
1069 arch_phys_wc_del(adev->gmc.vram_mtrr);
1070 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1074 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1075 * @bo: &amdgpu_bo buffer object
1076 * @vma: vma as input from the fbdev mmap method
1078 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1081 * 0 for success or a negative error code on failure.
1083 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1084 struct vm_area_struct *vma)
1086 return ttm_fbdev_mmap(vma, &bo->tbo);
1090 * amdgpu_bo_set_tiling_flags - set tiling flags
1091 * @bo: &amdgpu_bo buffer object
1092 * @tiling_flags: new flags
1094 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1095 * kernel driver to set the tiling flags on a buffer.
1098 * 0 for success or a negative error code on failure.
1100 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1102 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1104 if (adev->family <= AMDGPU_FAMILY_CZ &&
1105 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1108 bo->tiling_flags = tiling_flags;
1113 * amdgpu_bo_get_tiling_flags - get tiling flags
1114 * @bo: &amdgpu_bo buffer object
1115 * @tiling_flags: returned flags
1117 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1118 * set the tiling flags on a buffer.
1120 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1122 lockdep_assert_held(&bo->tbo.resv->lock.base);
1125 *tiling_flags = bo->tiling_flags;
1129 * amdgpu_bo_set_metadata - set metadata
1130 * @bo: &amdgpu_bo buffer object
1131 * @metadata: new metadata
1132 * @metadata_size: size of the new metadata
1133 * @flags: flags of the new metadata
1135 * Sets buffer object's metadata, its size and flags.
1136 * Used via GEM ioctl.
1139 * 0 for success or a negative error code on failure.
1141 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1142 uint32_t metadata_size, uint64_t flags)
1146 if (!metadata_size) {
1147 if (bo->metadata_size) {
1148 kfree(bo->metadata);
1149 bo->metadata = NULL;
1150 bo->metadata_size = 0;
1155 if (metadata == NULL)
1158 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1162 kfree(bo->metadata);
1163 bo->metadata_flags = flags;
1164 bo->metadata = buffer;
1165 bo->metadata_size = metadata_size;
1171 * amdgpu_bo_get_metadata - get metadata
1172 * @bo: &amdgpu_bo buffer object
1173 * @buffer: returned metadata
1174 * @buffer_size: size of the buffer
1175 * @metadata_size: size of the returned metadata
1176 * @flags: flags of the returned metadata
1178 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1179 * less than metadata_size.
1180 * Used via GEM ioctl.
1183 * 0 for success or a negative error code on failure.
1185 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1186 size_t buffer_size, uint32_t *metadata_size,
1189 if (!buffer && !metadata_size)
1193 if (buffer_size < bo->metadata_size)
1196 if (bo->metadata_size)
1197 memcpy(buffer, bo->metadata, bo->metadata_size);
1201 *metadata_size = bo->metadata_size;
1203 *flags = bo->metadata_flags;
1209 * amdgpu_bo_move_notify - notification about a memory move
1210 * @bo: pointer to a buffer object
1211 * @evict: if this move is evicting the buffer from the graphics address space
1212 * @new_mem: new information of the bufer object
1214 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1216 * TTM driver callback which is called when ttm moves a buffer.
1218 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1220 struct ttm_mem_reg *new_mem)
1222 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1223 struct amdgpu_bo *abo;
1224 struct ttm_mem_reg *old_mem = &bo->mem;
1226 if (!amdgpu_bo_is_amdgpu_bo(bo))
1229 abo = ttm_to_amdgpu_bo(bo);
1230 amdgpu_vm_bo_invalidate(adev, abo, evict);
1232 amdgpu_bo_kunmap(abo);
1234 /* remember the eviction */
1236 atomic64_inc(&adev->num_evictions);
1238 /* update statistics */
1242 /* move_notify is called before move happens */
1243 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1247 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1248 * @bo: pointer to a buffer object
1250 * Notifies the driver we are taking a fault on this BO and have reserved it,
1251 * also performs bookkeeping.
1252 * TTM driver callback for dealing with vm faults.
1255 * 0 for success or a negative error code on failure.
1257 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1259 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1260 struct ttm_operation_ctx ctx = { false, false };
1261 struct amdgpu_bo *abo;
1262 unsigned long offset, size;
1265 if (!amdgpu_bo_is_amdgpu_bo(bo))
1268 abo = ttm_to_amdgpu_bo(bo);
1270 /* Remember that this BO was accessed by the CPU */
1271 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1273 if (bo->mem.mem_type != TTM_PL_VRAM)
1276 size = bo->mem.num_pages << PAGE_SHIFT;
1277 offset = bo->mem.start << PAGE_SHIFT;
1278 if ((offset + size) <= adev->gmc.visible_vram_size)
1281 /* Can't move a pinned BO to visible VRAM */
1282 if (abo->pin_count > 0)
1285 /* hurrah the memory is not visible ! */
1286 atomic64_inc(&adev->num_vram_cpu_page_faults);
1287 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1288 AMDGPU_GEM_DOMAIN_GTT);
1290 /* Avoid costly evictions; only set GTT as a busy placement */
1291 abo->placement.num_busy_placement = 1;
1292 abo->placement.busy_placement = &abo->placements[1];
1294 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1295 if (unlikely(r != 0))
1298 offset = bo->mem.start << PAGE_SHIFT;
1299 /* this should never happen */
1300 if (bo->mem.mem_type == TTM_PL_VRAM &&
1301 (offset + size) > adev->gmc.visible_vram_size)
1308 * amdgpu_bo_fence - add fence to buffer object
1310 * @bo: buffer object in question
1311 * @fence: fence to add
1312 * @shared: true if fence should be added shared
1315 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1318 struct reservation_object *resv = bo->tbo.resv;
1321 reservation_object_add_shared_fence(resv, fence);
1323 reservation_object_add_excl_fence(resv, fence);
1327 * amdgpu_bo_gpu_offset - return GPU offset of bo
1328 * @bo: amdgpu object for which we query the offset
1330 * Note: object should either be pinned or reserved when calling this
1331 * function, it might be useful to add check for this for debugging.
1334 * current GPU offset of the object.
1336 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1338 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1339 WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1340 !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1341 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1342 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1343 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1345 return amdgpu_gmc_sign_extend(bo->tbo.offset);
1349 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1350 * @adev: amdgpu device object
1351 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1354 * Which of the allowed domains is preferred for pinning the BO for scanout.
1356 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1359 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1360 domain = AMDGPU_GEM_DOMAIN_VRAM;
1361 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1362 domain = AMDGPU_GEM_DOMAIN_GTT;